Download the full version of the textbook now at textbookfull.

com

Bayesian Prediction and Adaptive Sampling

Algorithms for Mobile Sensor Networks: Online
Environmental Field Reconstruction in Space
and Time 1st Edition Yunfei Xu
https://textbookfull.com/product/bayesian-
prediction-and-adaptive-sampling-algorithms-for-
mobile-sensor-networks-online-environmental-field-
reconstruction-in-space-and-time-1st-edition-
yunfei-xu/

Explore and download more textbook at https://textbookfull.com

 Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.

Clustering and Routing Algorithms for Wireless Sensor

Networks Energy Efficiency Approaches 1st Edition Pratyay
Kuila
https://textbookfull.com/product/clustering-and-routing-algorithms-
for-wireless-sensor-networks-energy-efficiency-approaches-1st-edition-
pratyay-kuila/
textbookfull.com

Mobile Wireless and Sensor Networks A Clustering Algorithm

for Energy Efficiency and Safety 1st Edition Amine Dahane
(Author)
https://textbookfull.com/product/mobile-wireless-and-sensor-networks-
a-clustering-algorithm-for-energy-efficiency-and-safety-1st-edition-
amine-dahane-author/
textbookfull.com

Mobile Agent Based Anomaly Detection and Verification

System for Smart Home Sensor Networks 1st Edition Muhammad
Usman
https://textbookfull.com/product/mobile-agent-based-anomaly-detection-
and-verification-system-for-smart-home-sensor-networks-1st-edition-
muhammad-usman/
textbookfull.com

Finn Noel The Bloodwood Academy 3 5 1st Edition Rae Foxx

Foxx Rae

https://textbookfull.com/product/finn-noel-the-bloodwood-
academy-3-5-1st-edition-rae-foxx-foxx-rae/

textbookfull.com
Urban Practices from Delicacy Management to Governance in
Contemporary China The Case of Xuhui District Shanghai
Gaohong Chen
https://textbookfull.com/product/urban-practices-from-delicacy-
management-to-governance-in-contemporary-china-the-case-of-xuhui-
district-shanghai-gaohong-chen/
textbookfull.com

The Routledge International Handbook of Clinical Hypnosis

1st Edition Julie H Linden

https://textbookfull.com/product/the-routledge-international-handbook-
of-clinical-hypnosis-1st-edition-julie-h-linden/

textbookfull.com

Formal Problems in Semitic Phonology and Morphology 2nd

Edition John J. Mccarthy

https://textbookfull.com/product/formal-problems-in-semitic-phonology-
and-morphology-2nd-edition-john-j-mccarthy/

textbookfull.com

The Victorian Ghost Story and Theology: From Le Fanu to

James 1st Edition Zoe Lehmann Imfeld (Auth.)

https://textbookfull.com/product/the-victorian-ghost-story-and-
theology-from-le-fanu-to-james-1st-edition-zoe-lehmann-imfeld-auth/

textbookfull.com

Fast Facts for the ER Nurse: Emergency Department

Orientation in a Nutshell Jennifer R. Buettner

https://textbookfull.com/product/fast-facts-for-the-er-nurse-
emergency-department-orientation-in-a-nutshell-jennifer-r-buettner/

textbookfull.com
Photocatalysis applications Dionysios D Dionysiou

https://textbookfull.com/product/photocatalysis-applications-
dionysios-d-dionysiou/

textbookfull.com
SPRINGER BRIEFS IN ELEC TRIC AL AND COMPUTER
ENGINEERING  CONTROL, AUTOMATION AND ROBOTICS

Yunfei Xu
Jongeun Choi
Sarat Dass
Tapabrata Maiti

Bayesian Prediction
and Adaptive Sampling
Algorithms for Mobile
Sensor Networks
Online Environmental
Field Reconstruction in
Space and Time

123
SpringerBriefs in Electrical and Computer
Engineering

Control, Automation and Robotics

Series editors
Tamer Başar
Antonio Bicchi
Miroslav Krstic
More information about this series at http://www.springer.com/series/10198
Yunfei Xu Jongeun Choi Sarat Dass
• •

Tapabrata Maiti

Bayesian Prediction
and Adaptive Sampling
Algorithms for Mobile
Sensor Networks
Online Environmental Field Reconstruction
in Space and Time

123
Yunfei Xu Sarat Dass
Michigan State University Department of Statistics
East Lansing, MI Michigan State University
USA East Lansing, MI
USA
Jongeun Choi
Michigan State University Tapabrata Maiti
East Lansing, MI Department of Statistics
USA Michigan State University
East Lansing, MI
USA

ISSN 2191-8112 ISSN 2191-8120 (electronic)

SpringerBriefs in Electrical and Computer Engineering
ISSN 2192-6786 ISSN 2192-6794 (electronic)
SpringerBriefs in Control, Automation and Robotics
ISBN 978-3-319-21920-2 ISBN 978-3-319-21921-9 (eBook)
DOI 10.1007/978-3-319-21921-9

Library of Congress Control Number: 2015950872

Springer Cham Heidelberg New York Dordrecht London

© The Author(s) 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
To our loving parents and beautiful families
Preface

We have witnessed a surge of applications using static or mobile sensor networks

interacting with uncertain environments. To treat a variety of useful tasks such as
environmental monitoring, adaptive sampling, surveillance, and exploration, this
book introduces a class of problems and efficient spatio-temporal models when
scalar fields need to be predicted from noisy observations collected by mobile
sensor networks. The book discusses how to make inference from the observations
based on the proposed models and also explores adaptive sampling algorithms for
robotic sensors to maximize the prediction quality subject to constraints on mem-
ory, communication, and mobility.
The objective of the book is to provide step-by-step progress in chapters for
readers to gain better understanding of the interplay between all the essential
constituents such as resource-limited mobile sensor networks, spatio-temporal
models, data-driven prediction, prediction uncertainty, and adaptive sampling for
making better predictions. The book builds on previous collective works by the
authors and is not meant to provide a comprehensive review of the topics of
interest. Specifically, materials from the previous publications by the authors [1–5]
make up a large portion of the book.
In this book, a spatio-temporal scalar field is used to represent the collection of
scalar quantities of interest, such as chemical concentration or biomass of algal
blooms (e.g., see Fig. 1.3), transported via physical processes. To deal with com-
plexity and practicality, phenomenological and statistical modeling techniques are
used to make inference from noisy observations collected, taking into account a
large scope of uncertainties. To this end, nonparametric models such as Gaussian
processes and Gaussian Markov random fields (GMRFs), along with their predic-
tion and adaptive sampling algorithms, will be explored and tailored to our needs.
The importance of selecting a Gaussian process prior via hyperparameters for given
experimental observations is illustrated (Chap. 3). Adaptive sampling to improve
the quality of hyperparameters is proposed (Chap. 3). Memory efficient prediction
based on truncated observations in space and time as well as the collective mobility
based on distributed navigation are discussed (Chap. 4). While the book starts with

vii
viii Preface

a rather simple empirical Bayes approach (Chap. 3), as we move through further
chapters, we discuss recent efforts with a fully Bayesian perspective to maximize
the flexibility of the models under various uncertainties while minimizing the
computational complexity (Chaps. 5 and 7). A fully Bayesian framework is adopted
here as it offers several advantages when inferring parameters and processes from
highly complex models (Chaps. 5 and 7). The Bayesian approach requires prior
distributions to be elicited for model parameters that are of interest. Once the priors
are elicited, the Bayesian framework is flexible and effective in incorporating all
uncertainties as well as information (limited or otherwise from data) into a single
entity, namely, the posterior. The fully Bayesian approach thus allows additional
sources and extent of uncertainties to be integrated into the inferential framework,
with the posterior distribution effectively capturing all aspects of uncertainties
involved. Subsequently, the practitioner needs only to focus on different compo-
nents of the posterior to obtain inference separately for the parameters of interest,
nuisance parameters, and hyperparameters. The fully Bayesian approach also
allows data to select the most appropriate values for nuisance parameters and
hyperparameters automatically and achieve optimal inference and prediction for the
scalar field. In this book, a fully Bayesian approach for spatio-temporal Gaussian
process regression will be formulated for resource-constrained robotic sensors to
fuse multifactorial effects of observations, measurement noise, and prior distribu-
tions for obtaining the predictive distribution of a scalar environmental field of
interest. Traditional Markov Chain Monte Carlo (MCMC) methods cannot be
implemented on resource-constrained mobile sensor networks due to high com-
putational complexity. To deal with complexity, the Bayesian spatio-temporal
models will be carefully tailored (Chap. 5). For example, we will approximate a
Gaussian process with a GMRF for computational efficiency (Chaps. 6 and 7).
A new spatial model is proposed via a GMRF (Chap. 6). In addition, ways to
improve computational efficiency will be proposed in form of empirical Bayes and
approximate Bayes instead of MCMC-based computation. For some special cases,
the developed centralized algorithms will be further refined in a distributed manner
such that mobile robots can implement distributed algorithms only using local
information available from neighboring robots over a proximity communication
graph (Chaps. 4–6).
We note that although regression problems for sensor networks under location
uncertainty have practical importance, they are not considered in this book. The
interested reader is referred to [6, 7] (centralized scheme) and [8] (distributed
scheme) for further information on this topic.

Organization

This book is organized as follows: Chapter 1 gives some background information

and a summary for each chapter. In Chap. 2, we introduce the basic mathematical
notation that will be used throughout the book. We then describe the general
Preface ix

Chapter 1: Background: Mobile sen-

sor network (MSN) and robotic sensors

Chapter 2: Preliminaries: Gaussian process (GP)

and Gaussian Markov random field (GMRF)

Chapter 3: GP, prediction, adaptive

sampling, & empirical Bayes appr.

Chapter 4: GP, prediction, & distributed strategy

Chapter 5: GP, prediction, adaptive sam-

pling, MCMC appr., & fully Bayesian appr.

Chapter 6: new spatial model, GMRF, predic-

tion, distributed strategy, & empirical Bayes appr.

Chapter 7: approximated GP, GMRF, predic-

tion, adaptive sampling, & fully Bayesian appr.

Fig. 1 Organization of chapters along with keywords

Gaussian process and its usage in nonparametric regression problems. The notations
for mobile sensor networks are also introduced in Chap. 2. In Chap. 3, we deal with
the case where hyperparameters in the covariance function is deterministic but
unknown. We design an optimal sampling strategy to improve the maximum
likelihood estimation of these hyperparameters. In Chap. 4, we assume the
hyperparameters in the covariance function are given; they can be obtained using
the approach proposed in Chap. 3. We then analyze the error bounds of prediction
error using Gaussian process regression with truncated observations. Inspired by the
error analysis, we propose both centralized and distributed navigation strategies for
mobile sensor networks to move in order to reduce prediction error variances at
points of interest. In Chap. 5, we consider a fully Bayesian approach for Gaussian
process regression in which the hyperparameters are treated as random variables.
Using discrete prior probabilities and compactly supported kernels, we provide a
way to design sequential Bayesian prediction algorithms that can be computed in
constant time as the number of observations increases. To cope with the compu-
tational complexity brought by using standard Gaussian processes with covariance
functions, in Chap. 6, we exploit the sparsity of the precision matrix by using
Gaussian Markov random fields (GMRFs). We first introduce a new class of
Gaussian processes with built-in GMRF and show its capability of representing a
wide range of nonstationary physical processes. We then derive the formulas for
x Preface

predictive statistics and design sequential prediction algorithms with fixed com-
plexity. In Chap. 7, we consider a discretized spatial field that is modeled by a
GMRF with unknown hyperparameters. From a Bayesian perspective, we design a
sequential prediction algorithm to exactly compute the predictive inference of the
random field. An adaptive sampling strategy is also designed for mobile sensing
agents to find the most informative locations in taking future measurements in order
to minimize the prediction error and the uncertainty in the estimated hyperpara-
meters simultaneously.
Keywords for chapters are summarized in Fig. 1. While each chapter is
self-contained and so can be read independently, arrows in Fig. 1 recommend
possible reading sequences for readers.

Acknowledgments

We would like to thank Songhwai Oh at Seoul National University for his

suggestions and contribution to Chap. 4. We also thank the National Science
Foundation RET teacher, Alexander Robinson, undergraduate student David York,
and Ph.D. student Huan N. Do at Michigan State University for collecting the
experimental data using a robotic boat in Chap. 3. We thank Jeffrey W. Laut,
Maurizio Porfiri (NYU Polytechnic School of Engineering), Xiaobo Tan (Michigan
State University), and Derek A. Paley (University of Maryland) for providing
pictures of their robots used in the introduction of Chap. 1.
The authors Yunfei Xu and Jongeun Choi have been supported in part by the
National Science Foundation through CAREER Award CMMI-0846547. This
support is gratefully acknowledged. Any opinions, findings, and conclusions, or
recommendations expressed in this book are those of the authors and do not
necessarily reflect the views of the National Science Foundation.

Santa Clara, California Yunfei Xu

East Lansing, Michigan Jongeun Choi
Seri Iskandar, Malaysia Sarat Dass
East Lansing, Michigan Tapabrata Maiti
May 2015
Visit https://textbookfull.com
now to explore a rich
collection of eBooks, textbook
and enjoy exciting offers!
Contents

1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Contents in Chapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1 Mathematical Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Physical Process Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 Gaussian Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Spatiotemporal Gaussian Process . . . . . . . . . . . . . . . . . . 13
2.2.3 Gaussian Markov Random Field . . . . . . . . . . . . . . . . . . 14
2.3 Mobile Sensor Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Gaussian Processes for Regression . . . . . . . . . . . . . . . . . . . . . . 17
3 Learning Covariance Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1 Selection of Gaussian Process Prior . . . . . . . . . . . . . . . . . . . . . 19
3.2 Learning the Hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Optimal Sampling Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 Simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4 Memory Efficient Prediction With Truncated Observations . . . . . . 27
4.1 GPR with Truncated Observations . . . . . . . . . . . . . . . . . . . . . . 28
4.1.1 Error Bounds Using Truncated Observations . . . . . . . . . . 28
4.1.2 Selecting Temporal Truncation Size . . . . . . . . . . . . . . . . 34
4.2 Optimal Sampling Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2.1 Centralized Navigation Strategy . . . . . . . . . . . . . . . . . . . 37
4.2.2 Distributed Navigation Strategy . . . . . . . . . . . . . . . . . . . 39
4.3 Simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.1 Centralized Sampling Scheme . . . . . . . . . . . . . . . . . . . . 46
4.3.2 Distributed Sampling Scheme . . . . . . . . . . . . . . . . . . . . 48

xi
xii Contents

5 Fully Bayesian Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

5.1 Fully Bayesian Prediction Approach . . . . . . . . . . . . . . . . . . . . . 54
5.1.1 Prior Selection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.1.2 MCMC-Based Approach. . . . . . . . . . . . . . . . . . . . . . . . 56
5.1.3 Importance Sampling Approach . . . . . . . . . . . . . . . . . . . 61
5.1.4 Discrete Prior Distribution. . . . . . . . . . . . . . . . . . . . . . . 64
5.2 Sequential Bayesian Prediction . . . . . . . . . . . . . . . . . . . . . . . . 64
5.2.1 Scalable Bayesian Prediction Algorithm . . . . . . . . . . . . . 65
5.2.2 Distributed Implementation for a Special Case. . . . . . . . . 66
5.2.3 Adaptive Sampling. . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 Simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.3.1 MCMC-Based Approach on a 1-D Scenario . . . . . . . . . . 70
5.3.2 Centralized Scheme on 1-D Scenario . . . . . . . . . . . . . . . 71
5.3.3 Distributed Scheme on 2-D Scenario . . . . . . . . . . . . . . . 74
6 New Efficient Spatial Model with Built-In Gaussian Markov
Random Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.1 Spatial Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.1.1 Spatial Model Based on GMRF . . . . . . . . . . . . . . . . . . . 78
6.1.2 Gaussian Process Regression . . . . . . . . . . . . . . . . . . . . . 79
6.1.3 Sequential Prediction Algorithm. . . . . . . . . . . . . . . . . . . 82
6.2 Distributed Spatial Prediction. . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.2.1 Distributed Computation . . . . . . . . . . . . . . . . . . . . . . . . 84
6.2.2 Distributed Prediction Algorithm . . . . . . . . . . . . . . . . . . 84
6.3 Simulation and Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.3.1 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.3.2 Centralized Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.3.3 Distributed Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.3.4 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7 Fully Bayesian Spatial Prediction Using Gaussian Markov
Random Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.1 Spatial Field Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7.2 Bayesian Predictive Inference . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.3 Sequential Bayesian Inference . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.3.1 Update Full Conditional Distribution . . . . . . . . . . . . . . . 97
7.3.2 Update Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.3.3 Update Predictive Distribution . . . . . . . . . . . . . . . . . . . . 99
7.4 Adaptive Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.5 Simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

Appendix A: Mathematical Background. . . . . . . . . . . . . . . . . . . . . . . 107

References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Chapter 1
Introduction

1.1 Background

Sensor networks are ubiquitous due to the recent technological breakthroughs in

micro-electro-mechanical systems (MEMS), wireless communications, and embed-
ded systems [9, 10]. A sensor network consists of a collection of low-cost, low-power,
and multifunctional sensing devices that communicate over finite distances. A flexi-
ble and application-specific operating system, TinyOS, was developed at UC Berke-
ley for sensor networks with severe memory and power constraints [11]. TinyOS
runs on small and cheap wireless sensor nodes (e.g., MICA2DOT from Crossbow
Technology, Inc., CA, USA) as shown in Fig. 1.1a. Such sensor nodes have been
equipped with various environmental and ambient sensors such as temperature sen-
sors, lighting sensors, chemical sensors, accelerometers, and RFID readers along
with the communication capability with neighbors via low-power wireless commu-
nication to form a wireless ad hoc sensor network with up to 100,000 nodes [10].
Endowing the nodes in a sensor network with mobility significantly increases the
sensor network’s sampling capabilities [12, 13]. The sensor networks which consist
of mobile sensing agents are more flexible than the ones with only static nodes. A
conceptual picture of a distributed mobile sensor network with a (R-disk) proximity
communication graph model is shown in Fig. 1.1b, which assumes that a robotic
sensor can communicate with its neighboring robots within distance R. Devised in
the Laboratory of Intelligent Systems at the Swiss Federal Institute of Technology,
the Swarming Micro Air Vehicle Network (SmavNet) depicted in Fig. 1.1c allows
a single operator to control an entire swarm of cheap unmanned aerial vehicles
(UAVs) for search and rescue operation [14]. 3D Robotics, Inc., Berkeley, USA
produces “personal drones” such as DIY drone kits as well as ready-to-fly quadrotors,
multirotors, and fixed wing UAVs based on open source UAV autopilot platforms. A
quadrotor with a camera from 3D Robotics is shown in Fig. 1.2.
Biologists and other scientists are interested in leveraging recent technological
advances [10, 15, 16] by deploying mobile sensor networks for environmental and

© The Author(s) 2016 1

Y. Xu et al., Bayesian Prediction and Adaptive Sampling Algorithms
for Mobile Sensor Networks, SpringerBriefs in Control,
Automation and Robotics, DOI 10.1007/978-3-319-21921-9_1
2 1 Introduction

(a) (b)

(c)

Fig. 1.1 a Wireless microsensor mote (MICA2DOT) from Crossbow Technology, Inc., CA, USA,
(www.xbow.com). b Distributed mobile sensor network with a (R-disk) proximity communication
graph for environmental monitoring (credit: Justin Mrkva). c Swarming Micro Air Vehicle Network
(SmavNet) developed in the Laboratory of Intelligent Systems at the Swiss Federal Institute of
Technology for search and rescue operation (Photo courtesy of Laboratory of Intelligent Systems
at EPFL, http://lis.epfl.ch)

wildlife monitoring. For example, one of the pressing societal concerns about water
quality is the proliferation of harmful algal blooms in ponds, lakes, rivers, and coastal
ocean worldwide. A satellite image of a 2011 significant harmful algal bloom in
western Lake Erie is shown in Fig. 1.3. The excessive growth of cyanobacteria leads
to a decaying biomass and oxygen depletion, which are detrimental to fish and other
aquatic life as well as to land animals and humans consequently (due to the produced
1.1 Background 3

Fig. 1.2 Personal drone (IRIS+) manufactured by 3D Robotics Inc., Berkeley, USA (credit: 3D
Robotics, http://3drobotics.com)

Fig. 1.3 Satellite image of 2011 significant harmful algal bloom in western Lake Erie in
Michigan, which impacted over half of the lake shore (credit: MERIS/ESA, processed by
NOAA/NOS/NCCOS, http://www.noaanews.noaa.gov)

toxins that deteriorate water quality) [17–19]. Deploying mobile sensor networks
can be a viable way to reconstruct and monitor such harmful algal blooms [20, 21].
Indeed, we have seen the increasing exploration of robotic technologies in aquatic
sensing [20, 22–25]. A robotic boat was used in concert with stationary buoys to form
an aquatic microbial system [20]; spatiotemporal aquatic field reconstruction was
implemented using inexpensive, low-power, robotic fish in [21] (see also Fig. 1.5a
for gliding robotic fish [26]); and low-cost, self-sustained mobile surface vehicles
have been designed for environmental monitoring as part of the citizen science project
Brooklyn Atlantis [23] (see Fig. 1.4). A robotic boat equipped with a depth sensor, as
4 1 Introduction

(a)

(b)

Fig. 1.4 Robots developed by NYU Polytechnic School of Engineering for image and water quality
data collection as part of the citizen science project Brooklyn Atlantis (http://www.brooklynatlantis.
poly.edu) [23] (credit: Jeffrey W Laut)

shown in Fig. 1.5b, can sample the depth of a lake for its estimation [27]. Autonomous
underwater vehicles (AUVs) are being developed as an important tool in oceanog-
raphy, marine biology, and other maritime applications [28–30]. Autonomous sea
gliders are another noteworthy example. These battery-powered, buoyancy-driven
vehicles can travel thousands of miles horizontally, for many months, without chang-
ing or recharging batteries [31–34]. With the networks of gliders as shown in Fig. 1.6,
adaptive sampling has been demonstrated in Monterey Bay, California [35–37].
1.1 Background 5

(a)

(b)

Fig. 1.5 a Gliding robotic fish “Grace” sampling harmful algae in the Wintergreen Lake, Michigan
[26] (credit: Xiaobo Tan). b Robotic boat sampling depth near Hawk Island, Lansing, Michigan
(credit: Jongeun Choi)

The robotic sensor technologies have then brought an increasing exploitation of

navigation of mobile sensor networks and robotic sensors interacting with uncertain
environments [2, 35–42]. A necessity in such scenarios is to design algorithms to
process collected observations from environments (e.g., distributed estimators) for
robots such that either the local information about the environment can be used for
local control actions or the global information can be estimated asymptotically.
The approach of designing such algorithms takes two different paths depend-
ing on whether it uses an environmental model in space and time or not. Without
6 1 Introduction

Fig. 1.6 Gliders used for adaptive sampling [35–37] (credit: Derek A. Paley)

environmental models, extremum seeking control has been proven to be very effective
for finding a source of a signal (chemical, electromagnetic, etc.) [38, 39]. Distributed
algorithms for stochastic source seeking with mobile robot networks have been devel-
oped for both cases with and without the mutual information model between their
expected measurements and the expected source location [43]. A unifying frame-
work of distributed stochastic gradient algorithms that can deal with coverage control,
spatial partitioning, and dynamic vehicle routing problems in the absence of a priori
knowledge of the event location distribution has been presented in [40].
A drawback of the spatial model free approach is that it limits its task to finding
the maximum (or minimum) point of the environmental field. To tackle a variety
of useful tasks such as the exploration, estimation, prediction, and maximum seek-
ing of a scalar field, it is essential for robots to have a spatial (and temporal) field
model [2–4, 36, 41, 42, 44–50]. Although control algorithms for mobile robots have
been developed based on computationally demanding, physics-based field models
[51], for resource-constrained mobile robots, recently, phenomenological and statis-
tical modeling techniques such as kriging, Gaussian process regression, and kernel
regression have gained much attention. Among phenomenological spatial models,
adaptive control of multiple robotic sensors based on a parametric approach needs
a persistent excitation (PE) condition for convergence of parameters [42, 50], while
control strategies based on Bayesian spatial models do not require such conditions
(e.g., by utilizing priori distributions as in Kalman filtering [41] or Gaussian process
regression [2]). Hence, control engineers have become more aware of the useful-
ness of nonparametric Bayesian approaches such as Gaussian processes (defined by
1.1 Background 7

mean and covariance functions) [52, 53] to statistically model physical phenomena
for the navigation of mobile sensor networks, e.g., [2–4, 36, 44–47]. Other more
data-driven approaches have also developed (without statistical structure used in
Gaussian processes) such as using kernel regression [48] and in reproducing kernel
Hilbert spaces [49]. However, without a statistical structure in a random field, such
an approach (as in [48, 49]) usually requires a great number of observations than the
one with a statistical structure for a decent prediction quality.
In a mobile sensor network, the resource-limited sensing agents are required to
collaborate in order to achieve a specific objective. The cooperative control becomes
essential. The most popular applications are in networks of autonomous ground
vehicles [54, 55], underwater and surface vehicles [23, 36, 56–58], or aerial vehicles
[59–61]. Emerging technologies have been reported on the coordination of mobile
sensing agents [41, 62–70].
The mobility of mobile agents can be designed in order to perform the optimal
sampling of the field of interest. Optimal sampling design is the process of choosing
where to take samples in order to maximize the information gained. Recently, in
[36], Leonard et al. developed mobile sensor networks that optimize ocean sampling
performance defined in terms of uncertainty in a model estimate of a sampled field.
A typical sensor placement technique [71] that puts sensors at the locations where
the entropy is high tends to place sensors along the borders of the area of interest
[44]. In [44], Krause et al. showed that seeking sensor placements that are most
informative about unsensed locations is NP-hard, and they presented a polynomial
time approximation algorithm by exploiting the submodularity of mutual information
[72]. In a similar approach, in [73], Singh et al. presented an efficient planning of
informative paths for multiple robots that maximize the mutual information.
To find these locations that predict the phenomenon best, one needs a model
of the spatiotemporal phenomenon. To this end, we use Gaussian processes (and
Gaussian random fields) to model fields undergoing transport phenomena. Nonpara-
metric Gaussian process regression (or Kriging in geostatistics) has been widely used
as a nonlinear regression technique to estimate and predict geostatistical data [52, 53,
74, 75]. A Gaussian process is a natural generalization of the Gaussian probability
distribution. It generalizes a Gaussian distribution with a finite number of random
variables to a Gaussian process with an infinite number of random variables in the
surveillance region [53]. Gaussian process modeling enables us to efficiently pre-
dict physical values, such as temperature, salinity, pH, or biomass of harmful algal
blooms, at any point with a predicted uncertainty level. For instance, near-optimal
static sensor placements with a mutual information criterion in Gaussian processes
were proposed in [44, 76]. A distributed Kriged Kalman filter for spatial estimation
based on mobile sensor networks was developed in [45]. Multiagent systems that are
versatile for various tasks by exploiting predictive posterior statistics of Gaussian
processes were developed in [77, 78].
Gaussian process regression, based on the standard mean and covariance func-
tions, requires an inversion of a covariance matrix whose size grows as the number
of observations increases. The significant computational complexity in Gaussian
8 1 Introduction

process regression due to the growing number of observations (and hence the size of
covariance matrix) has been tackled in different ways [2, 79–83].
Unknown hyperparameters in the covariance function can be estimated by a max-
imum likelihood (ML) estimator or a maximum a posteriori (MAP) estimator and
then be used in the prediction as the true hyperparameters [1]. However, the point
estimate (ML or MAP estimate) itself needs to be identified using a sufficient amount
of measurements and it fails to incorporate the uncertainty in the estimated hyper-
parameters into the prediction in a Bayesian perspective. The advantage of a fully
Bayesian approach is that the uncertainty in the model parameters is incorporated in
the prediction [84]. In [85], Gaudard et al. presented a Bayesian method that uses
importance sampling for analyzing spatial data sampled from a Gaussian random
field whose covariance function was unknown. However, the solution often requires
Markov Chain Monte Carlo (MCMC) methods, which greatly increases the com-
putational complexity. In [46], an iterative prediction algorithm without resorting
to MCMC methods has been developed based on analytical closed-form solutions
from results in [85], by assuming that the covariance function of the spatiotemporal
Gaussian random field is known up to a constant.
There have been growing efforts to fit a computationally efficient Gaussian
Markov random field (GMRF) on a discrete lattice to a Gaussian random field on a
continuum space [4, 86–88]. It has been demonstrated that GMRFs with small neigh-
borhoods can approximate Gaussian fields surprisingly well [86]. This approximated
GMRF and its regression are very attractive for the resource-constrained mobile sen-
sor networks due to its computational efficiency and scalability [89] as compared to
the standard Gaussian process and its regression. Fast kriging of large datasets using
a GMRF as an approximation of a Gaussian field has been proposed in [88].

1.2 Contents in Chapters

A brief summary for each subsequent chapter is as follows. Chapter 2 gives an

introduction to Gaussian processes and Gaussian Markov random fields for general
domains as well as the space-time domain.
In Chap. 3, we develop covariance function learning algorithms for the sensing
agents to perform nonparametric prediction based on a properly adapted Gaussian
process for a given spatiotemporal phenomenon. By introducing a generalized covari-
ance function, we expand the class of Gaussian processes to include the anisotropic
spatiotemporal phenomena. Maximum likelihood (ML) optimization is used to esti-
mate hyperparameters for the associated covariance function as an empirical Bayes
method. The proposed optimal navigation strategy for autonomous vehicles will max-
imize the Fisher information [90], improving the quality of the estimated covariance
function.
In Chap. 4, we first present a theoretical foundation of Gaussian process regres-
sion with truncated observations. In particular, we show that the quality of prediction
based on truncated observations does not deteriorate much as compared to that of
Visit https://textbookfull.com
now to explore a rich
collection of eBooks, textbook
and enjoy exciting offers!
1.2 Contents in Chapters 9

prediction based on all cumulative data under certain conditions. The error bounds to
use truncated observations are analyzed for prediction at a single point of interest. A
way to select the temporal truncation size for spatiotemporal Gaussian processes is
also introduced. Inspired by the analysis, we then propose both centralized and dis-
tributed navigation strategies for mobile sensor networks to move in order to reduce
prediction error variances at points of interest. In particular, we demonstrate that
the distributed navigation strategy produces an emergent, swarming-like, collective
behavior to maintain communication connectivity among mobile sensing agents.
In Chap. 5, we formulate a fully Bayesian approach for spatiotemporal Gaussian
process regression under practical conditions such as measurement noise and
unknown hyperparameters (particularly, the bandwidths). Thus, multifactorial effects
of observations, measurement noise, and prior distributions of hyperparameters are
all correctly incorporated in the computed posterior predictive distribution. Using
discrete prior probabilities and compactly supported kernels, we provide a way to
design sequential Bayesian prediction algorithms that can be computed (without
using the Gibbs sampler) in constant time (i.e., O(1)) as the number of observations
increases. An adaptive sampling strategy for mobile sensors, using the maximum
a posteriori (MAP) estimation, has been proposed to minimize the prediction error
variances.
In Chap. 6, we propose a new class of Gaussian processes for resource-constrained
mobile sensor networks that build on a Gaussian Markov random field (GMRF) with
respect to a proximity graph over the surveillance region. The main advantages of
using this class of Gaussian processes over standard Gaussian processes defined by
mean and covariance functions are its numerical efficiency and scalability due to
its built-in GMRF and its capability of representing a wide range of nonstationary
physical processes. The formulas for predictive statistics are derived and a sequential
field prediction algorithm is provided for sequentially sampled observations. For a
special case using compactly supported weighting functions, we propose a distributed
algorithm to implement field prediction by correctly fusing all observations.
In Chap. 7, we consider a discretized spatial field that is modeled by a GMRF with
unknown hyperparameters. From a Bayesian perspective, we design a sequential pre-
diction algorithm to exactly compute the predictive inference of the random field.
The main advantages of the proposed algorithm are (1) the computational efficiency
due to the sparse structure of the precision matrix, and (2) the scalability as the num-
ber of measurements increases. Thus, the prediction algorithm correctly takes into
account the uncertainty in hyperparameters in a Bayesian way and also is scalable to
be usable for the mobile sensor networks with limited resources. An adaptive sam-
pling strategy is also designed for mobile sensing agents to find the most informative
locations in taking future measurements in order to minimize the prediction error
and the uncertainty in the estimated hyperparameters simultaneously.
Chapter 2
Preliminaries

2.1 Mathematical Notation

Standard notation is used throughout this book. Let R, R≥0 , R>0 , Z, Z≥0 , Z>0 denote
the sets of real numbers, nonnegative real numbers, positive real numbers, integers,
nonnegative integers, and positive integers, respectively.
Let E, Var, Corr, Cov denote the expectation, variance, correlation, and the covari-
ance operators, respectively.
Let AT ∈ R M×N be the transpose of a matrix A ∈ R N ×M . Let tr(A) and det(A)
denote the trace and the determinant of a matrix A ∈ R N ×N , respectively. Let
rowi (A) ∈ R M and col j (A) ∈ R N denote the ith row and the jth column of a matrix
A ∈ R N ×M , respectively.
The positive definiteness and the positive semi-definiteness of a square matrix A
are denoted by A  0 and A  0, respectively.
Let |x| denote the absolute value of a scalar x. Let x denote the standard
Euclidean norm (2-norm) of a vector x. The induced 2-norm of a matrix A is denoted
by A. Let x∞ denote the infinity norm of a vector x.
Let 1 denote the vector with all elements equal to one and I denote the identity
matrix with an appropriate size. Let ei be the standard basis vector of appropriate
size with 1 as its ith element and 0 on all other elements.
The symbol ⊗ denotes the Kronecker product. The symbol ◦ denotes the
Hadamard product (also known as the entry-wise product and the Schur product).
A random vector x, which is distributed by a normal distribution of mean μ and
covariance matrix C, is denoted by x ∼ N(μ, C). The corresponding probability
density function is denoted by N(x; μ, C).
The relative complement of a set A in a set B is denoted by B \ A := B ∩ Ac ,
where Ac is the complement of A. For a set A ∈ I, we define z A = {z i | i ∈ A}.
Let −A denote the set I \ A.

© The Author(s) 2016 11

Y. Xu et al., Bayesian Prediction and Adaptive Sampling Algorithms
for Mobile Sensor Networks, SpringerBriefs in Control,
Automation and Robotics, DOI 10.1007/978-3-319-21921-9_2
12 2 Preliminaries

An undirected graph G = (V, E) is a tuple consisting of a set of vertices V :=

{1, · · · , n} and a set of edges E ⊂ V × V. The neighbors of i ∈ V in G are denoted
by Ni := { j ∈ V | {i, j} ∈ E}.
Other notations will be explained in due course.

2.2 Physical Process Model

In this section, we review important notions for the Gaussian process which will be
used to model the physical phenomenon. In particular, we introduce a class of spa-
tiotemporal Gaussian process model with anisotropic covariance functions. The prop-
erties of Gaussian Markov random fields (GMRF) are also briefly reviewed.

2.2.1 Gaussian Process

A Gaussian process can be thought of a generalization of a Gaussian distribution over

a finite vector space to function space of infinite dimension. It is formally defined as
follows [53, 91]:

Definition 2.1 A Gaussian process (GP) is a collection of random variables, any

finite number of which have a consistent1 joint Gaussian distribution.

A Gaussian process, denoted by

 
z(x) ∼ GP μ(x), C(x, x ; θ ) (2.1)

is completely specified by its mean function μ(x) and covariance function C(x, x ; θ )
which are defined as

μ(x) = E [z(x)] ,
 
C(x, x ; θ ) = E (z(x) − μ(x)) (z(x ) − μ(x ))|θ .

Although not needed to be done, we take the mean function to be zero for notational
simplicity,2 i.e., μ(x) = 0. If the covariance function C(x, x ; θ ) is invariant with
respect to translations in the input space, i.e., C(x, x ; θ ) = C(x − x ; θ ), we call it
stationary. Furthermore, if the covariance functionis a function of only the distance
between the inputs, i.e., C(x, x ; θ ) = C(x − x  ; θ ), then it is called isotropic.

1 Itis also known as the marginalization property. It means simply that the random variables obey
the usual rules of marginalization, etc.
2 This is not a drastic limitation since the mean of the posterior process is not confined to zero [53].
2.2 Physical Process Model 13

Fig. 2.1 Realization of a

two-dimensional (D = 2)
Gaussian process with
σ 2f = 5, σ1 = 2.5, and
σ2 = 1.5.

In practice, a parametric family of functions is used instead of fixing the covariance

function [84]. One common choice of a stationary covariance function is
 D  2
 x − x
C(x, x ; θ ) = σ 2f exp − , (2.2)
=1
2σ2

where x is the th element of x ∈ R D . From (2.2), it can be easily seen that the
correlation between two inputs decreases as the distance between them increases.
This decreasing rate depends on the choice of the length scales {σ }. A very large
length scale means that the predictions would have little bearing on the correspond-
ing input which is then said to be insignificant. σ 2f gives the overall vertical scale
relative to the mean of the Gaussian process in the output space. These parame-
ters play the role of hyperparameters since they correspond to the hyperparame-
ters in neural networks and in the standard parametric model. Therefore, we define
θ = (σ 2f , σ1 , · · · , σ D )T ∈ R D+1 as the hyperparameter vector. A realization of a
Gaussian process that is numerically generated is shown in Fig. 2.1.

2.2.2 Spatiotemporal Gaussian Process

In this section, spatiotemporal Gaussian processes are of particular interest. Spa-

tiotemporal Gaussian processes are obtained as a special case of (2.1) by setting
x ⊂ R D × R≥0 , where R D is for spatial locations and R≥0 is the temporal domain.
A spatiotemporal Gaussian process can be written as

z(s, t) ∼ GP(μ(s, t), C(s, t, s , t ; θ )),

14 2 Preliminaries

(a) (b) (c)

Fig. 2.2 Realization of a spatiotemporal (D = 2) Gaussian process with σ 2f = 5, σ1 = 2.5,

σ2 = 1.5, and σt = 8 at a t = 1, b t = 5, and c t = 10.

where x = (sT , t)T ∈ R D × R≥0 . We consider the following generalized

anisotropic covariance function C(x, x ; θ ) with a hyperparameter vector θ :=
(σ 2f , σ1 , · · · , σ D , σt )T ∈ R D+2 :


D
(s − s )2 (t − t )2
C(x, x ; θ ) = σ 2f exp − exp − , (2.3)
=1
2σ2 2σt2

where s, s ∈ Q ⊂ R D , t, t ∈ R≥0 . {σ1 , · · · , σ D } and σt are kernel bandwidths

for space and time, respectively. (2.3) shows that points close in the measurement
space and time indices are strongly correlated and produce similar values. In reality,
the larger temporal distance two measurements are taken with, the less correlated
they become, which strongly supports our generalized covariance function in (2.3).
This may also justify the truncation (or windowing) of the observed time series data
to limit the size of the covariance matrix for reducing the computational cost. A
spatially isotropic version of the covariance function in (2.3) has been used in [36].
A realization of a spatiotemporal Gaussian process that is numerically generated is
shown in Fig. 2.2.

2.2.3 Gaussian Markov Random Field

The Gaussian Markov random field is formally defined as follows [92]:

Definition 2.2 (GMRF, [92, Definition 2.1]) A random vector z = (z 1 , · · · , z N )T ∈

R N is called a GMRF with respect to a graph G = (V, E) with mean μ and precision
matrix Q  0, if and only if its density has the form

|Q|1/2 1
π(z) = /2
exp − (z − μ)T Q(z − μ) ,
(2π ) N 2
Random documents with unrelated
content Scribd suggests to you:
 Treatment. The first consideration is to remove the causes of local
irritation and infection, give a clean sweet stall, with dry floor, and
allow no contact with putrid liquids, mud, cold water, melting snow
or other irritant. If exercise is needful to obviate stocking of the legs
give it on dry clean ground.
If inflammation runs high with fever and costiveness a laxative will
be valuable and it may be well to follow this in some cases with
cooling diuretics. When the animal has been on a heavy grain ration
this should be largely cut down in keeping with enforced idleness, or
restricted work. If on the other hand condition is low, and the
discharge profuse a more generous ration may be desirable.
Local treatment is essentially soothing and antiphlogistic, and in
view of the infection should be antiseptic. White lotion (acetate of
lead and sulphate of zinc of each ½ oz., water 1 quart) has been long
used with fair success, for although lead sulphate is thrown down, it
is in part freed again through contact with the exudate. It will be
materially improved by the addition of 1 dr. carbolic acid, creolin,
lysol or chloro-naphtholeum or by some other antiseptic. Lead
acetate alone with an antiseptic is an excellent substitute. In mild
cases the surface may be wet with the lotion several times a day:
while in severe ones the lotion may be applied on a bandage kept
constantly wet. When secretion is well established it may be better to
use dry applications, as calomel; salicyclic acid 10 parts, with iodine
10 parts; calomel and lamp black; carbolated oxide of zinc or burned
alum, salol, etc. After washing and drying the skin, dust this freely
even into the deepest wrinkles and cover with cotton and bandage.
Dress twice daily. For very fœtid cases, Robertson recommends zinc
chloride 1 oz. in 1 qt. water with the addition of glycerine and phenol,
and again a saturated solution of copper sulphate with carbolized
glycerine. For profuse secretion after the subsidence of active
inflammation Renal highly recommends wood tar with 5 to 10 drops
of sulphuric acid to the ounce. Under these circumstances a powder
of gloss starch 5 ozs. with iodine ½ oz. may be employed twice daily.
Or again, 1 oz. each of carbolic acid, tincture of iodine and glycerine
may be employed. When one agent seems to be losing effect, it is well
to change for another and never to neglect the regular dressing, until
full recovery has been secured.
 In case of grapes the actual cautery is the most efficient measure.
Heat a blacksmith’s fire shovel to a bright red and use this to cut
through the pedicles, a cool shovel being kept constantly beneath it
and in contact with the pedicles, so as to protect the adjacent skin
from injury by the radiated heat. The lower shovel must be dipped in
cold water at very frequent intervals to cool it and prevent
cauterization of the skin between the pedicles. This not only removes
the diseased and infected masses, but leaves the stumps of the
pedicles aseptic. Another method is to cut off the “grapes” and
staunch the blood with the actual cautery at a dull red heat. Still
another is to tie the pedicle of each excrescence separately so as to
cut off circulation and secure sloughing. This is, however, a long,
tiresome process, and entails prolonged contact with much infecting
dead tissue. After either method the parts must be dressed with
antiseptics, and dealt with generally like cases in which the
excrescences had not formed.
 CUTANEOUS HEMORRHAGE: BLOODY SWEAT:
HÆMATIDROSIS. HÆMATOPEDESIS.

Forms of cutaneous hæmorrhage; in specific diseases; in parasitism; in insect

bites; in congestions of sweat glands; in deranged innervation; in hæmophilia.
Section of sympathetic. Salt on sciatic. Hysteria. Sclerosis of cord. Inflammation.
Symptoms: drops, crusts. Hæmorrhagic nodules. Treatment: styptics, cold, ice,
snow, tannin, matico, iron chloride, alum, gelatine, atropine, ergot, lead acetate,
quinia. Gravitation.

The escape of the blood by the skin is seen in a variety of morbid

conditions, due it may be to profound changes in the blood and
capillary walls, as in petechial fever, anthrax, scorbutus, septicæmia,
swine erysipelas, etc., in which this is only a subsidiary phenomenon
of a general disorder:—to the presence of parasites (Filaria
hæmorrhagica,) in the skin:—to insect bites:—to violent congestions
implicating the sweat glands (bloody sweat):—or to deranged
innervation of the part as in cases of trauma of the sympathetic or
sciatic nerve, or disease of the nerve centres. It may further be a
manifestation of hæmophilia in which any slight lesion becomes the
occasion of persistent hæmorrhage.
Cases that appear in the course of specific contagious diseases and
those dependent on filaria will be considered under these headings,
and we may confine our attention here to the forms of sweating and
oozing of blood from independent causes. German writers draw
attention to its frequency in eastern horses, attributing it to the great
development of the vascular system especially of the skin, but its
comparative infrequency in the English racer and American trotter
would throw doubt on this doctrine. It may be questioned whether
the frequency of the disease in Oriental horses is not to be ascribed
rather to filariasis. This idea is not contradicted by the especial
prevalence of the bleeding in summer when the filaria is most active,
but when also the skin is the most vascular and its tissues most
relaxed.
Of nervous hæmorrhages we have the experimental examples of
Bouchard and Simon from section of the sympathetic nerve in
animals, also those of Glen and Mathieu from irritation of the sciatic
in dogs with common salt. In man the nervous causation has been
seen in hysteria, under profound nervous shock, in sclerosis of the
cord, and even as the result of auto-suggestion. This influence is
constantly operative in violent inflammations in which diapedesis
and minute hæmorrhages into the affected tissues are marked
phenomena, and under such a cause the gland ducts especially are
the seat of transudation. When the skin is abraded, cracked, or
blistered it occurs also on the surface of the exposed derma.
Symptoms. With active local congestion or inflammation the blood
usually oozes in drops from the surface, and drying concretes into
dark red crusts. In other instances, however, it drops from the
surface, or even flows, producing anæmia and even death. Into such
cases hæmophilia presumably enters. Hæmorrhagic swellings like
wheat kernels or beans also form in the skin.
Treatment. Apart from the contagious and parasitic diseases, and
scurvy, the general treatment will be styptic. Cold water, ice, snow, a
stream from a hose, solutions of tannin, matico, iron chloride or
sulphate, alum or gelatine may be employed. Internally the iron
salts, gelatine, atropine, ergot, lead acetate, or quinia may be given.
In hæmophilia the gelatine especially should be tried both locally and
generally. When it is possible, as in the case of the head, gravitation
should be availed of. Elsewhere a compress bandage may be used.
 ULCERATION. GANGRENE. BED SORES.

Causes: inflammation, exudation, obstructed circulation, lesions in trophic nerve

centres, sclerosis, toxins, ergot, caustics, freezing, gangrene, microbes, cryptogams,
spoiled fodder, white skins, buckwheat, insolation. Symptoms: inflammation,
molecular disintegration, dry sloughs. Treatment: camphorated spirit or vaseline,
antiseptics, phenol, salicylic acid, iodoform, iodine, creolin, lysol, tar, detach
sloughs.

In all cases in which the skin is violently inflamed, and particularly

when the seat of an abundant exudation or infiltration which blocks
circulation and retards nutrition, the tissues are especially liable to
death, molecular or by sloughing, and formation of bedsores. As a
general cause lesions of the trophic centres in the medulla and cord
must be accepted as a cause of the imperfect nutrition and lack of
vitality. This is seen in sclerosis of the cord, but may appear as the
result of poisoning of the myelon as well as the gangrenous tissues by
absorbed toxins. Again a common cause of circumscribed cutaneous
gangrene is the capillary contraction and obstruction of ergotism.
This usually involves all the tissues, soft and hard, at the distal end of
a member or organ, causing the separation of all at one common
level, but in less severe forms the skin only sloughs, in the form of
round or irregular masses, usually around the coronet, and the
resulting sores heal up under an appropriate diet. Cauterization and
freezing may be a further occasion of gangrene. Finally, the local
operation of the microbes of gangrene, determines both ulceration
and sloughing. Cryptogams on spoiled fodders (trefoil, lupins,
vetches, rusty gramineæ) are also charged with developing gangrene.
White skins or white patches on the skin are especially liable to
suffer as in cases of fagopyrism and “grease”. The action of the solar
rays in summer must therefore be accepted as a concurrent cause.
 Symptoms. The first symptoms are usually those of cutaneous
congestion or inflammation. Redness, swelling, pitting on pressure,
or tension, are accompanied or followed by vesicles, chaps or
erosions. The margins of the sores become thick and irregular, often
undermined, and they gradually increase by breaking down of tissue
in their depth or on their margin. In other cases patches of skin dry
or wither up, either in superficial layer or throughout its entire
thickness, and these dried extra vascular sloughs are gradually
detached by granulation beneath. The surrounding tumefaction is
always extensive and the sores may expose the deep seated
structures—tendons, ligaments, fascia, bones, joints—causing
widespread destruction.
Treatment. If the disease is due to capillary occlusion of nervous
origin, compresses with camphorated spirit, followed by
camphorated vaseline may be of advantage. If otherwise, antiseptics
will be in order: carbolated vaseline, salicylic acid cream, iodoformed
vaseline, a weak iodine ointment, creolin, or lysol in water, tar water.
When the dead tissues are partially separated the detachment may
be hastened with knife or scissors and the sores treated like a septic
sore.
 CUTANEOUS HYPERTROPHY. ELEPHANTIASIS.
PACHYDERMIA.

Chronic thickening of skin and lymph plexuses, horse hind limb after eczema,
grease, glanders, ox neck and head, knees, shoulder. Calcification. Treatment:
laxatives, diuretics, exercise, elastic bandage, friction, astringents, iodine.

Chronic thickening of the skin is most commonly seen in horses as

a sequel of lymphangitis in the hind limb, the engorgment of the
lymph plexus and thickening of its walls being associated with a
general productive inflammation and thickening of the derma until
the fetlock may be thirty inches or more in diameter. It may follow
eczema, grease or chronic glanders. In cattle a productive dermatitis
in the region of the head and neck, has led in the experience of the
author to a similar distention of the lymph vessels and morbid
thickening of the skin. The pads and calluses which form on the
knees of the camel and on other parts subject to friction, furnish
examples of hypertrophy of another kind. Again the thick dense
cutaneous plates on the shoulders of the old boar may serve to
illustrate a physiological hypertrophy. The writer has seen thickening
of the skin in the seat of an incision made in spaying the pig and the
deposition of earthy salts so as to form a distinct calcification.
Treatment is very unsatisfactory, yet something may be done by
laxatives, diuretics, regular exercise, an evenly applied elastic
bandage when in the stable, massage and the use of astringent and
dilute iodine ointments. It is much more important to prevent the
lesion by cutting short the morbid condition on which it depends.
When developed, attention is usually given to prevent its
advancement and to utilize the animal at slow work.
 ICHTHYOSIS. FISH-SKIN DISEASE.

This consists in a scaly formation of the epidermis which is also

formed in excess, and is supposed to be dependent on disordered
trophic innervation. In new born calves suffering in this way Van
Stettin found an excess of phosphate of lime in the epidermis. The
calves usually die in a few days.
 FURUNCULUS. BOIL.

Definition: pustule with necrotic core. Digital region. Microbes. Symptoms:

Prevention: antisepsis. Treatment: phenol, iodine, alcohol, boric or salicylic acid,
iron, bitters, calcium sulphide, sulphur, sodium sulphite.

A deep seated inflammation of the derma resulting in suppuration

with the formation of a central adherent necrotic slough or core.
Boils are not uncommon on the digital region of horses in winter;
or where the parts are exposed to street mud containing an
abundance of decomposing organic matter. They are unquestionably
due to pyogenic microbes, and have been largely associated with
staphylococcus pyogenes aureus. As they often come out in
successive crops, it may be assumed that the second focus is infected
from the first, or that there is a special susceptibility in the particular
animal system, or that both these factors contribute to the result.
Debility, and traumas contribute to bring about the infection.
Symptoms. A nodular, hot, and very painful swelling, implicating
the substance of the true skin, and surrounded by a hot, swollen
zone, progresses to suppuration in the centre, yet when it bursts, or
is opened, a core or small mass of necrotic, tough, fibrous tissue is
found to be firmly adherent in the center of the bottom of the sore.
Prevention. Must be sought in sustaining the general health and in
preserving the greatest cleanliness of stables and skin. Washing with
a weak antiseptic solution when returning from muddy streets may
be of use.
Treatment. When developing, the application of carbolic acid in
crystal or on the end of a glass rod may often relieve the pain and
destroy the microbes. Next day the part may be painted with tincture
of iodine. If already opened the phenol may still be applied and
followed after five minutes by alcohol to check the caustic action. Or
boric or salicylic acid may be applied. To counteract the
constitutional tendency, iron, bitters and other tonics and
antiseptics, calcium sulphide, sulphur and sodium sulphite may be
given.
 HYPERPLASIA OF THE SKIN.

Productive inflammations and tumors of the skin are common in

domestic animals but as they are essentially surgical lesions, they
will be better considered under that head. We may name papilloma,
warts, fibroma, melanoma, epithelioma, sarcoma, adenoma,
carcinoma, dermoid cysts. Parasitic and microbian diseases of the
skin will be found under their respective headings.
 ALOPECIA CONGENITA. CONGENITAL
BALDNESS.

Cases of this kind have been met with in foals and calves, which
were born entirely bald or with only a few thin delicate hairs
scattered over the surface. In a calf observed by the author, and
which lived for several months, the body was almost absolutely bald,
and the mouth remained edentulous, a coincidence which has been
observed in other cases. The teeth belong to the same class of
embryonic tissues as the epidermis, and a failure in the development
of the one is likely to entail a corresponding failure of the other.
Schneidemühl observed that the few hairs present in such cases were
especially delicate and brittle.
 ALOPECIA. POST PARTEM. ALOPECIA AREATA.

Normal shedding. Shedding out of time; laminitis, dropsy, exudative dermatitis,

acariasis, ringworm, traumas, folliculitis. Debility, excessive lactation, starvation,
petechial fever, spoiled fodder. Without apparent cause, alopecia areata, neurosis,
micrococci, disease of derma. Symptoms: general disorder has general shedding.
Local disorder extends from a centre. In horse with increased pigmentation.
Treatment: correct general causes, use hair stimulants, cantharides, kerosene, tar,
pilocarpin, mercuric chloride, cresol, iodine, balsam of Peru, silver nitrate. Arsenic.

Acquired baldness is recorded in horse, cattle, sheep and dog.

Causes. The simple shedding of hair occurs physiologically in
animals with the change of season, and if anything interferes with
the growth of the new hair a transient baldness may ensue. If such
shedding occurs from any cause at the wrong season, before the new
hair has started, the baldness may be accentuated. Thus shedding
may occur in some forms of indigestion, in laminitis, in dropsical
swelling of the limbs or ventral aspect of the body, in dermatitis with
an exudation which concretes around the hairs and raises them out
of their follicles, in mange, in demodectic acariasis, in circinate
ringworm, in traumas as on the elbows of dogs, etc., from lying on
them, and in inflammation of the hair follicles from a variety of
infections. It has been charged on general debility in excessive
lactation, in gestation, and in starvation, on poisons in the blood as
in petechial fever, and on musty or spoiled fodders in bad seasons, or
from low damp lands.
When in the absence of such appreciable causes it commences at
one or more points and gradually extends, and persists, it constitutes
alopecia areata. This has been attributed to a disorder of the
cutaneous nerves (tropho-neurosis), but the progressive advance of
the disease, without limitation to areas representing the distribution
of given cutaneous nerves, and the complete absence of other
derangement of nerve function, throw doubt on this conclusion.
Another doctrine attributes it to a microbe, but though micrococci
and other organisms have been found, they have not been proved to
be constant nor to be absolutely causative of the disease. Still another
theory holds that it is a disease of the derma and not of the hair at all,
the evulsion of the hair following the implication of the tissues
around the follicles.
Symptoms. The baldness dependent on a general disorder occurs
at once over an extended area. That of ringworm, acariasis, and of
the specific alopecia areata, advances gradually and often slowly
from a given point, until it may include a large area. Röll has seen it
extend from a few points to nearly the whole body of the horse in a
single year. In this, as in other cases in horse and dog, the baldness
was followed by a considerable increase of the pigmentation of the
skin.
Treatment. In cases that occur as the result of other diseases, the
rational treatment is to deal with these diseases, and then to
stimulate the growth of hair by some one or other of the known
stimulants (dilute tincture of cantharides, kerosene, tar water,
solution of pilocarpin hydrochlorate). In the more specific form, no
treatment has been very successful, yet the best results on the whole
appear to have come from local germicide applications. Mercuric
chloride in alcohol and water (1 ∶ 500); cresol 1, alcohol 20; tincture
of iodine reduced to half its strength by addition of alcohol; balsam
of Peru 1, alcohol 5; nitrate of silver 1, alcohol 15, serve as examples.
As general treatment arsenic has been employed, but with no very
encouraging results.
 TRICHORRHEXIS NODOSA. NODULAR
SWELLING AND SPLITTING OF HAIR.

Debility, ringworm, nodular hair. Hair bursts with brush like end. Cases in horse
like singed hair. Causes: infection probable, disturbed innervation, dry air,
impaired nutrition. Treatment: shave, oil, vaseline, petroleum, cantharides,
sulphur, tar, favor shedding coat.

In a variety of conditions the hairs split up and break, leading to

bristling or baldness. In debilitated conditions, when the hair is
badly nourished, in ringworm when it is invaded by a cryptogam,
and in the various nodular or ringed conditions of the hair this
brittleness and splitting appears. The term Trichorrhexis (rexis a
bursting forth) has been applied to one of these affections, in which
the hair swells and bursts into a number of filaments giving it a
broom-like termination. Trofimo describes this as affecting two-
thirds of the horses of an artillery brigade, and showing upon the
back, croup, mane and tail in isolated patches. Megnin, Trasbot and
Roy have observed similar cases. The affected hair a short distance
from the skin swells into a grayish white nodule and breaks across,
leaving a tuft of fine filaments. The patches look to the naked eye as
if singed, but when the hair is magnified the difference is easily made
out, as there may be several nodular enlargements on the same hair,
and the brushlike spread of filaments at the broken end is entirely
unlike the solid node on the end of a singed hair.
The causes of the affection have not been demonstrated, though
indications point to an infection which gradually extends from the
first point of attack. This extension is not limited to the area presided
over by particular nerves, and there is no other indication of
disordered innervation. Trasbot has seen the disease transmitted
from horse to horse by the use in common of combs, brushes and
rubbers, and in other cases experimentally by rubbing the scurf from
a diseased animal on the skin of a healthy one. From observations on
the human subject Montgomery looks on it as a result of extreme
dryness and brittleness of the hair, and charges dry climates and
seasons, and an abuse of soap in washing as conducive causes.
Debility and impaired nutrition of the hair may operate in the same
way. No specific microbe has been constantly found in these cases
and until such factor can be proved, it may be held that among the
etiological agents impaired nutrition and dry air are important.
In the way of treatment shaving of the affected parts and oily or
vaseline applications may be tried, together with such slight
cutaneous stimulants as petroleum, weak tincture of cantharides,
sulphur, tar ointment, etc. Roy noticed that cases that had lasted all
winter recovered after shedding of the coat. It might be assumed that
the evulsion of the infected hair, and the vascular stimulus necessary
to the new growth contributed to the result.
 CONSTITUTIONAL DISEASES.
RHEUMATISM.

Definition. Past views. Causes: heredity, age, cold, damp, cold climates, seasons,
exposure, buildings, cellars, night chills, weather vicissitudes, valleys, wading,
checked perspiration, lactic acid, metabolic products, acid phosphate of soda,
vegetable acids, neuropathic causation, infection, microbes, injuries, overwork.
Lesions: in joints, synovia, serosa, articular cartilage, fibro-cartilage, articular
lamella, bone, eburnation, ligaments, joints affected in horse and ox, blood
changes, albumen, fibrine, blood globules, pericardium, endocarditis, valvular
disease, myocarditis, embolisms in lungs, pleura, nervous lesions, digestive
system.

Definition. A constitutional, inflammatory affection, probably

toxic, tending to localization in the joints, muscles, tendons, fascia,
skin, heart and serous membranes and with a marked disposition to
shift from place to place.
The word is derived from rhein (γειν) to flow, and was originally
employed to indicate that an acrid humor, generated in the brain or
elsewhere escaped mainly by the nose and eyes as a catarrh. The idea
naturally followed that the retention of this humor caused
inflammation in the joints, muscles, heart or elsewhere. The
connection of these various conditions with exposure to cold, led to
the association of the name with the various internal inflammations
in which chilling appeared to have been a factor, until it was difficult
to limit it by any definite line. Finally infectious diseases implicating
the joints or muscles (influenza, contagious pneumonia, omphalitis,
gonorrhœal rheumatism), and diseases of metabolism (gout and
possibly rheumatoid arthritis) have added to the general confusion.
General Causes. Heredity. This has been more definitely traced
in man than in the lower animals, the line of family descent being
more easily followed in man. Children of rheumatic parents are more
obnoxious to the disease in the ratio of 5 to 1. In the absence of data
for animals, we may say that it is probable that the influence of
heredity will hold good for live stock in this matter. Whether the
disease results from faulty metabolism or from microbian invasion
or toxins, the hereditary defensive powers of the animal cell are likely
to be an important factor.
Age. Genuine rheumatism is not common at a very early age, but
in animals as in man, attacks especially those in the prime of life,
accustomed to vigorous exercise, free perspiration and succeeding
chills. There is as yet no direct evidence of an exhaustion of the
susceptibility of early maturity and of any consequent immunity of
the old.
Cold and damp. The association of rheumatism with cold and
damp has always been so notorious that no argument is needed to
enforce the truth. It is pre-eminently the disease of cold climates and
cold situations exposed to the north and east in Europe, or to the
north and west in North America on the Atlantic slope. It is very little
prevalent in the tropics. Buildings and yards that lack sunshine and
exposed storm-swept hillsides show cases most numerously. Dogs
kept in cold, damp sunless cellars often suffer. Cold, damp or frosty
nights start or aggravate cases in animals left out in the field, while
warm sunshine usually brings immediate improvement. A sudden
change of weather to cold and wet is the signal for aggravation of the
existing disease, though this may have set in before the change of
weather has been fully developed. Late autumn and early spring with
their sudden changes, their sleet and rain are especially hurtful and
particularly to chronic cases. The cold, raw ocean winds on the
Pacific coast contrasting with the warm sunshine is a marked
contributing cause. The currents of cold air that draw up through
shaded valleys, gullies and ravines expose their denizens to attacks.
Rheumatism is more rare among well housed and carefully treated
family horses, and less so in draught, dray and other hard working
horses that are made to cross streams and stand with feet and legs in
cold water while the body is drenched with perspiration. Dogs suffer
in the same way while hunting. Haycock attributes many attacks to
washing the legs with ice cold water when the animal has just
returned to the stable heated and perspiring. If then left undried and
unbandaged in a cold draught these conditions are still farther
aggravated. This was formerly attributed to the suppression of
perspiration, but it occurs in an animal which has just perspired
freely and is now in no special need of further elimination through
the skin. It appears to be due rather to the action of the cold on the
nerve endings, the susceptibility of which has been enhanced by the
free secretion and general relaxation.
But potent as cold is in precipitating or aggravating an attack, it
cannot be looked on as the sole or essential cause of rheumatism.
The great majority exposed to the cold escape. The animal which has
stood in the stall, or shed, is less likely to be attacked than the one at
work in which the heat production has been more active. Indeed an
immunity has been claimed for the Arctic regions, provided the
subject is not exerted so as to cause perspiration. One might readily
conclude that this apparent immunity, depended on the absence of
an essential germ, as is also claimed for catarrhs, yet Nansen after
his prolonged bath in the frigid waters seems to have had an attack of
rheumatism. Cold appears to be one of the most potent accessory
causes, but evidently not in itself the essential cause.
Cold undoubtedly affects profoundly the metabolism of the body
and especially when the vital powers have been reduced by severe
exertion, fatigue and perspiration. The same applies to many other
affections especially such as are associated with an infecting element,
which takes occasion of the debility caused by the cold to establish
itself in the tissues.
Lactic acid and other metabolic products. Prout long ago claimed
that rheumatism was caused by an excess of lactic acid in the system,
and this was supported by the fact that an exclusive diet of
buttermilk given in diabetes, is likely to produce acute articular
rheumatism. Again the production of rachitis in the young can be
traced in certain cases to excess of this acid. The strongly acid odor of
the sweat in certain rheumatic attacks is thought to support this
theory. The excess of lactic acid is variously explained by the
overwork of the muscles of which it is the normal product, and by the
imperfect oxidation of the muscular tissue and its product into lactic
acid (C3H4O(OH)2), instead of carbonic acid (CO(OH)2). But in spite
of the perfect theory, there is the fact that as a rule no special
increase of lactic acid can be found in blood or urine in acute
rheumatism and the improbability that an excess of this acid caused
by sudden excessive muscular waste could be kept up during a long
attack of acute much less of chronic rheumatism. Again the
comparative immunity of sucklings in which there is the greatest
opportunity for the production of lactic acid, would imply that that
alone cannot be accepted, as the one essential cause of the affection.
That its excess in the system will aggravate rheumatism, or even
produce it under favorable circumstances must be freely
acknowledged; also that acidity of the saliva is a marked feature of
rheumatism.
The acid phosphate of soda may be assigned a somewhat similar
rôle. In strongly predisposed subjects the ingestion of citric or other
organic acid will sometimes precipitate rheumatic articular pains.
Still other products, the result of imperfect oxidation or
metabolism must be allowed a place as probable factors in
rheumatism. The occurrence of gout in connection with the excess of
uric acid in the system is strongly suggestive of this, and the
frequency of muscular rheumatism in pampered, overfed dogs with
diseased livers and abnormal hepatic products, seems to give further
support. Even in man, the subject of acute rheumatism, often leads a
luxurious life and suffers from inactive or disordered liver, while in
man and animals alike, a low grade of health and imperfect
functional activity of important organs, are often precursors and
accompaniments of acute rheumatism.
Neuropathic theory. The doctrine of a neurotic cause originated by
Dr. Mitchell, of Philadelphia, has considerable basis in theory. The
primary chill tends to nervous derangement, which may readily
affect the overworked or already diseased and debilitated joints.
Similar peripheric disturbance of nutrition occurs in locomotor
ataxia in which the central nervous lesion is very evident. There is a
strong disposition in rheumatism to show a bilateral symmetry,
which points directly to a central nervous control. The frequent
violence of the pain, disproportionate to the moderate structural
changes, points in the same direction, and the free perspiration
present in some severe attacks, points alike to its origin in cutaneous
chill and to derangement of the centres presiding over perspiration.
It may be added that the development of rheumatic symptoms in the
advanced stages of infectious diseases, when the toxins are
accumulating in the system, suggests that they are the direct result of
a toxic action on the nerve centres. In man the influence of severe
nervous shock in developing and maintaining rheumatism is
recognized.
Theory of infection. This has been advocated by Cornil and Babes
and Friedberger and Fröhner. The former quote the frequent
presence of microörganisms and above all of micrococci in the
liquids of the affected joints, serosæ and valvular exudates and
ulcers, and even in the blood, and the occurrence of rheumatism in
the course of various septic affections. The latter quote Auer and
others as to the frequent supervention of articular rheumatism on
the retention of the afterbirth in the cow. The weak point in these
theories is the multiplicity and variety of the microbes found in the
exudates. Either we must accept the rheumatic lesions as an
occasional result of many microbes, which habitually act differently
on the system, or we must look upon them as mere accessory causes
or accidental complications. It might even be, that the invasion of
these microbes are made possible by the inflammation and debility
of the tissues, without being directly chargeable with the rheumatic
process. Even then there remains the possibility that a specific
microbe is present, which by reason of its infinitesimal size, or other
physical property, has as yet escaped recognition. If such specific
microbe is present, it manifestly requires a very special
predisposition, since it is not seen to advance from one individual to
another unless such favoring conditions are present. The symptoms
and lesions of rheumatism are not incompatible with the idea of such
an essential, specific germ, but as yet no such germ has been
satisfactorily demonstrated as present in all cases.
Local injuries. Rheumatism seems to attack by preference parts
that have already been debilitated by disease, a disposition which is
also true of distinctly microbian infections. A pre-existing sprain,
blow, bruise or an inflammation arising from any other cause seems
to practically invite the localization of the morbid process at that
point, and thus what was at first a purely traumatic lesion becomes
the seat of active and perhaps permanent rheumatism. Unusual
overwork and fatigue of given joints and muscles induce a similar
predisposition, and habitual overexertion, sprain, injury or
inflammation affecting repeatedly the rheumatic organ tends to fix
the process in chronic form.
 Articular lesions. These tend to concentrate as a form of
inflammation in the synovial membranes, but usually implicate all
the constituent structures of the joint, capsular and binding
ligaments, cartilage and fibro-cartilage, articular lamella and osseous
tissue. The synovial membrane may show only slight hyperæmia, or
in severe cases it may be congested, red, thickened or even
extensively infiltrated with a serogelatinoid liquid. These lesions are
most marked around the line of attachment on the articular surface
and in the synovial fringes. The synovia is usually in excess,
distending the capsule and is whitish, opaque, flocculent or more or
less deeply colored with red. It contains flakes of fibrine, leucocytes,
albumen and it may be red blood globules. Pus cells are usually
absent unless in distinctly infective cases. Coagula and false
membranes floating from or adherent to the solid tissues, may be
present in considerable amount and if these become organized they
tend to lay the foundation for future stiffening and rigidity. In and
beneath the serosa, cell proliferation may go on actively, especially in
the synovial fringes. The synovia is usually neutral or slightly
alkaline, though in rare instances it has proved to be acid.
The inflammation of the synovial membrane of the joint often
extends to those of the adjacent tendons, implicating at the same
time the tendons and their fibrous sheaths. Softening and rupture of
the tendons have been noted by different observers (perforans, Oger;
gastrocnemii, Trasbot; suspensory ligament, Olivier).
The articular cartilages and the fibro-cartilages may be the seat of
congestion, with ramified or uniform redness, and areas of swelling,
softening, absorption, erosion and ulceration, the ulcers varying in
size from a millet seed upward. The nuclei increase in size and the
cartilage cells multiply. In chronic forms calcification is not
uncommon.
The articular lamella and subjacent bone may show inflammation
with increased vascularity, softening and even ulceration. In the
chronic forms on the parts denuded of cartilage, the surface of the
bone may show the clear, polished condition known as eburnation.
The binding ligaments may show congestion, thickening,
exudation, softening, with increased rigidity, or in some cases
relaxation. Calcification is not uncommon in chronic cases. The soft
parts around the joints are often extensively infiltrated and swollen.
 In horses these lesions are specially common in the fetlock, but
occur also in the knee, elbow, shoulder, hips, stifle and hock. In
cattle they tend to attack the same parts with a preference perhaps
for the hocks and fetlocks.
Changes in the blood. The blood becomes profoundly altered, the
albumen reduced, the fibrinogenous elements increased (5 to 10
parts per 1000 instead of three), the red blood globules disappear (in
man 2,850,000 per cubic millimeter instead of 4,500,000), the
hæmoglobin and oxidizing power of the blood are reduced by about
50 per cent., the leucocytes are relatively increased, and coagulation
takes place with unusual firmness, a cupped surface and an excess of
buffy coat. No excess of urea, uric acid, nor lactic acid, has been
found, nor has acidity been found save in very rare cases. In rare and
severe cases petechiæ have appeared on the skin and mucosa.
Lesions of the heart. The implication of the fibrous structures of
the pericardium and heart and especially of the valvular structures is
a common lesion, and to be dreaded more than all others. In all
animals this tendency to cardiac lesion is well marked, but especially
in solipeds in which the great demands made on the heart during
rapid paces, heavy draught, jumping, etc., severely strain the cardiac
mechanism. In dogs there is not only the violent exertion and high
blood pressure, but also the great irritability of the nervous
mechanism presiding over the heart and the tendency to irregularity
and intermissions in the rhythm and palpitations even when the
organ is sound. There is every reason to conclude with Trasbot,
Megnin, Heu and Laurent that in many cases the heart is primarily
attacked, and that this heralds the articular rheumatism. In other
cases undoubtedly the cardiac affection is secondary, following the
articular attack.
Endocarditis is the most frequent, being found in a large
proportion of fatal cases, and thickening of the valves, wart-like
exudates and coagula are especially common. The clots may fill
nearly the entire ventricular cavity, or at times the auricular, and
show a preference for the left side, probably because of the more
vigorous systole and the higher blood tension. The clots as a rule are
firmly adherent to the diseased valve. Ulceration of the valve is rare
in rheumatism. Other parts of the ventricular endocardium may be
involved, becoming red, congested, rough or thickened, with
adherent blood clots.
Pericarditis is less common though it may exist on either or both
the cardiac and visceral folds. It is shown by vascularization,
thickening, fibrinous exudate, and serous effusion. Haycock found in
a horse a quart of reddish serous exudate with floating fibrous shreds
and false membranes. Pus has been found in exceptional cases
manifestly indicating a complex infection. Like endocarditis it may
precede, follow, or coincide with an articular attack (Leblanc,
Cadeac).
Myocarditis is usually seen as a complication and extension of
rheumatic inflammation of the visceral pericardium, or of the
ventricular endocardium. The muscular tissue may appear parboiled
and friable, and shows granular or fatty degeneration.
Pulmonic and pleuritic lesions. Embolism of the lungs and pleura
may occur from the transference of clots from the right heart, yet the
sequence is much more commonly an articular rheumatism
following infective disease of the lungs. Cadeac suggests that the
impaired nutrition in pneumonic and pleuritic animals predisposes
to the rheumatic arthritis, and again that the microbes of the
infectious chest affection, colonizing the joints and other synovial
sacs, determine the rheumatism. This last theory has the weakness of
identifying influenza and contagious pneumonia with articular
rheumatism, and is negatived by the experience that these two
pulmonary affections never develop de novo from simple
rheumatism. The rheumatism which follows influenza and
contagious pneumonia therefore must either be considered as a
pseudo-rheumatism, or a rheumatism occurring only concurrently
and accidently with the pulmonary affection. Apart from this,
pleurisy or even pneumonia occurs as a simple extension from a
rheumatic pericarditis.
The occurrence of the rheumatoid affection as a complication of
influenza and contagious pneumonia in the horse, usually appears
not earlier than 15 days after the outset of the pulmonary affection
and may be delayed, according to Palat, for 102 days. Palat who had
excellent opportunities for observation in army horses found that
about one in ten was the ratio in which the rheumatic affection
followed these pulmonary diseases.
 The pulmonary lesions in these affections are essentially those of
uncomplicated influenza, or contagious pneumonia.
Cerebro-Spinal Lesions. Nervous disorders are occasionally seen
in rheumatism in man (dullness, prostration, delirium, coma,
spasms) and traced in different cases to hyperthermia, congestion,
exudation, embolism and toxins. In a few cases in the horse, cerebral
complications have been observed. Olivier saw a horse with
lachrymation, closed eyelids and hot, tender forehead, which showed
at the necropsy articular inflammation, and sanguineous effusion in
the cranium, encephalon, frontal and maxillary sinuses and ethmoid
cells. Jacob records two cases (mare and horse) in which rheumatism
was complicated by meningo-encephalic congestion but without
necropsy as both recovered.
Digestive system. In man rheumatism has been exceptionally
preceded by pharyngitis, dysphagia, and diarrhœa. In the horse
Haycock has seen concurrent congestion of the pharyngeal mucosa,
Olivier congestion of the stomach and intestine, and Jacob diarrhœa
and abdominal pain, Leblanc and Palat record cases of peritonitis
accompanying articular rheumatism in the horse.
 SYMPTOMS OF ACUTE ARTICULAR
RHEUMATISM IN THE HORSE.

Chill, hyperthermia, lameness in fetlock, hock, knee, shoulder, stifle, hip, elbow,
tendons, sheaths, bursæ, bilateral, shifting, swelling, heat, tenderness, decubitus,
joint semi-flexed, pain on extension or flexion, signs of cardiac disease. Course:
rapid to recovery, improvements and relapses, metastasis, debility.

The attack is usually sudden and marked by a slight chill, shivering

or staring coat with hyperthermia and lameness. The temperature
may vary from normal to 107° or 108° F. in severe attacks. In many
cases the fever and lameness appear simultaneously, the former
being in ratio with the extent and severity of the latter, but not
infrequently the elevation of temperature precedes the articular
symptoms, and then it is to be considered as concurrent with the
internal lesions—cardiac, pericardial or otherwise. In other cases the
articular lesions and lameness precede by several days the
appearance of the fever. So far as we know the fever never antedates
the occurrence of some local lesion, external or internal.
The joints affected are very varied. The fetlock is the most
frequently attacked, but some of the other larger joints, the hock,
knee, shoulder, stifle, hip and elbow are often involved or exclusively
affected. The adjacent tendons and their synovial sheaths are very
often implicated; the attack is very prone to show a bilateral
symmetry, the same joints (right and left) on the corresponding fore
or hind limbs, suffering at once, or, as in the case of the fetlocks, all
four are simultaneously attacked. A joint that is weak by reason of
previous injury or disease is especially liable to suffer, and is then
less likely than joints that had been previously healthy to undergo
speedy improvement. When the symptoms wander from joint to joint
or from joint to muscle, or fascia or tendon, the disease in its earlier
seat seeming to undergo almost complete resolution, while it
advances with great intensity in the newly affected joint or part, the
occurrence is highly significant. These transitions often take place
with great rapidity. Thus the centre of suffering may shift from one
joint to another in an hour, (Megnin, Cadeac), or from a limb to the
loins in a quarter of an hour (Lewis). The fact that the inflammation
remains fixed in one or several joints, is not, however, proof of the
absence of rheumatism. A joint with a primary weakness or injury
may remain the seat of disease through even a chronic rheumatism.
The affected joint is usually swollen, hot and tender; the
tenderness being as a rule greatest where the capsular ligament is
pressed upon. These symptoms are very evident in joints that
approach the surface, and obscure in such as are thickly covered by
muscle (shoulder, hip). The swelling is soft, or tense and elastic
(especially over the synovial membrane), or œdematous and pitting
on pressure. On white skins sparsely covered by hair there is marked
redness and congestion, the veins standing out prominently and the
arteries above the joint pulsating strongly. When the tendinous
sheaths are involved, they stand out as elastic lines following the
course of the tendons, and with more or less pasty swelling adjacent.
In some cases, however, the swelling may be entirely absent, and
the trouble is to be located only by the local tenderness and pain
during motion.
Small, hard, pea like, subcutaneous nodosities were first noticed
by Floriep, in 1843, in rheumatism in man, and have been met with
in different cases in the horse. Rodet fils met with great numbers of
these nodules crepitating under pressure, in a horse that had
suffered for months from chronic rheumatism.
Acute pain on moving the affected joint or tendon is a most
characteristic symptom. The horse goes dead lame, walks on three
legs, or with great stiffness, and avoids as far as possible all flexion of
the joint. If left alone the animal stands stock still, never moving
from the place, or in the worst cases lies down and refuses to rise. If
compelled to walk his suffering is shown by hastened breathing,
dilated nostrils, anxious, pinched countenance and low plaintive
neighing. The affected joint is held semi-flexed, to relieve the
tension, the pastern is habitually more upright, and if the foot is
lifted and the affected joint bent or extended, the animal winces, or
resists, and tries to draw away the limb and groans. The movements
of the affected joint in walking or under manipulation, are often
attended by cracking which may be both felt and heard. It has been
variously attributed to lack of synovial lubrication (Cadeac), and
ulceration of the articular cartilage (Lafosse), yet it may occur from
the constrained position assumed, as in the case of a man attempting
to walk noiselessly on tip toe, and in other cases from the extension
of false membranes, or of rigid or contracted binding ligaments.
Rheumatism of the fetlock and sesamoid pulley, as the most
common form in solipeds, demands a special notice. The swelling of
one fetlock, of the two fore, of the two hind or of all four at once,
extends beyond the limits of both joint and sesamoid pulley and may
form a general engorgement or stocking which serves to hide the
synovial distension. Pressure however shows that while all is tender,
the extreme tenderness is referable to the joint, the synovial sheath
of the flexors, to the flexors or suspensory ligaments. As the general
swelling subsides the rounded or ovoid synovial distensions become
more patent. The swelling and tenderness may extend to the knee in
which case the synovial membrane of the carpal arch is especially
distended and tender from the carpus down, or in the hind limb the
synovial membrane of the tarsal arch is distended showing in this
case on the inner and outer sides in front of the calcis, and not
infrequently implicating the summit of that bone as a capped hock.
In rheumatism of the shoulder the coraco-radial tendon and pulley
may be involved, causing a diffuse swelling on the point of the
shoulder. If the hip is the seat of disease the median gluteal may
suffer, causing an indefinite swelling over the joint. If the stifle is
affected the patellar capsule suffers and not infrequently the tendon
and pulley of the flexor metatarsi are involved.
The most important internal complications, pericarditis and
endocarditis, are manifested by their usual symptoms, sharp,
variable, irregular, unequal or intermittent pulse, blowing murmur
usually with the first heart sound, oppressed breathing, and it may
be muffled heart sounds, or dropsies.
Course. This is exceedingly uncertain. Some cases make a rapid
progress to complete convalescence; others make partial
improvements interrupted by relapses; others have the morbid
process subside in great measure in one joint or organ only to
reappear in full force in another; others leave complications on the
part of the heart especially and are rendered permanently useless.
Even should the heart escape, the health often suffers so much in
connection with the destruction of the red globules, the malnutrition,
and the local disease, swellings and distortions of the joints that a
perfect recovery seems distant and problematical.
 SYMPTOMS OF CHRONIC ARTICULAR
RHEUMATISM IN THE HORSE.

Larger joints, muscles, heart, false membranes, indurations, thickenings,

calcifications, remittent, weather changes, cold, damp beds, winds, open windows
or doors, draughts, cold sponging, clipping. Diagnosis: lameness variable, shifting,
electric and meteoric storms.

Chronic rheumatism may be a sequel of the acute, or it may occur

from the same causes acting with lessened force, or on a less
susceptible animal. It tends to attack the larger joints especially,
though it may implicate the muscles as well. Coincident affection of
the heart is less common than in the acute, and when it does arise
seems to advance slowly. It is liable to cause permanent distensions
of the affected joint capsules, as well as false membranes, articular
abrasions, degenerations and ulcerations and less frequently bony
enlargements and calcifications, the latter implicating the soft tissues
in the vicinity.
The attendant lameness is liable to be remittent or intermittent,
subsiding in warm buildings and during genial, clear sunny weather,
and relapsing in connection with cold, raw nights and mornings,
exposure in the dew or rain, and before and during great changes of
weather. Cold, damp beds, chilling draughts between open doors or
windows, washing with iced water, sudden intense cooling of the
body after perspiration, clipping during cold weather, any cause of
sudden rigor, when followed by stiffness, lameness and articular
swelling, serves to identify the latter as rheumatic. Even the warmth
induced by judicious exercise, may cause improvement, so that a
horse, starting out stiff or lame, may drive out of it after going a mile
or two. The formation of subcutaneous nodules, though rare, appears
to be more frequent than in acute rheumatism.