Lecture 1:

Course Introduction

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 1

 Welcome
- This is the 2nd offering of BIODS 220 Artificial Intelligence in Healthcare (CS
271, BIOMEDIN 220)
- What we hope you will get out of this course:
1. Broad knowledge of opportunities for AI in healthcare
2. Fluency in cutting edge deep learning algorithms, and practical ability to
develop models for diverse types of healthcare data
3. Understanding of real-world considerations and challenges for deploying
AI algorithms in healthcare

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 2

 Today’s agenda
- A brief overview of AI in healthcare
- Course logistics

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 3

 AI in healthcare: a rapidly exploding field

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 4

 AI in healthcare: a rapidly exploding field

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 5

 AI in healthcare: a rapidly exploding field

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 6

 A journey back in time… brief history of modern AI
1956: Birth of AI as a modern
research discipline

John McCarthy

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 7

 Early progress in the late 50s and 60s

Perceptron model: Rosenblatt, 1958 ELIZA chatbot: Weizenbaum, 1966

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 8

 Progress and excitement in the late 50s and 60s

Capabilities of early projects soon reached

limitations they could not surpass.
Enthusiasm for AI dwindled in the 70s.

Perceptron model: Rosenblatt, 1958 ELIZA chatbot: Weizenbaum, 1966

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 9

 Resurgence in the 80s

...

Expert systems, 1970s and 80s.

Feigenbaum, etc. Backpropagation. Rumelhart, 1986.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 10

 First appearances of modern neural networks

LeCun, 1990s.

Schmidhuber, 1997.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 11

 First appearances of modern neural networks

But limitations reached again… -> AI

winter of the 1990s and 2000s
LeCun, 1990s.

Schmidhuber, 1997.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 12

 2012: Deep learning breakthrough

Krizhevsky et al. 2012. 8-layer “AlexNet”.

ImageNet Visual Recognition Challenge results.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 13

 Convergence of key ingredients of deep learning
Algorithms Compute

Data

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 14

 2015: Very deep convnets and challenging vision tasks

He et al. 2015. ResNet.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 15

 2018: Breakthroughs in deep learning for natural
language processing (sequences)
Transformer architectures and pre-training -> fine-tuning. State-of-the-art on 11 NLP benchmarks.

Devlin et al. 2018. BERT.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 16

 Deep learning for healthcare: the rise of medical data

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 17

 Deep learning for healthcare: the rise of medical data

***All zoom polls in this class are anonymous***

Zoom poll: When was x-ray invented?

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 18

 Deep learning for healthcare: the rise of medical data

X-rays (invented 1895).

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 19

 Deep learning for healthcare: the rise of medical data

X-rays (invented 1895).

Zoom poll: When was CT invented?

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 20

 Deep learning for healthcare: the rise of medical data

X-rays (invented 1895). CT (invented 1972).

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 21

 Deep learning for healthcare: the rise of medical data

X-rays (invented 1895). CT (invented 1972).

Zoom poll: When was MRI invented?

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 22

 Deep learning for healthcare: the rise of medical data

X-rays (invented 1895). CT (invented 1972). MRI (invented 1977).

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 23

 Deep learning for healthcare: the rise of medical data

Q: What are other examples of medical data?

(Raise hand or type in chat box)

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 24

 Electronic health records -- making patient data available

Clinical notes
Imaging data

1960s: invention
1980s: increased effort
2009: 51% adoption, HITECH Act
2017: 98% adoption Lab results
Patient measurements

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 25

 Genomics data

1953 - Watson and Crick 1977 - Fred Sanger 1990 - 2003: Human Genome
discover double helix sequences first full genome Project sequences full human
structures of DNA of a virus genome

2003: ENCODE project launched to 1000 Genomes Project: UK100,000 Genomes

identify and characterize genes in 2008 - 2015 Project: 2012 - 2018
human genome

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 26

 Wearables and other sensor data

First iPhone: 2007 Fitbit: 2009 Apple Watch: 2014

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 27

 AI in healthcare: biomedical image interpretation

Wu et al. 2019 Liu et al. 2017

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 28

 AI in healthcare: clinical event prediction

Harutyunyan et al. 2019

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 29

 AI in healthcare: genomic analysis

Zhou et al. 2015 Poplin et al. 2016

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 30

AI in healthcare: drug discovery and drug interaction prediction

Torng et al. 2019 Ryu et al. 2018

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 31

AI in healthcare: intelligent healthcare spaces and environments

Yeung et al. 2019 robinhealthcare.com

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 32

 AI in healthcare: mobile health and wearables

Tariq et al. 2018 Menictas et al. 2019

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 33

 AI in healthcare: recent applications for COVID-19

Q: What are ways AI could be used to help

tackle the COVID-19 crisis?

(Raise hand or type in chat box)

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 34

 AI in healthcare: recent applications for COVID-19

Harmon et al. 2020

Yan et al. 2020 Jumper et al. 2020

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 35

 The promise is great…
but many open challenges
in deployment as well
 Uncertainty and AI / human collaboration

Rosenberg et al. 2018

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 37

Bias and fairness

Obermeyer et al. 2019

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 38

Privacy and security

Price et al. 2019 Figure: https://news.developer.nvidia.com/first-privacy-preserving-federated-learning-system/

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 39

 In this class
1st part: developing DL algs for health data 2nd part: deploying AI for health

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 40

 Speed Breakouts
Get to know your classmates

Can opt-out by not pressing join

Private message TAs if you have any questions

- 2x 4-minute breakouts (4 students each)

- Name, program, year
- What’s one thing you hope to get out of this class?
- What kind of healthcare tasks or data are you most interested in?

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 41

 Who are your classmates?

Zoom Polls:
- PhD, MS, Undergrad, Other
- Department / major
- Where in the world are you?
- What type of medical data are you currently
most interested in?

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 42

 Course Logistics
Lectures: MW 1-2:20pm, on Zoom

- Zoom links available through Canvas

- Lectures will be recorded and posted afterwards on Canvas

A few review sessions (e.g., Tensorflow, Project Partner Finding): select Fridays
1-2:20pm, Zoom

- First one will be a session for finding project partners, Fri 9/18
- Stay tuned for announcements of future Friday reviews

Course materials will be hosted on website: http://biods220.stanford.edu/

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 43

 Teaching team

Office hours will start week 2

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 44

 Prerequisites
1. Proficiency in Python, or significant experience with a different programming language
and ability to self-learn. Python will be used for homework assignments and the course
project.
2. Basic familiarity with college calculus (e.g. Math 19 or 41, comfortable taking
derivatives), linear algebra (e.g. Math 51 or EE 103 / CME 103, comfortable with
common matrix vector operations and notation), and probability and statistics (e.g. CME
106 or CS 109, comfortable with common probability distributions, mean, standard
deviation, etc).
3. Familiarity with machine learning, e.g. comfortable with the framework of machine
learning and experience training a machine learning model.
4. Familiarity with deep learning is highly recommended, e.g. prior experience training a
deep learning model.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 45

 Piazza
- Will be used for all course communications.
- Sign up using link on course website “logistics” tab
- If it is a personal matter (e.g., OAE), please make a private post to the
teaching team or instructor.
- Note: we will not be using canvas in this course, with the exception of zoom
lecture links. Communications will be through Piazza, and grading will be
through gradescope.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 46

 Grading
- Sign up for Gradescope through the “logistics tab”
- Breakdown:
- Assignment 1: 20%
- Assignment 2: 20%
- Assignment 3: 20%
- Course project: 40%

** Note: No in-class midterm this year due to COVID-19 remote learning. Instead,
the class will focus on supporting the opportunity to pursue an in-depth
AI+healthcare project of your choice.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 47

 Assignments
- Main objective to build conceptual and practical foundations in using deep learning for
biomedical data
- A0 (Data access prerequisites): Out Tue 9/15, due Tue 9/22. No grade, but required by due
date to gain data access required for later assignments.
- A1 (Medical images): Out Tue 9/22, due Tue 10/6.
- A2 (EHR and text data): Out Tue 10/6, due Tue 10/20.
- A3 (Genomics): Out Tue 10/20, due Tue 11/3.

- In this class, all deadlines refer to 11:59pm PST on the stated day.
- A limited amount of Google cloud credits will be provided for the assignments. Should be
sufficient, but use wisely.
- Collaboration policy: please read on course website. Study groups are allowed, but each
student must produce independent assignment and write names of group on assignment.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 48

 Project
- Opportunity to gain in-depth experience developing an AI-based approach to a healthcare
problem.
- Worth 40% of grade. Can work in groups of 1-3. (Grades will be calibrated by group size)
- Since large part of course is focused on deep learning, must involve implementation and
training of at least one deep learning model on health data. Otherwise, significant flexibility in
technical component (compare DL vs. non-DL models, analyze DL model in depth, novel DL
architectures, etc.).
- Can use any health-related data of your choice. Options include public datasets and
challenges (e.g., start from a published paper!), ongoing projects at Stanford (if applicable),
projects suggestions from Stanford Medical School, etc.
- Will release detailed project guidelines and suggestions, and discuss in lecture on Mon 9/20.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 49

 Project (cont.)
- Graded components:
- Proposal: Due Fri 10/9.
- Milestone: Due Fri 10/30.
- Project milestone presentations (4-5 min): During Mon 11/2 class time.
- TA project advising sessions: Sign-up by Fri 11/6.
- Final project presentations (4-5 min): During Wed 11/18 class time.
- Final report due: Fri 11/20.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 50

 COVID-19 / asynchronous class participation
- Live attendance for classes highly recommended if possible for best learning
experience
- We understand that COVID-19 causes many unexpected challenges; will do
our best to support students who need asynchronous participation, e.g. due to
different time zones
- Lectures will be recorded and posted afterwards on canvas
- Project milestone and final presentations will have option of submitting the presentation in
video form beforehand, to be played during class
- Stay tuned for details on asynchronous accommodations
- Please feel free to reach out to the course staff at any time if you are
having challenges. We know these are difficult times and want to do our
best to help and support you.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 51

 Late days
- Can be used on A1, A2, A3, project proposal, project milestone report.
- Cannot be used on project milestone presentation, project final presentation,
or final project report.
- 6 late days total, 2 max for any assignment.
- Grades will be deducted by 25% for each additional late day.

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 52

 Course schedule

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 53

 Next time
- Review of deep learning fundamentals

Serena Yeung BIODS 220: AI in Healthcare Lecture 1 - 54