MAASTRO Knowledge Engineering: The Fun(ction) of Medical Physics in Cancer Care
2 likes1,259 views
The document discusses the challenges and advancements in medical physics, highlighting the need for personalized treatment in healthcare due to the complexity of disease factors. It explores the concept of knowledge engineering in developing decision support systems to optimize patient care based on extensive data analysis. The document emphasizes the importance of addressing real-world clinical problems and engaging in translational research to improve healthcare outcomes.
![22
MAASTRO
©2013
Rapid Learning
In [..] rapid-learning [..] data routinely
generated through patient care and
clinical research feed into an ever-
growing [..] set of coordinated
databases.
•Abernethy, J Clin Oncol 2010;28:4268
•[..] rapid learning [..] where we can
learn from each patient to guide
practice, is [..] crucial to guide
rational health policy and to contain
costs [..].
•Lancet Oncol 2011;12:933](https://image.slidesharecdn.com/2013-08-06aapmerikroelofs96ppi-130828032103-phpapp01/85/MAASTRO-Knowledge-Engineering-The-Fun-ction-of-Medical-Physics-in-Cancer-Care-22-320.jpg)
![33
MAASTRO
©2013
Problems, problems…
Barriers
Administrative
(time to capture,
time to curate)
Political
(value,
authorship)
Ethical
(privacy)
Technical
[..] the problem is not really
technical […]. Rather, the
problems are ethical,
political, and administrative.
Lancet Oncol 2011;12:933
Solutions: Distributed learning
from federated databases](https://image.slidesharecdn.com/2013-08-06aapmerikroelofs96ppi-130828032103-phpapp01/85/MAASTRO-Knowledge-Engineering-The-Fun-ction-of-Medical-Physics-in-Cancer-Care-33-320.jpg)
![42
MAASTRO
©2013
Data>Model>Decision Support
•Prediction Models: Revolutionary in Principle, But Do They Do
More Good Than Harm?
•“we are drowning in prediction models [..] more than 100
prediction models on prostate cancer alone”
•“currently [..] a large number of models [..] are not
independently validated at all”
•J Clin Oncol 2011;29:2951](https://image.slidesharecdn.com/2013-08-06aapmerikroelofs96ppi-130828032103-phpapp01/85/MAASTRO-Knowledge-Engineering-The-Fun-ction-of-Medical-Physics-in-Cancer-Care-42-320.jpg)
MAASTRO Knowledge Engineering: The Fun(ction) of Medical Physics in Cancer Care
- 1. MAASTRO©2013 The fun(ction) of Medical Physics… Erik Roelofs, MSc. M e d i c a l P h ys i c ist , Info. / Prog. Manager A A P M – 2 0 1 3 - 0 8 - 0 6 MAASTRO Knowledge Engineering
- 2. 2 MAASTRO ©2013 Overview • Background • What we want to do • Where it became “Knowledge Engineering” • Conclusion / tips
- 3. 3 MAASTRO ©2013 Setting the stage (my background) • Born 1971 • Education – Bachelor Physics – Master “Mechatronics” linkedin://roelofserik
- 4. 4 MAASTRO ©2013 Setting the stage (my background) • Born 1971 • Education – Bachelor Physics – Master “Mechatronics” • System engineer (3 yr) – Integrator: cross-border, optics, electronics, mechanics – Informatics: control systems architecture & design
- 5. 5 MAASTRO ©2013 Setting the stage (my background) • Education – Bachelor Physics – Master “Mechatronics” • System engineer (3 yr) – Integrator: cross-border, opto-, electro-mechanics, system control & architecture • MAASTRO clinic (2003) – Project leader (1 yr) – Med.Phys. trainee (4yr) – QMP & PhD candidate (5? yr) – Information / program manager (1 yr) • Substantial time for side-tracks – some to optimize treatment quality, workflow, proceure, etc. – others to start a new research line
- 10. 10 MAASTRO ©2013 Overview • Background • What we want to do • Where it became “Knowledge Engineering” • Conclusion / tips
- 12. 12 MAASTRO ©2013 Life Sciences and Healthcare are converging Predictive, Preventive, Personalized and Participatory Healthcare HEALTHCARELIFE SCIENCES “Trial and Error” Healthcare “Evidence Based” Healthcare “Precision” Healthcare Blockbusters and mass- production of novel drugs Targeted Therapies Increased regulation and efficacy standards Analytics LIFE SCIENCES HEALTHCARE DNA chemistry and advanced technology “Managed” Healthcare Paper based Records Electronic Data Capture Pharmacovigilance and Risk Mgmt Safety at Point of Care Electronic Medical Records Paper based Systems Personalized Healthcare Patient Care and Disease Mgmt Translational Med © 2010 Oracle and/or its affiliates. All rights reserved.
- 13. 13 MAASTRO ©2013 The Problem is in the patient • Remember the girl? – Young smoker – Numerous co-founding factors! – It’s not just a tumor to treat… – but a complex, cross-domain system! • Need for individualized treatment
- 14. 14 MAASTRO ©2013 European Journal of Cancer, Volume 48, Issue 4, March 2012 Towards individualized treatment
- 15. 15 MAASTRO ©2013 European Journal of Cancer, Volume 48, Issue 4, March 2012 Entering the OMICS era…
- 16. 16 MAASTRO ©2013 European Journal of Cancer, Volume 48, Issue 4, March 2012 Entering the OMICS era…
- 17. 17 MAASTRO ©2013 The doctor is drowning • Explosion of data • Explosion of decisions • Explosion of „evidence‟* • 3 % in trials, bias *2010: 1574 & 1354 articles on lung cancer & radiotherapy = 7.5 per day Half-life of knowledge estimated at 7 years J Clin Oncol 2010;28:4268 JMI 2012 Friedman, Rigby
- 18. 18 MAASTRO ©2013 The problem: Individualized Medicine It is unethical to ask any person (even a doctor) to predict the best treatment… You fool!
- 19. 19 MAASTRO ©2013 Prediction by MDs? Two year survival • Non Small Cell Lung Cancer • 2 year survival • 30 patients • 8 MDs • Retrospective • AUC: 0.57 Cary Oberije et al.
- 20. 20 MAASTRO ©2013 Results: Models always significantly better then RO Death at 2 years Dyspnea Dysphagia RO’s models
- 21. 21 MAASTRO ©2013 Decision support in Radiation Oncology
- 22. 22 MAASTRO ©2013 Rapid Learning In [..] rapid-learning [..] data routinely generated through patient care and clinical research feed into an ever- growing [..] set of coordinated databases. •Abernethy, J Clin Oncol 2010;28:4268 •[..] rapid learning [..] where we can learn from each patient to guide practice, is [..] crucial to guide rational health policy and to contain costs [..]. •Lancet Oncol 2011;12:933
- 23. 23 MAASTRO ©2013 Computer Assisted Theragnostic (CAT): Predictive model allowing treatment individualization - A holistic approach
- 24. 24 MAASTRO ©2013 Treatment administered Real Outcome (Complications, S urvival) Prospective gathering of pre-treatment data (+CI) Feed-back Loop Computer Assisted Theragnostic (CAT): Predictive model allowing treatment individualization - A holistic approach - Survival: National Database (GBA) - Complications: Module EHR, (e)Questionnaire CTC like GP- Patients-Long specialist Biological Data Clinical Data Image Data Data-based & Knowledge based models: Probability of Survival & Complications (+ CI) for treatment x, y, z…
- 25. 25 MAASTRO ©2013 The five components of Radiation Oncology Clinic Biology Physics Molecular Imaging Computer science Radiation Oncology
- 26. 26 MAASTRO ©2013 Overview • Background • What we want to do • Where it became “Knowledge Engineering” • Conclusion / tips
- 27. 27 MAASTRO ©2013 CAT ~ 2005 = MAASTRO Knowledge Engineering Build Decision Support Systems to individualize patient care by using machine learning to extract multifactorial personalized prediction models from existing databases containing all data on all patients that are validated in external datasets Theme 2 Learning (5%) Theme 1 Data (95%)
- 29. 29 MAASTRO ©2013 Centralized Data for Research Hospital 1 Research System data domains clinical imaging biobanking integrated data e.g. tranSMART e.g. caTissue NBIA Open Clinica HIS PACS LIS Hospital 2 HIS PACS LIS
- 31. 31 MAASTRO ©2013 MISTIR framework : www.mistir.info Multicenter In Silico Trials In Radiotherapy ROCOCO: Photon, Proton, C-ion Comparison project Roelofs, et al. Radiother. Oncol. Dec 2010 Reporters Principal Investigators Data centre Participants Secure DB TP Initialisation Collaboration Protocol MTA Collaboration Protocol MTA Reporting Dummy Run** Preparation Analysis Institute n Perform statistics Biological modelling Derive parameters Generate DVH Perform statistics Biological modelling Derive parameters Generate DVH Database (DB) CT/PET Calibration DICOM datasets* Database (DB) CT/PET Calibration DICOM datasets* Institute 1 QA Limited nr. of slices Contour names Orientation, offsets Grid spacing Limited nr. of slices Contour names Orientation, offsets Grid spacing Roelofs, et al. J. Thorac. Onc., Jan 2012 Van der Laan, et al. Acta Oncol. Apr 2013
- 32. 32 MAASTRO ©2013 ROCOCO network Europe (13 partners): •Aken •Amsterdam (NKI) •Vienna •Darmstadt •Gent •Groningen United States (3 partners): •Boston •Pennsylvania •Madison Wisconstin Japan: •Chiba •Hasselt •Heidelberg •Luik •Lyon •Maastricht (PI) •Paris •Villigen Switserland
- 33. 33 MAASTRO ©2013 Problems, problems… Barriers Administrative (time to capture, time to curate) Political (value, authorship) Ethical (privacy) Technical [..] the problem is not really technical […]. Rather, the problems are ethical, political, and administrative. Lancet Oncol 2011;12:933 Solutions: Distributed learning from federated databases
- 34. 34 MAASTRO ©2013 Data extraction system - Federated
- 35. 35 MAASTRO ©2013 Federated Data for Research Hospital 1 integrated data e.g. euroCAT HIS PACS LIS Hospital 2 HIS PACS LIS integrated data e.g. euroCAT
- 36. Distributed Learning Architecture Update Model Learn Model from Local Data Central Server Model Server RTOG Send Model Parameters Final Model Created Learn Model from Local Data Learn Model from Local Data Model Server MAASTRO Model Server Roma Send Model Parameters Send Model Parameters Send Average Consensus Model Send Average Consensus Model Send Average Consensus Model Only aggregate data is exchanged between the Central Server and the local Servers
- 37. 37 MAASTRO ©2013 The realiztion of the dream: euroCAT (see www.eurocat.info), ameriCAT, duCAT Active or funded CAT partners (10) Prospective centers (4) 2 5
- 38. 38 MAASTRO ©2013 Data>Model>Decision Support 1. Modeling “Learn a model from data” 2. Validation “Estimate model performance” 3. Decision Support “Impact of the model on clinical practice”
- 39. 39 MAASTRO ©2013 Dehing-Oberije, IJROBP 2009;74:355 Learn a model from data •Training cohort – 322 patients (MAASTRO) •Clinical variables •Support Vector Machines •Nomogram
- 40. 40 MAASTRO ©2013 Estimate model performance (survival) • INDEPENDENT Validation cohort – 36 patients (Leuven) – 65 patients (Ghent) • Discrimination, Calibra tion, Reclassification • AUC 0.75 Dehing-Oberije (MAASTRO), IJROBP 2009;74:355
- 41. 41 MAASTRO ©2013 Power of DSS (compare to TNM) Stage IIIA 10 (14%) Stage IIIB 13 (19%) T4 12 (17%)
- 42. 42 MAASTRO ©2013 Data>Model>Decision Support •Prediction Models: Revolutionary in Principle, But Do They Do More Good Than Harm? •“we are drowning in prediction models [..] more than 100 prediction models on prostate cancer alone” •“currently [..] a large number of models [..] are not independently validated at all” •J Clin Oncol 2011;29:2951
- 43. 43 MAASTRO ©2013 Models built & validated : PredictCancer • Lung cancer – Survival – Lung dyspnea – Lung dysphagia • Rectal cancer – Tumor response – Local recurrences – Distant metastases – Overall survival • Larynx cancer – Local recurrences – Overall survival www.predictcancer.org
- 44. 44 MAASTRO ©2013 Cost-effectiveness at www.predictcancer.org The increased effectiveness of IMPT does not seem to outweigh the higher costs for all head-and- neck cancer patients. However, when assuming equal survival among both modalities, there seems to be value in identifying those patients for whom IMPT is cost-effective.
- 45. 45 MAASTRO ©2013 No No Proton therapy reimbursement decision tree for the Netherlands Treatment with PROTONS Yes Yes Isthisamodel basedindication? Create“stateoftheart” PHOTONandPROTON treatmentplans No NoYes Clinicallyrelevant benefitexpected? Comparebothplans accordingtoprotocol Treatment with PHOTONS Integratecomplication probabilitymodelling Yes Isthisdiseasea standardindication? Evidentdosimetric benefitwithprotons? PRODECIS : clinical grade decision support system • Up-front dosimetric and complication rate comparison • Referrers needs answers fast • Reuse as much information (data) as possible
- 46. 46 MAASTRO ©2013 PRODECIS : clinical grade decision support system
- 47. 47 MAASTRO ©2013 Overview • Background • What we want to do • Where it became “Knowledge Engineering” • Conclusion / tips
- 48. 48 MAASTRO ©2013 You! Yes, you! (Stand still laddie!) Some word of advise from my sponsor: André Dekker
- 49. 49 MAASTRO ©2013 How to start a successful research line • Don‟t choose physics research, choose medical physics research – Take an engineering approach – Choose translational research • Pick a real world clinical problem • Make friends – Cross-border: companies, other researchers • Make sure the problem has not been solved • Make sure you have or can generate the data
- 50. 50 MAASTRO ©2013 Pick a real world problem that your department has • Pick a cancer – Don’t think every cancer is the same • Pick a subtype/subgroup in that cancer – Don’t think every cancer is the same • Pick a problem – Ours: In inoperable stage I-IIIB Non-Small Cell Lung Cancer 2 year overall survival is below 50% • Make sure it is a problem for your department • Make sure it is still a problem in 10 years (it will take time to develop your expertise)
- 51. 51 MAASTRO ©2013 Focus • It will take year before you become an expert in a certain problem • Be ruthless in your focus, say no etc.
- 52. 52 MAASTRO ©2013 Other • Make sure your department is representative / state-of-the-art in that problem – Don’t try to do research to improve your process • Publish or perish is still the norm • Develop a steady stream of BSc, MSc, PhD students and ultimately post-docs and first focus on pumping out manuscripts
- 53. 53 MAASTRO ©2013 Why choose a real problem? • You can make friends with a radiation oncologist • You now have sessions to go to at a conference • You can do the research on-the-job • Your department might actually use the results • You can more easily identify experts • Almost every grant mechanism requires you to actually apply what you dreamt up • Last but not least: Patients should benefit from it!
- 54. 54 MAASTRO ©2013 DO NOT… • choose a problem that is only interesting for physicists • fiddle around with the last % of dose distribution, monte carlo models etc unless you can prove that it has a clinical benefit • try to do research in your spare time, it will not work. Do it on-the-job or in protected research time. • try to do it alone
- 55. 55 MAASTRO ©2013 Thank you for your attention Visit us at: www.maastro.nl www.eurocat.info www.predictcancer.org www.mistir.info www.cancerdata.info
Editor's Notes
- #4: I try to have hobbies…Mechatronics : a lot of INGS, ICS and TICS
- #6: Info / prog man : privilege of having my share in managing Medical Physics Engineers & Data Analysts
- #10: MAASTRICHT: capitol of province South Limburg (122.000 inhabitants)Habitat of MAASTRO clinic (Maastricht Radiation Oncology)Lead Inspirational Professor Philippe LambinPay close attention to one thing!No, not the once-in-a-lifetime clean desk…The Personalized Treatment Decision Support prototype
- #12: Decision Support Systems as foundation of our organizationWhy is this needed?
- #13: From population-based to personal healthcare.However, population data needed for individual decision makingSlide taken from ORACLE presentatoin:The intersection of the 2 industries starts with PV on LS side and Safety at Point of Care on HC side. So it is not the end result, it is first and most active step in move toward personalized health.Future & Oracle vision: multiple data sources from LS & HC co-exist and one can apply all the traditional reactive engines and predictive event-based engines for real-time information on impact of the product. Feed knowledge back into drug development lifecycle mgmt. Patient safety is immediate benefit, l/t benefit is better understanding of patient population.