The talk deals with the recent success in AI, and how it is transferable to the medical domain. Despite the progress in machine learning, the medical domain has not profited so much from the hype.

1. Artifical Intelligence
for Medicine
Tassilo Klein
SAP Connected Health
Image credits: bestfreejpg.com
January 19, 2017

2. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Talk Outline
• Recent success in machine learning
• Why deep neural networks became so popular
• Situation in medical data domain
• Limitations of current machine learning approaches
• Machine learning of the future
2

3. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Recent Development - Deep Learning
3Source: http://www.waymo.com/, http://www.apple.com/, http://www.uber.com

4. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Recent Success in Deep Learning
Deep Learning Technology
Human
Performance
Computers can „see“ at
human performance level

5. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Computers can play games
• Go requires more elements that mimic human intelligence
than chess
• 2015 Google DeepMind AlphaGo beat 18-time world
champion Lee Sedol
• Prior to 2015, best Go programs only managed amateur
level
• AlphaGo is using Deep Learning
• Evaluation heuristics are to a large extent learned by the
program itself
Silver et al., Mastering the game of Go with deep neural networks and tree search, Nature 2015
Source: http://www.nature.com/

6. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Algorithmic
improvements
E.g. convolutional
networks, recurrent
networs, ReLUs etc.
Huge
Datasets
E.g. ImageNet (14M
images)2, WordNet
(155k words in
taxonomy)3)
Massive
Parallelization
TFLOPs on consumer
GPUs
Deep Learning - Three Key Enablers
$17B
ML & AI investments
since 2009
$1.9B
Funding for 670 companies in 13
categories in 2014
62%
Annual private investment growth for
the past 4 years
What’s Different Now? New Momentum
Core technology
dates back to
the 80s / 90s

7. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
What‘s going on in the medical domain?

8. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Medical data / knowledge
• Immense growth in medical knowledge
Densen, Challenges and Opportunities Facing Medical Education, Transactions of the American Clinical and Climatological
Association, 2011
• Dimensionality growing due to Big Data trend, multimodality, multi-omics, etc.
• Manual analysis difficult and getting more and more impossible
• Humans are excellent in pattern recognition with low dimensions (≤3)
Marr et al., Vision: A computational investigation into the human representation and processing of visual information‘, MIT
Press, 2010
8

9. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Data Challenges in Medical Domain
Privacy
Integration
Harmony
Erroneous Data
Missing Data
Imprecise Data
Johnson et al., Machine Learning and Decision Support in Critical Care, IEEE Transactions on Biomedical Engineering, 2015
• No data stored for observation
• Random or systematic
• Data collection comes at a cost
• Precision limits of measuring device
• Data collected for other purpose
• Artifacts
• Data partly or completely false
• Endo- and exogenous noise sources
• Concept inconsistency
• Patient information scattered in
disconnected data sources
• Protect patient confidentialy
• HIPAA

10. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Why Deep Learning in near future will not
solve all problems medical domain?

11. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Key characteristics for ML adoption
• Accurate – the models have to get it right
• Interpretability – Patients and doctors have to understand decisions
• Actionable – results have to lead to conrete actions
• Credible – Outputs have to agree with what is known in clinical literature
• Robust – adapt to changes over time and population
• Data - Learning must be done with the data that is available, not the data one would want

12. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
ML Situation in Medical Domain
• Recent advances in machine learning largely due to publicly available Big Data
• Learning must be done with the data that is available, not the data one would want
• Scattered stakeholder-ship
• Almost no openly large available datasets
• Data often recorded for other reasons such as billing rather than analysis
• Often little incentives for accurate and detailed data acquisition
• Impossible or very expensive to obtain labels
• Typically no benchmark dataset
• Biological diversity of individuals, environment and pathogenesis
• Variance in clinical practices
• Millions of patient records are needed to cope with complexity and variability
• Medicine is more than math
Consequence: Very very little research work ends up in healthcare

13. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Limitations – Uncertainty Modeling
• It is important to understand when model is falsely over-confident (particular silent and unpredictable bad
decisions)
• Deep Learning approaches unreliable when test examples lie outside of data distribution
• Currently, uncertainty only poorly modeled
• Particularly important when learning from small data or incomplete (e.g. observational studies)
• Data from observational studies: no control or even full understanding of action mechanism
How would a patient who obtained medication A, respond to medication B?
Maybe the patient was prescribed medication A only because it was cheaper
Which one
works better?
Counterfactual
Inference
?

14. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Limitations: Deep Learning can be fooled I
Szegedy et al., Intriguing properties of neural networks, International Conference on Learning Representations, 2014
Prediction
Noise
99,9% certainty

15. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Limitations: Deep Learning can be fooled II
Ngyen et al., Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images, Computer vision and Pattern Recognition (CVPR), 2015
Evolutionary
algorithms
modify images
≥ 99.6 certainty in
class prediction

16. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Deep Learning: Data Hungry Black Box
• Is a black box – intransparent decision making process
• Machine learning: Trade-off between accuracy and interpretability
• More or less only does pattern recognition
• Needs a lot of data
• Training takes a lot of time
• Training process needs lots of tweaks
• Only weakly bio-inspired
• Training with 14m images
• Classification of 1k classes
• Training on 28.4m
positions/moves from 160k
games
• In addition: playing against itself
30m times
• For comparison: the Go
champion Lee Sedol played
~50k games in his life
Source: http://www.nature.com/, http://www.deepmind.com/, /

17. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
AI Safety
• Poor objective function: Wrong or poor definition can lead to harmful results
• Negative side effects: Focusing on one task and neglecting other, may cause harm to environment
• Reward hacking: Cost function admits some clever ‚easy‘ solution perverting design objective
• Safe exploration: Exploratory actions should not lead to negative consequences; exploring involves taking
actions, with unknown consequences; in complex environment risk anticipitation harder
• Robustness to distributional shift: Avoiding of making bad decisions when training/test distributions differ
[„Mostly prominent in the domain of reinforcement learning / robotics,
where an agent has to acquire knowledge by actively exploring an
environment“]
Amodei et al., Concrete Problems in AI Safety, arXiv, 2016

18. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Human learning
Lake et al., Building Machines That Learn and Think Like People, Behavioral and Brain Sciences, 2016
• Natural intelligence is still the best example of intelligence
• Distinguished by richness and efficiency
• Machines may rival or exceed performance on particular tasks
• Human learning
• Big data vs. Small data: People learn a lot more from a lot less
• Model building: Learning of rich representations
• Concept learning (boundary of the infinite set of possibilities)
• Beyond pattern recognition
• Generalization without additional training
• Possession of intuition for domains such as physics and psychology

19. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Future: Machines that learn like humans
Lake et al., Building Machines That Learn and Think Like People, Behavioral and Brain Sciences, 2016
• Main ingredients for better learning
• Composability: Explicit representation of objects, identity, relations through
combination of primitive elements
• Causality: Representing hypothetical real world processes producing real world
observations
• Learning-to-learn: Learning new tasks is accelerated through previous learning
• Future generations of neural networks will be very different from state-of-the art
Machines will have to learn more like a human:
• To cope with complex data
• Support reason for inference
• Cope with small data

20. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
Conclusion
Contact: Tassilo.Klein@sap.com
https://www.linkedin.com/in/tassiloklein
• Medical-data access will be solved, e.g. platforms
• Machine learning will from big data to small data paradigm
• Deep learning and other learning paradigms probably move closer to human-like learning
Thank you for your attention!

21. © 2015 SAP SE or an SAP affiliate company. All rights reserved.
© 2015 SAP SE or an SAP affiliate company. All rights reserved.
No part of this publication may be reproduced or transmitted in any form or for any purpose without the express permission of SAP SE or an SAP affiliate company.
SAP and other SAP products and services mentioned herein as well as their respective logos are trademarks or registered trademarks of SAP SE (or an SAP affiliate
company) in Germany and other countries. Please see http://global12.sap.com/corporate-en/legal/copyright/index.epx for additional trademark information and notices.
Some software products marketed by SAP SE and its distributors contain proprietary software components of other software vendors.
National product specifications may vary.
These materials are provided by SAP SE or an SAP affiliate company for informational purposes only, without representation or warranty of any kind, and SAP SE or its
affiliated companies shall not be liable for errors or omissions with respect to the materials. The only warranties for SAP SE or SAP affiliate company products and
services are those that are set forth in the express warranty statements accompanying such products and services, if any. Nothing herein should be construed as
constituting an additional warranty.
In particular, SAP SE or its affiliated companies have no obligation to pursue any course of business outlined in this document or any related presentation, or to develop
or release any functionality mentioned therein. This document, or any related presentation, and SAP SE’s or its affiliated companies’ strategy and possible future
developments, products, and/or platform directions and functionality are all subject to change and may be changed by SAP SE or its affiliated companies at any time
for any reason without notice. The information in this document is not a commitment, promise, or legal obligation to deliver any material, code, or functionality. All forward-
looking statements are subject to various risks and uncertainties that could cause actual results to differ materially from expectations. Readers are cautioned not to place
undue reliance on these forward-looking statements, which speak only as of their dates, and they should not be relied upon in making purchasing decisions.