|Date/Time||Mar 27 2019 16:00 GMT|
|9:00am PDT/12:00pm EDT|
|4:00pm GMT/5:00pm CET|
|Co-Champions:||Ram D. Sriram and David Whitten|
Ontology Summit 2019 Medical Explanation Session 2
The speaker today is:
- Ugur Kursuncu and Manas Gaur
- Explainability of Medical AI through Domain Knowledge
- Video Recording
- Wright State University
Conference Call Information
- Date: Wednesday, 27-March-2019
- Start Time: 9:00am PDT / 12:00pm EDT / 5:00pm CET / 4:00pm GMT / 1600 UTC
- ref: World Clock
- Expected Call Duration: 1.5 hours
- The Video Conference URL is https://zoom.us/j/689971575
- iPhone one-tap :
- US: +16699006833,,689971575# or +16465588665,,689971575#
- Dial(for higher quality, dial a number based on your current location): US: +1 669 900 6833 or +1 646 558 8665
- Meeting ID: 689 971 575
- International numbers available: https://zoom.us/u/Iuuiouo
- iPhone one-tap :
- Chat Room
[12:09] Ken Baclawski: The recording will be posted after the session, and an outline of the slide content is posted below.
[12:16] RaviSharma: with help of live chat I can now see his slides.
[12:18] ToddSchneider: How should we understand the notion 'concept based information'?
[12:22] RaviSharma: Ugur - what is either the improvement in diagnosis with use of All AI data, multimodal medical data vs only social media data?
[12:23] RaviSharma: or improvement in probability of social media based only data?
[12:26] ToddSchneider: Is the mapping of the 'personal data' into/with the medical knowledge based on natural language terms or phrases?
[12:26] RaviSharma: how does openness of patient in soc media affect the result?
[12:28] RaviSharma: medical entity data?
[12:28] TerryLongstreth: At what point do the subjects (patients..?) know that their social media accounts are being scanned/extracted? Did the research control for intentional misdirection on the part of subjects after they learned? Or was the use of social media data covert/hidden from the subjects?
[12:32] ToddSchneider: Is there an assumption of the existence of social media data for a person?
[12:32] RaviSharma: if you limit the social interaction among the similar patients what do you expect the result to be compared to social media data?
[12:36] ToddSchneider: Perhaps a better question is 'How much personal data is needed' (for the system to be 'useful')?
[12:50] Ken Baclawski: Arash Shaban-Nejad Semantic "Analytics for Global Health Surveillance" will be speaking on April 17. Slides are available at http://bit.ly/2YvlHLK
[12:53] TerryLongstreth: PSQ9 - questionnaire
[12:54] RaviSharma: thanks ken
[12:59] ToddSchneider: I have to get to another meeting. Thank you.
[13:02] RaviSharma: please upload speaker slides, thanks
[13:06] RaviSharma: thanks to speakers
The following is an outline of the slide content, not including the images.
1. Explainability of Medical AI through Domain Knowledge
- Ugur Kursuncu and Manas Gaur
with Krishnaprasad Thirunarayan and Amit Sheth
- Kno.e.sis Research Center
- Department of Computer Science and Engineering
- Wright State University, Dayton, Ohio USA
2. Why AI systems in Medical Systems
- Growing need for clinical expertise
- Need for rapid and accurate analysis for growing healthcare big data including Patient Generated Health Data and Precision Medicine data
- Improve productivity, efficiency, workflow, accuracy and speed, both for doctors and for patients
- Patient empowerment through smart (actionable) health data
3. Why Explainability in Medical AI Systems
- Trust in AI systems by clinicians and other stakeholders
- Major healthcare consequences
- Legal requirements; need to adhere to guidelines/protocols
- More significant for some specific medical fields, such as mental health
4. Patient-Doctor Relationship
- Cultural and political reason for ownership of personal data
- Privacy concerns for personal data:
- Two stages for permission to use: model creation, personal health decision-making
- Incomplete data due to privacy concerns
- How would AI systems treat patients?
- For personalized healthcare: Researchers or analyzers or doctors need such personal data to provide explainable decisions supported by AI systems
5. How will AI assist humans in medical domain?
- Intelligent assistants through conversational AI (chatbots)
- Multimodal personal data
- Text, voice, image, sensors, demographics
- Help physician burnouts
- Legal implications
- Common ground and understanding between machines and humans.
- Forming cognitive associations
- Big Multimodal data
- Ultimate goal: Recommending or Acting?
7. Problem: Reasoning over the outcome
- How were the conclusions arrived at?
- If some unintuitive/erroneous conclusions were obtained, how can we trace back and reason about them?
8. A Mental Health Use Case
- Recommendations on:
- Relevance to disease
- Recommendations on:
- Big multimodal data for humans!
- Explainability is required as to how data is relevant and significant with respect to the patient situation
9. Explainability vs Interpretability
- Explainability is the combination of interpretability and traceability via a knowledge graph
10. A Mental Health Use Case
- From Patient to Social Media to Clinician to Healthcare
11. A Mental Health Use Case
- From Patient to Clinician via a Black box AI system
12. A Mental Health Use Case
- What is the severity level of suicide risk of a patient?
- ML can be applied to a variety of input data: Text, image, network, sensor, knowledge
13. Explainability with Knowledge
- Explainability through incorporation of knowledge graphs in machine learning processes
- Knowledge enhancement before model is trained
- Knowledge harnessing after model is trained
- Knowledge infusing while model is trained
14. Explanation through knowledge enhancement
15. Relevant Research: Explaining the prediction of mental health disorders (CIKM 2018)
16. Relevant Research: Explaining the prediction of mental health disorders (CIKM 2018)
- Explanation through word features that is created through Semantic Encoding and Decoding Optimization technique
- Semantic encoding of personal data into knowledge space
- Semantic decoding of knowledge into personal data space
17. Relevant Research: Explaining the prediction of mental health disorders (CIKM 2018)
18. Relevant Research: Explaining the prediction of mental health disorders (CIKM 2018)
19. Relevant Research: Explaining the prediction of severity of suicide risk (WWW 2019)
20. Relevant Research: Explaining the prediction of severity of suicide risk (WWW 2019)
- Progression of users through severity levels of suicide risk
21. Explanation through Knowledge Harvesting
22. Relevant Research: Explaining the prediction wisdom of crowd (WebInt 2018)
23. Explanation through Knowledge Infusion
24. Explanation through Knowledge Infusion
- Learning what specific medical knowledge is more important as the information is processed by the model
- Measuring the importance of such infused knowledge
- Specific functions and how they can be operationalized for explainability
- Knowledge-Aware Loss Function (K-LF)
- Knowledge-Modulation Function (K-MF)
- ROC & AUC
- Assessments of true positives and false positive rates, to properly measure feature importance
- Inverse Probability Estimates
- Estimate the counterfactual or potential outcome if all patients in dataset were assigned either label or have close estimated probabilities
- PRM: Perceived Risk Measure
- The ratio of disagreement between the predicted and actual outcomes summed over disagreements between the annotators multiplied by a reduction factor that reduces the penalty if the prediction matches any other annotator
27. Mental Health Ontology
- Extensively used in this research
- Built based on DSM-5, which is the main guideline documentation for psychiatrists
- Includes: SNOMED-CT, Drug Abuse Ontology and Slang terms
28. Key Takeaways
- Medical explainability is a necessity to form trust for medical community
- Three ways of explainability with knowledge
- Interpretability and traceability are necessary and sufficient conditions for explainability
- Infusing knowledge would further enhance the reasoning capabilities