Ignoring the mirage of the disposable clinician for the successful deployment of AI in medicine

Go back to programme

Ignoring the mirage of the disposable clinician for the successful deployment of AI in medicine

Medicine is driving many investigators from the machine learning community to the exciting opportunities presented by applying their methodological tool kits to improve patient care. They are inspired by the impressive successes in image analysis (e.g. in radiology, pathology and dermatology) to proceed to broad application to decision support across the time series of patient encounters with healthcare. I will describe the central role of the clinician in this process, and I will examine closely some of the under-appreciated assumptions in that research/engineering agenda. Lastly, I focus on how ignoring these will limit success in medical applications and conversely how these assumptions define a necessary and ambitious research program in shared human-ML decision making.

Shownotes

 

04:58 Introduction by Maha Farhat

 

07:04 Introduction by Isaac Kohane

 

07:38 Keeping the doctor in the loop for ML

 

08:24 Disclosure

●     Board member: Inovalon, Canary Medical, Activate CAre, i2b2TransMART Foundation

●     Scientific Advisory Board Member: Danaher, PulseData, Medaware, Rady CHildren’s Genomics Institute

●     Conflict of interest: I am passionately biased towards findings and hypotheses that I believe are sound. 

 

09:11 THE ML SPECTRUM 

●     What is machine learning? 

●     Used for decision making.

 

11:33 Translating Artificial Intelligence into clinical care

●     There are advances in automatic image recognition. For instance, Diabetic Retinopathy. 

●     The work of a doctor of these images is to add labels (they are the persons behind detecting diseases by using ML). 

 

14:00 AI helps clinicians predict when COVID-19 patients might need to go to the Intensive Care

●     When using an AI model in health, it is important to understand how the healthcare system works. 

 

16:09 Survival 3 Years After a WBC Test

●     Physiology of patients vs. physiologies of healthcare system (doctor responses) 

 

19:00 Predicting survival from ordering a lab test

●     If a test is ordered at an odd time, it is because there is greater danger, and thus lower survival rates

 

20:00 Machine learning for patient risk stratification

●     ML is learning from the senior doctor

 

24:28 Case-control studies (shape of 1000s of studies on medical ML studies)

●     Example: patients with heart attacks and without. (HA vs. control) 

●     ML would design a way to specify the feature of a HA or a control patient. 

 

25:50 Risk as a trajectory 

●     Non-disease state to disease state in an instantaneous way (measuring at the beginning and at the end, but not in the path to the disease)

 

26:17 Prediction implied a time dimension 

●     Protein A to predict a heart attack.

●     ⅛-⅕  effect of the the protein effect size relative to reported baseline.

 

27:15 Simulating prospective deployment 

●     Measurement and finding samples of patients with the trajectory of past and future of the disease. 

 

27:29 Poor leaders in machine learning community 

●     Domestic abuse can be detected on average 2 years earlier (up to six years)

●     Such a system is yet to be deployed, despite it being a clear danger

 

29:33 Why doctors hate their computers 

●     Digitization is prone to make medical care easier and more efficient. But  screens are coming between the doctor and patients.

 

30:03 Machine Learning in medicine 

●     Daily clinical workflow: Audio recording of a conversation between a doctor and a patient. Provides automated generation of sections of a clinical note or automatic assignment of billing codes from an encounter. 

 

31:30 Is this the future of medicine in 10 years? 

●     Treatment of a patient as a game? 

 

33:11 When does medicine succumb to AlphaZero?

●     A system should be: Deterministic; fully observed; discrete; simulated; short; evaluative and based on a huge dataset of human work

●     All these characteristics are different from databases in the healthcare sector. 

●     Which medical "game" fits the bill?: Reimbursement 

 

35:32 Where is the game now clearer? 

●     Reinforcement learning to maximize reimbursement

●     Reinforcement learning to minimize/maximum care utilization 

●     Referral approval is discrete

 

36:49 Adversarial attacks 

●     Attacks on medical machine learning 

●     Original image —> Noise -> ML model is tricked 

●     This affect the classification of an image 

 

39:10 Who completes the loop?

●     Kohane monitored his mothers weight, and in case of weight increase, advised an additional dose of medicine

●     Why did this work despite RCTs not being able to show this?

●     Patient trusting the doctor

●     Daily monitoring

●     Doctor needs to listen to the patient, increased nightly bathroom visits affect weight measure in the morning

 

43:20 Coronavirus update

●     SAIL 2021: 18-20 October Fairmont Hamilton Princess Hotel, Hamilton Bermuda: http://sail.health/

 

44:01 Analysing remarks from Maha Farhat. 

 

44:28 Start of the Q&A Session 

 

44:33 Q: Doctors should have the role of checking the data for sandness for a given task. 

 

●     Yes in order to have better data. But sometimes you don't know where more data is required. To do this we need smarter programmes and personnel who understand the data. Therefore we will have the opportunity to deliver high quality care. 

 

46:50 Q: Do you think there is a role for computers and helping collect or label patient data?

 

●     Yes, a lot of ways of finding the performance of doctors by using, for example, cameras. When you use a camera you can watch if they watch their hands or not and you can supervise this by ML 

 

48:34 Q:What do you think of smartwatches using ML to diagnose Heart conditions? 

 

●     This is something that would happen at some point. These technologies will deliver less in the short term and more in the long term. I am optimistic. 

 

51:25 Q: Better design roles out of technology to ensure already marginalized groups, do you have something similar in your work? 

 

●     Several efforts addressing issues to marginalised individuals which do not have access to some ICT’s.

 

55:58 Q: Can you provide a framework for a doctor to critique each prediction case by case using the model internal properties? Will it help in the same fashion like ordering a lab test but it’s for AI prognostication?

 

●     I think that’s a wonderful idea. IN the NI of health would be a successful grant. Doctors critique is a framework currently and it's right. This is exactly what we need.

 

57:28 Closing of the Q&A session

 

 

57:40 Closing remarks

Share this session
Scroll Up