AI in Health Care–National Academy of Medicine’s Perspective

Here is the full report

Here are some highlights:

1. Promoting population-representative data with accessibility, standardization, and quality is imperative
2. Prioritize ethical, equitable, and inclusive health care AI while addressing explicit and implicit bias
3. Contextualizing the dialogue of transparency and trust requires accepting differential needs.
4. Near-term focus is needed on augmented intelligence vs AI autonomous agents
5. Develop and deploy appropriate training and educational programs to support health care AI.
6. Leverage frameworks and best practices for learning health care systems, human factors, and implementation science to address the challenges in operationalizing health care AI
7. Balance innovation with safety via regulation and legislation to promote trust.

Artificial Intelligence in a Smartwatch for Atrial Fibrillation

An interesting study by Perez et al in NEJM found some very interesting things:

1. A large number of people opted-in to be part of the study (over 400K)
2. In about 4 months, about 0.5% participants were noted to have irregular heart rate – a relatively small percentage than one would expect.
3. Of the people who returned 7-day monitoring after being notified of irregular heart beat, about one-third had atrial fibrillation. This is a significant number of atrial fibrillation in otherwise asymptomatic individuals.

The study highlights several important things, some noted above. However, what it does not tell us if the use of smartwatch to identify atrial fibrillation in otherwise healthy people results in improved health outcomes or not. A future study should be able to evaluate this question. For now, at least we know we can depend on smartwatch to identify atrial fibrillation.

Are Machine Learning Tools Better than Standard Tools in Predicting Readmissions

Just saw this study where authors compared standard readmission tools (HOSPITAL score, modified LACE score, and Maxim/RightCare score) with a model developed using machine learning. Authors found that machine learning score (they called it Baltimore Score or B-score) performed much better than standard tools.

While I agree that machine learning tools will likely outperform standard methods. Standard methods are quite a bit of oversimplification of the real life, machine learning tools less so. However, I doubt that authors have got their model right. Two reasons: One, their sample size is relatively small. Two, they have not validated their tool in a new dataset.

Deep Learning in Pathology

Here is an intersting article, showing the role of AI in pathological diagnoses. Medical speciliaties most dependent on pattern recognition, such a pathology, dermatology, or radiology, are more likely to get AI assistance sooner. With time, it is quite possible that AI can replace such physicians. For now, it appears that AI will help them work faster. Remember, increasing efficiency means need for fewer physicians in such specilaties and possibly decliing salary.

Artificial Intelligence & Medicine

Some recent articles ……

Goldberg JE, Rosenkrantz AB. Artificial Intelligence and Radiology: A Social Media Perspective. Current Problems in Diagnostic Radiology. 2018 Jul 23.
An interesting study examining the types of conversations on Twitter about the role of artificial intelligence in radiology. It appears that most tweets (or linked websites) were upbeat and wanted radiologists to continue supervision of AI-run diagnostics.

Meskó B, Hetényi G, Győrffy Z. Will artificial intelligence solve the human resource crisis in healthcare?. BMC health services research. 2018 Dec;18(1):545.
This article expresses the hope that artificial intelligence may be able to help in mitigating human resource crisis in health sector.

dos Santos DP, Giese D, Brodehl S, Chon SH, Staab W, Kleinert R, Maintz D, Baeßler B. Medical students' attitude towards artificial intelligence: a multicentre survey. European radiology. 2018 Jul 6:1-7.
This article examines the attitudes of undergraduate medical students towards artificial intelligence in radiology and medicine. Students think that artificial intelligence will revolutionize the practice of medicine, in particular radiological diagnosis.

Machine Learning Predicts Clinical Deterioration

Edelson and her colleagues trained a gradient boosted machine model on a dataset of almost 300K patients and were able to predict clinical deterioration better than the modified early warning (MEWS) system. Here is the link for the abstract that was presented at ATS 2018.

The simplicity of the tool which uses only age, heart rate, and respiratory is really impressive. Interestingly, respiratory rate is often the least accurately measured vital sign in the hospitals despite its simplicity. A more accurate measurement of respiratory rate can potentially help better predictive models to predict clinical worsening in a patient’s condition. Better prediction can not only help to patients by focusing on the cause of clinical worsening but will also help to more optimally utilize resources within the hospital.

Productivity Paradox and EHR

Bob Wachter and Michael Howell are optimistic about the future improvements in health information technology and its impact on physicians. I share their enthusiasm and I am optimistic about an increase in productivity in healthcare sector. And why not? Its about time.

Productivity in every other sector has resulted in decreasing utilization of human resources and increasing production of goods. However, such is not the case in healthcare. A physicians still sees about the same number of patients in a day as he/she used to see 50 or 100 years ago. It is quite possible that the number patients seen might have dropped due to documentation and other requirements. The result is evident; healthcare sector now accounts for almost one-fifth of the GDP in the US. Much of this has to so with the lack of growth in productivity in healthcare sector as compared to other sectors.

What is promising is that artificial intelligence will likely to increase productivity in the healthcare. I don’t think that this will come through improvements in health information technology (HIT). I believe it will come with better utilization of HIT as data repository for learning algorithms. Technology will also help in data capture, and it does so in many ways even now, but most importantly in data inference. However, it will not be good for an average physician.

Remember, productivity can increase only if more patients can be seen by fewer physicians (or healthcare workers). That is not possible with one physician simultaneously seeing several patients. It is possible only when machines see most patients and physician sees patients only in those situations where algorithms have not gotten enough data to develop a routine.

And what those situations will occur in relatively rare conditions. These will not our usual common diseases, what we might call ‘bread-and-butter’ medicine. Rather these will be either rare diseases or rare manifestations of common diseases. A physician overseeing such a highly productive healthcare delivery system will have to be a master clinician and a superb diagnostician.

In other words, an increase in productivity in healthcare sector will spell the doom for majority of physicians ….. happy to hear what you think……….

Artificial Intelligence versus Human Doctors

While it is relatively easy to imagine artificial intelligence taking over pattern recognition type of physician work (such as reading images like radiologists or identify skin conditions by looking at the skin like dermatologists), the job of a general internist is much more demanding and requires assimilation of data inputs from a large number of sources, hence making it difficult for an artificial intelligence system to accurately predict a disease (or prepare a list of differential diagnoses).

On the other hand, the process is not, and should not be, very difficult for a well-crafted system. Razzaki et al tested such an artificial intelligence algorithm and compared it with physicians in its ability to predict diagnosis. For their study, they adopted a semi-naturalistic, role-play paradigm that simulated a realistic consultation between a patient and either their artificial intelligence system or a human doctor. Their study was designed to assess both the clinical (diagnostic and triage) accuracy and the ability to gather all of the relevant history from the patient. What they found was that their artificial intelligence algorithm was able to provide diagnostic and triage advice with a level of accuracy and safety approaching that of human doctors. If there study results are replicated then such systems may not only reduce costs and improve access to healthcare worldwide but may also provide better standardization of care delivered to patients.

A Fantastic Weekly Update on Artificial Intelligence

I have been enjoying this weekly update by Azeem Azhar. And I invite you to try it too. Here is the link

http://www.exponentialview.co