Hacking Healthcare: How AI and the Intelligence Revolution Will Reboot an Ailing System

Book Preface

The COVID-19 pandemic exposed the profound vulnerabilities of American healthcare, with disproportionately poor outcomes in underrepresented minorities and the indigent. The United States had the highest per capita mortality rate of any industrialized nation, with more than 1 million deaths. Although only comprising 4% of the world’s population, US fatalities accounted for more than 16% of the pandemic deaths through 2021. Poor clinical outcomes for the US pandemic were not at all new for this country. Even with over $4 trillion spent on healthcare in 2021, outstripping all countries in the world, the lifespan of its residents, along with maternal and infant mortality metrics, are the worst of all 38 Organisation for Economic Co-operation and Development (OECD) member countries. If there is a better example of when you don’t get what you pay for, I don’t know of it.

In Hacking Healthcare, Tom Lawry, the National Director for AI, Health and Life Sciences at Microsoft, explores the intersection of the pandemic and the Intelligent Health Revolution. There are few examples for AI having a substantive impact on the pandemic. One medication, baricitinib, an anti-inflammatory JAK-kinase inhibitor used for treating rheumatoid arthritis, was identified by data mining and ultimately proved, by randomized clinical trials, to achieve mortality reduction for severe COVID. Many other AI data mining efforts to repurpose existing drugs have not been successful, despite subsequent testing in clinical trials. Another way AI made a difference in the pandemic was in Greece, whereby an algorithm was launched to determine which travelers should be tested for COVID. It turned out to be two to four times more efficient than random testing.

Literally, thousands of studies claimed that AI could accurately diagnose COVID via a chest X-ray or CT scans. But an in-depth review of the best 62 papers (of 2,212 preprints or publications) categorized them as “Frankenstein datasets” and concluded: “None of the models identified are of potential clinical use due to methodologic flaws and/or underlying biases.” That exemplifies the hype of AI, that its overall impact for the pandemic was modest at best, and that it is still early in the real-world clinical validation and acceptance of AI tools.

This book, however, gets into the remarkable promise that AI has for the future. Let’s take the potential for AI to reduce health inequities. In Africa, India, and many remote parts of the world, smartphone ultrasound is now being used to make a diagnosis. Noteworthy is that the person obtaining the ultrasound is not necessarily a doctor or clinician; the AI is starting to direct an uninitiated individual to obtain the desired image and do an auto-capture of a picture or video loop of interest. Algorithms are also being developed to provide immediate and accurate interpretations of the images obtained. Once fully validated, this creates a “closed-loop” without the need for trained personnel to rapidly and inexpensively obtain preliminary diagnoses anywhere in the world. We’re not there yet, but the early experience suggests it may be eminently achievable. Even now there is leveling of the earth with the smartphone attached probe that acquires an image that can be shared in the cloud and interpreted by an expert remotely.

It’s important to underscore, however, that AI tools could easily make inequities worse, if they are only available to affluent people. We have already seen too many examples of how bias can creep into algorithms and how AI tools can be used in practice that may unwittingly promote discrimination and unfairness. Attention to these concerns is vital if AI is ever going to reduce health inequities.
It took decades for clinical researchers to finally determine that the social determinants of health were just as important, if not more so, than traditional risk factors such as diabetes or hypertension. While helpful, a person’s zip code is far too rudimentary to get a handle on this risk factor. Early studies that are noted in the book show how such data are starting to get imputed with natural language processing of electronic medical records including education, economics, neighborhood, access to health, health literacy, and social context.

During the pandemic, the Hospital at Home (HaH) concept was tested by several health systems in the United States by default, since the faculties were overloaded with patients and the idea was to use remote monitoring of those who were not as sick with COVID at presentation to the Emergency Department. Some health systems used a multi-sensor device from Current Health placed on the upper arm that continuously monitors all vital signs except blood pressure: temperature, oxygen saturation, heart rate and rhythm, and respiratory rate. There haven’t been any randomized clinical trials or publications that prove remote monitoring with this device or other similar biosensors is as safe as admitting patients to the hospital for monitoring but the HaH concept and the tools to test it – the hardware and AI analytics – are now becoming available. Interestingly, Best Buy acquired Current Health, so that surprise should be an indicator of some unpredictable combinations and directions AI will take us in the years ahead. Hopefully less need for hospitalizations will be one of them.

The cost of American healthcare is such an outlier in the world. Lawry reviews many ways AI can help this dire situation such as with streamlining operations, billing, detection and prevention of fraud, optimizing staffing, avoiding hospital readmissions, and clinic no-shows. Reducing hospitalizations would certainly put a dent in it. They alone account for over $1 trillion in the annual US healthcare expenditures, but they also represent important revenue streams for hospitals. The American Hospital Association is one of the largest lobbyist organizations in the country. Unlike countries with universal healthcare, there is no incentive in the United States to actively promote HaH, owing to perverse incentives. AI cannot fix our lack of universal healthcare, a deep lesion that is holding us back in so many ways. The lack of access to healthcare for the under-represented and indigent is a by-product of our fee-for-service model. The points that the promise of AI has limitations, that the field is still early, and there are critical obstacles to overcome cannot be emphasized enough.

That brings me to the human capital narrative in American healthcare. We keep adding jobs to this sector at a torrent pace, which is already the largest labor force and well over the traditional major sectors of retail and manufacturing. Yet there’s still a marked shortage of personnel. The pandemic has magnified the issue greatly with not only intensification of clinician burnout, which had already manifest as a global crisis, but now there are resignations or early retirements at scale. We are losing an enormous number of nurses, doctors, and other clinicians from the American healthcare workforce. This is an area we urgently need AI to come to the rescue, to augment the productivity efficiency and accuracy of each clinician remaining, along with the new ones to join. On top of my list, and what we came to when I did the review of the National Health Service of the United Kingdom is keyboard liberation. That is using natural language processing and machine learning to eliminate or greatly reduce any need for clinicians to function as data clerks, with synthetic notes generated by the conversation between the patient and doctor – AI, not human, scribes. Just a first base for what AI can conceivably do to make the practice of medicine the way it was before clunky electronic medical records for billing purposes took over.
However, if AI is misused such that doctors and nurses are getting  “squeezed” to do even more, that will only exacerbate the conditions  that the pandemic ushered in. Therein lies the rub of managers and administrators, the overlords of the American health system, who focus on revenue and not morale, who are interested in implementing AI tools but may not be aware of what must be regarded as the cardinal objective. The people who went into medicine did so because they, for the most part, want to take care of patients. Reciprocally, patients want to know their doctors have a presence and have their back, that they really care for them. If clinicians feel they cannot provide care effectively, which has clearly been the case during the pandemic, their interest for staying in medicine will be lost. Using AI to make doctors see more patients faster, or read more scans, or slides, is not the fix that we need. The overarching goal
of AI in healthcare, as I wrote extensively in Deep Medicine, needs to be to restore the human connection, the precious patient-doctor relationship, the essentiality of medicine, and its humanity. Let’s hope as healthcare is rebooted and hacked, we will get there someday

Eric J. Topol, MD La Jolla, California