As a newly minted cardiologist in the late 1980s, one of my first jobs was screening corporate executives for heart problems. It was also my introduction to the health insurance system. To my dismay, I found a system with an overarching focus on reducing the utilization of overused medical services, sometimes at the expense of promoting the appropriate use of evidence-based diagnostic and therapeutic interventions. I saw many employees who were on the wrong treatment, or no treatment, despite the presence of serious conditions for which guideline-directed medical approaches were well established.
Ultimately, a colleague and I decided to address this challenge. We developed software that analyzed insurer-derived clinical data and alerted physicians when clinical improvement opportunities surfaced. It may sound counterintuitive, but as it turned out (and was confirmed in a peer-reviewed randomized prospective trial), more care led to lower total costs for our employer and insurance partners. By merely doing the right thing medically, we pre-empted the costly future adverse consequences of disease. A national payer acquired our clinical decision support company in 2005.
That experience opened my eyes to the power of data and computational sophistication, and its potential to markedly improve patient outcomes. While I served as the chief medical officer for that payer, we saw the start of widespread adoption of digital patient records, the advent of patient wearable devices, and the beginnings of a genetic revolution. These innovations have unleashed an avalanche of previously inaccessible data for analysis and customized patient interventions. Artificial intelligence and machine learning, which are transforming other industries, have now become the rage in healthcare with some estimating a marketsize of 6.6 billion by 2021 — up 40 percent annually since 2014.
AI is expected to perform clinical miracles such as predicting which patients will survive following an intensive care unit stay and unearthing the optimal therapy for a particular cancer. Some of these claims have been validated in real-world settings, but many have generated false hope and frustration among the public and in the clinical community.
Perhaps we should learn to walk before we run.
There are a few essential requirements that will accelerate AI-enabled strides in clinical care:
Start with (mostly) structured date
Parsing doctors’ notes is a tricky undertaking. Extracting lab values or physiological parameters is invaluable, but in many instances natural language processing may miss context. Was it you, for example, or your mother, who has a history of heart failure? Unless you can stand up an army of people to comb through clinical notes, AI may reflect potentially misleading information.
Validate the data
The adage ‘garbage in, garbage out’ is true. Claims data are intended to facilitate reimbursement, not necessarily reflect underlying clinical reality. Perhaps your doctor coded for atrial fibrillation to reflect your presentation with heart palpitations. The claim for atrial fibrillation would propagate through your record, even if your palpitations turned out to be a harmless, rare heart rate irregularity associated with nervousness. Check and cross-check for validation of claims veracity, looking for repeated diagnoses or the use of certain interventions or drugs associated with a “true” diagnosis. If a patient has a claim for a heart attack on a particular date with no associated hospitalization, that’s another red flag.
Begin with the evidence
Conclusions drawn from well-conducted clinical trials may serve as a better foundation for artificial intelligence and machine learning than raw statistical inference. By first starting with established, efficacious interventions (e.g., anticoagulation for stroke reduction in high risk atrial fibrillation patients), we may then deploy AI to personalize interventions for individual patients. This approach supports the optimization of costly, or risky, interventions for specific patients most likely to benefit.
To prove AI works, conduct clinical trials designed to demonstrate superior outcomes. AI is making significant advances in imaging interpretation, with encouraging results in areas like diabetic retinopathy. As we extend to the domain of highly complex and nuanced clinical decision making, let’s ensure similar proof of benefits (or flaws). Evidence and credibility are essential as we endeavor to augment the skills of practicing clinicians.
With ever-expanding patient data and the sheer power of AI, novel approaches to clinical decision making will naturally emerge. However, a rigorous focus on data-driven and evidence-based approaches is paramount. This will instill confidence in the medical community and ensure we keep patients’ best interest at heart.
Photo: ipopba, Getty Images