A Stepwise Approach to Diagnostic Excellence
Most AI-driven algorithms take the clinician from A—the ailing patient—to D—the final diagnosis—ignoring intermediate steps B and C in the diagnostic journey. There is a better way.
By John Halamka, M.D., president, Mayo Clinic Platform, and Paul Cerrato, senior research analyst and communications specialist, Mayo Clinic Platform
The digital approach to better disease screening, diagnosis, and treatment has been nothing short of revolutionary. Well-documented AI-enabled algorithms are now available to help identify patients most likely to develop colorectal cancer (ColonFlag), improve the detection of diabetic retinopathy (IDx-DR), and make insulin-assisted management of Type 2 diabetes easier (DreaMed). However, while these new tools have the potential to transform patient care, each of them moves the clinician from A to D, i.e., from initial signs and symptoms to final diagnostic or therapeutic decision. But as most experienced practitioners know, there are usually intermediate steps in the process. They too can benefit from innovative digital tools. This more nuanced, refined approach allows clinicians to find their way, step by step, through the diagnostic or therapeutic journey.
Julie Adler-Milstein, Ph.D., and her colleagues at University of California, San Francisco and Stanford University provide a brilliant description of this “wayfinding” approach to diagnostic excellence in a recent JAMA paper. In their words, the current AI tools have had a “long-standing focus … on predicting final diagnostic labels instead of helping clinicians navigate the dynamic refinement process of diagnosis.” They recommend “… shifting the role of diagnostic AI from predicting labels to “wayfinding” (interpreting context and providing cues that guide the diagnostician).” What would this shift in emphasis look like?
It would start with the realization that the diagnostic process is often a slow, complex journey. When a patient tells you their signs and symptoms, the next step is to gather data, including an assessment checklist, a detailed conversation that allows the patient to help create a narrative, and a physical exam. Those initial steps may or may not trigger the order for an EKG, blood work, any number of other procedures, or a referral to a specialist. Once all that data are collected, the clinician will form a working hypothesis about the root cause of the patient’s chief complaint. At this point, there’s often several forks in the road and the clinician may change direction as each puzzle piece falls in place. Each of these alternate pathways refines the diagnostic process, and there are AI-related tools that can be of assistance. Along one of these pathways, a physician might benefit from an algorithm that helps interpret an EKG result, or one that takes advantage of a natural language processing (NLP) program to extract more intelligence from the patient interview. The right NLP-enhanced algorithm might suggest additional questions to ask the patient that the physician had not considered. Similarly, Adler-Milstein, et al point out: “AI could detect patterns in patient data from digital monitoring devices (such as sleep quality or walking distance) and suggest additional questions (e.g., screening for sleep apnea) or testing (e.g., ankle brachial index measurement).”
The wayfinding approach could offer several other potential benefits. For instance, a digitally enabled platform that incorporates polygenic risk scores (PRS), , can not only help predict risk for complex diseases but also refine the diagnostic process. Unlike diseases caused by a mutation in a single gene—cystic fibrosis, for instance—there are many disorders in which several mutations contribute to their onset. PRSs enable diagnosticians to home in on these diseases. Several studies have shown that these genetic scores, coupled with traditional clinical risk factors, can more accurately identify patients at high risk of coronary artery disease. Similarly, there’s evidence to suggest that PRSs can help refine the diagnosis of diabetes mellitus, differentiating between Type 1 and Type 2. Combining machine learning with PRSs and adding these tools to the wayfinding journey will likely improve the diagnostic process even further.
It is also possible to expand this diagnostic wayfinding approach to preventative medicine and digital wellness. Adler-Milstein et al applied their methodology to symptomatic patients, but it is just as important to create a dynamic refinement process that starts with an asymptomatic patient as the first node. The journey would then include several intermediate nodes that utilize a variety of other algorithms. These tools can assess their need for disease screenings, as well as evaluate nutritional status, psychosocial stress, sleep hygiene, and physical activity level; the final node on this preventative journey would be high level wellness. There are several online tools available to capture and analyze this kind of data, and in the hands of committed clinicians and patients, this approach can radically transform our healthcare ecosystem.
The IT revolution in health care continues to generate AI-based tools to help arrive at definitive diagnoses and care plans. But let’s not forget that most journeys require several stops along the way.