Predicting Neurological Disease Risk Using Machine Learning Models: Evidence from Electronic Health Records Across the United States
Abstract
lectronic health records (EHRs) were primarily designed to document care, rather than forecast disease. Within the past five years, research experts have queried whether machine learning can be used to mine the same records, including lab values, years of diagnosis codes, clinical notes, and medication orders, to identify who is going to have a stroke, Alzheimer’s disease, or other neurological conditions before the symptoms are obvious. This review scans recent evidence on the question and synthesises multiple studies to identify the gap. In analyses of several dozen United States studies, machine learning models built on routine hospital and health-system data have predicted the risk of stroke with accuracy (area under the curve, or AUC) of 0.90–0.99, consistently outperforming tools such as the CHA2DS2-VASc score. The onset of Alzheimer’s disease and related dementias 1-5 years in advance is predicted with a gradient-boosted model trained on de-identified HER data, with AUCs between 0.809 and 0.833, while surfacing sleep apnea and headache as previously underappreciated risk signals. A separate model built on over 52,000 Cleveland Clinic records indicated that ordinary clinical data comprising patient-reported outcomes, sleep assessments, fall histories, and weight trends are characterised by easily detectable signs of impending neurological disease. Similarly, an increasing body of fairness research shows that these models can demonstrate worse performance, or different behaviour, especially for Black and other minority patients. The reason for this is that the training data in itself is a reflection of unequal access to care. The study, therefore, concludes that machine learning has the potential to predict neurological disease with EHR data, which are already sitting in hospital systems. Meanwhile, that power will only translate into equitable benefit if bias auditing is routinely carried out as much as accuracy testing.