Capturing multi-disease states on a spectrum with machine learning and routine clinical data

Background: Diseases exist on spectra of risk factors, cellular perturbations, organ dysfunction, and clinical manifestations. It is unknown whether the analysis of routine laboratory tests and vitals using artificial intelligence presents a scalable and portable system for capturing the spectral nature of common diseases. Methods: We constructed and validated machine learning models targeting seven common diseases—atrial fibrillation, breast cancer, coronary artery disease, migraine, rheumatoid arthritis, schizophrenia, and type 2 diabetes—using routine clinical measurements from 394,957 electronic health records (EHRs) in the BioMe Biobank and UK Biobank. The Resulting model outputs, termed spectral health index from machine measurements of electronic records (SHIMMER), were assessed for association with disease diagnosis, risk factors, biomarkers, onset, survival, complications, and medications in two cohorts. Findings: SHIMMER was associated with disease diagnosis, known risk factors, and biomarkers in expected directions in both cohorts. With greater SHIMMER, the prevalence of risk factors, complications, and medications continuously increased; for instance, age and hypertension, stroke risk and cardiac arrest, and beta blockers increased, respectively, with atrial fibrillation SHIMMER. Biomarker levels for type 2 diabetes, such as glucose, hemoglobin A1c, C-reactive protein, and triglycerides, changed stepwise as SHIMMER increased. Rising SHIMMER also revealed gradations of earlier disease onset and decreased survival, particularly for coronary artery disease and schizophrenia. Conclusions: A holistic, non-invasive marker derived from machine learning trained on routine clinical measurements snapshots multiple common diseases on a spectrum, quantifying disease risk, severity, onset, survival, sequela, and treatment. Funding: This study was supported in part by the National Institutes of Health.