Rapid screening for acute rheumatic fever using machine learning analysis of host tissue reactive antibodies

Abstract

Acute Rheumatic Fever and Rheumatic Heart Disease (ARF/RHD) affect over 45 million people globally. ARF/RHD are autoimmune complications following group A streptococcal infections. Current diagnosis of ARF requires thorough medical examination, echocardiography and laboratory tests that are unavailable in most primary care settings where patients with ARF typically first present. This pilot study was conducted to determine whether machine learning-based predictive models could be used to stratify host tissue protein reactive antibodies associated with ARF, that could be incorporated into a lateral flow point-of-care (POC) platform for ARF screening. We investigated serum antibody levels against four host tissue proteins (cardiac myosin, laminin, keratin, and tropomyosin) known to increase in ARF. Serum samples were obtained from: (i) a rat autoimmune valvulitis model (RAV) of RHD (30 streptococcal M protein-injected rats versus 30 controls); and (ii) human samples (25 newly diagnosed ARF patients versus 50 healthy controls). Four machine learning algorithms (logistic regression, decision tree, random forest, and AdaBoost) predicted ARF status using antibody levels detected by enzyme-linked immunosorbent assay (ELISA). Rats injected with streptococcal M protein developed cardiac pathology and demonstrated three-fold higher optical density values for all four host tissue protein reactive antibodies compared to controls. ARF patients showed significantly elevated antibody levels against all host tissue proteins tested (p < 0.01). Random forest achieved optimal performance for rat data (sensitivity 100%, specificity 92.5%, AUROC = 0.97), while AdaBoost excelled for human samples using binary biomarkers (sensitivity 85.0%, specificity 82.8%, AUROC = 0.87). We demonstrate the screening potential of known host tissue protein reactive antibodies and propose lateral flow assay POC technology as a possible advancement toward improved early screening for ARF in resource-constrained environments.