Temporal self-attention for risk prediction from electronic health records using non-stationary kernel approximation

Rawan AlSaad, Qutaibah M. Malluhi, Alaa A. Abd-Alrazaq, Sabri Boughorbel

Abstract: Highlights•Modeling temporal data in EHR is important for medical condition characterization.•Our method extends self-attention mechanism with non-stationary kernel approximation.•Method evaluated using EHR of 76925 patients to predict diagnosis codes in next visit.•Results demonstrate that non-stationary time modeling boosts healthcare predictions.