ACAT: Confidence-Aware Transformer for Incomplete EHRs

Abstract

Electronic Health Records frequently contain incomplete laboratory measurements and physiological observations or physiological data and forecasting the disease is often unpredictable. To address this, our team put forward Adaptive Confidence-Aware Transformer (ACAT). This solution does multiple things: it infers missing data, provides some information about interactions between clinical features, allows to estimate the level of certainty, and makes predictions more transparent. Developed an adaptive mechanism which observes the pattern in missing data, but not only does it fill it in, but it learns from it. The transformer then performs multiple functions: disease classification optimization, confidence estimation, calibration and explainability. In our evaluation of ACAT model based on such datasets as MIMIC-IV, eICU and an external benchmark, we noticed its high performance even with 10-50% of missing data. The benefits of ACAT include an improvement of the ROC-AUC of up to 4.2%, a decrease in Expected Calibration Error of up to 28% and maintain a comprehensible and credible prediction.