EvolveFNN: An interpretable framework for early detection using longitudinal electronic health record data

Abstract The extensive adoption of artificial intelligence in clinical decision support systems necessitates a significant presence of ML models that clinicians can easily interpret. Therefore, we developed an RNN-based interpretable method, combining the fuzzy concepts and recurrent units, to train accurate and explainable models on high-dimensional longitudinal electronic health records data. Through supervised learning, our method allows the identification of variable encoding functions and significant rules. To demonstrate its performance and capabilities in classification and rule discovery, we first tested it on a simulation dataset. The proposed methods achieved the best model performance, and the rules learned are almost identical to what we used to generate the synthetic data. Furthermore, we showcased a pilot application that proved its potential in the early detection of cardiac event onset. We constructed a prediction window of length from 30 days to 180 days and a 360-day retrospective observation window to evaluate our algorithm. Our proposed algorithm EvolveFNN with GRU unit achieved an AUPRC of 0.695 and an AUROC of 0.682 and EvolveFNN with LSTM unit achieved an AUPRC of 0.696 and an AUROC of 0.684 when the prediction window size is 180 days. The proposed algorithm obtained a comparable model performance to vanilla GRU models and remains relatively stable when the prediction window size changes. Examining the rules generated by our EvolveFNN model with the GRU unit, we found that the extracted rules align with clinical practices and existing literature. This provides potential risk factors not explored before in the population, such as chloride. To assess the generalizability of the algorithm, we conducted experiments on the MIMIC-III benchmark dataset for in-hospital mortality prediction. The results showed our proposed algorithms achieved model performance comparable to the top-performing GRU model and surpassed other approaches by providing transparent decision reasoning. In conclusion, our approach, EvolveFNN, can effectively train accurate, interpretable, and reliable models using large longitudinal electronic health records datasets, thereby offering valuable insights for clinicians.