Bridging the Gap Between Accuracy and Interpretability: A Hybrid LSTM Approach with SHAP for ICU Mortality Prediction Using EHR Data

While deep learning models have achieved remarkable performance, their adoption in healthcare faces a critical challenge due to a lack of interpretability. Interpretability is a serious issue in high-stakes environments like Intensive Care Units (ICUs), where transparency in decision-making is not just mandatory but also essential. Several studies have ascertained the trade-off between performance and interpretability, with interpretability being sacrificed for performance and vice versa. Therefore, this study aims to demonstrate that performance and interpretability need not be mutually exclusive by proposing a hybrid framework that integrates LSTM, a deep learning architecture, with explainable models such as SHAP for ICU mortality prediction using Electronic Health Record (EHR) data. The study employs publicly available ICU datasets such as MIMIC-III or MIMIC-IV, which contain comprehensive EHR data for ICU patients. The LSTM achieved an accuracy of 98.6% and a recall of 87.5% on unseen data, but recorded low Precision, indicating that the model was biased toward the majority class (No Mortality). When the LSTM results were compared with baseline models (Random Forest and Logistic Regression), it generally outperformed the baseline models. The major limitation is the presence of class imbalance within the dataset, as shown by the precision. Despite this, the LSTM model successfully maintained interpretability through SHAP without compromising predictive performance, thereby achieving a balance between accuracy, transparency, and clinical relevance.