Improving Mortality Risk Prediction Using an LSTM Network Combined With Self-Attention

Reliable mortality risk prediction help professional medical staffs make better decisions for patient healthcare. While previous death risk prediction tools and models for critically ill patients have always been depended on labor-intensive collection of data and mainly focused on static features, which may affect the applicability and accuracy of the models. By utilizing routine time-series clinical data collected during regular workflows of patient care, this study developed an mortality risk prediction model based on an LSTM network which incorporates with a dual self-attention design, which is designed specifically to enhance the model’s focus on critical periods and important features. The model’s performance was further compared with the traditional machine learning algorithms in an independent testing dataset. The results suggested that the LSTM model incorporating dual self-attention performed best in predicting risk of death. With an excellent AUROC value of 0.9583, and an AUPRC value of 0.8191, It has a very good calibration performance verified by a much lower brier score. Furthermore, we confirmed that physiologic variables related to respiration and circulation are key predictors of mortality. The desirable predictive performance of this model provides additional evidence for its use in clinical practice.