Explainable Deep Learning with Human Feedback for Perioperative Complications Prediction

Health problems are very common among pregnant women, and seemingly normal pregnant women may experience physiological disorders, which can lead to perinatal complications, greatly endangering the health of pregnant women and their newborns. Timely identification, provision of relevant resources, and timely response are the key to preventing serious complications and mortality in delivery women. The current predictive models used in medicine have an imbalance between interpretability and accuracy. In addition, there is a lack of utilization of knowledge in the field of medical expert domain knowledge, which is a waste. The method proposed in this article combines deep learning with regular decision trees to ensure high accuracy while improving its interpretability. In addition, adding expert domain knowledge and providing additional useful information can improve model performance.