An interpretable framework to discover the medical decision rules of the length of ICU stay for patients undergoing craniotomy based on electronic medical records (Preprint)

BACKGROUND Craniotomy is a kind of operation with great trauma and many complications, and patients undergoing craniotomy should enter the ICU for monitoring and treatment. Based on the electronic medical records (EMR), the discovery of high-risk multi-biomarkers rather than a single biomarker that may affect the length of ICU stay (LoICUS) can provide better decision-making or intervention suggestions for clinicians in ICU to reduce the high medical expenses of these patients and the medical burden as much as possible. The multi-biomarkers or medical decision rules can be discovered according to some interpretable predictive models such as ensembling methods. OBJECTIVE Our study aims to develop an interpretable framework based on the real-world EMRs to predict the LoICUS and discover some high-risk medical rules of patients undergoing craniotomy. METHODS The EMR data sets of patients undergoing craniotomy in ICU are separated into pre-operative features and post-operative features. The paper proposes a framework called Rules-TabNet (RTN) based on the data sets. It is a rule-based classification model and the high-risk medical rules can be discovered from RTN, and a risk analysis process is implemented to validate the rules discovered by RTN. RESULTS The performance of post-operative model is significantly better than that of pre-operative model. The post-operative RTN model has the best performance compared with other baseline models, and it achieves 0.93 accuracy and 0.98 AUC for the task. Seventeen key decision rules that may have impact on the LoICUS of patients undergoing craniotomy are discovered and validated by our framework. CONCLUSIONS The proposed post-operative RTN model in our framework can precisely predict whether the patients undergoing craniotomy are hospitalized for too long (more than 12 days) in ICU. We also discover and validate some key medical decision rules from our framework.