No-Code Training of Machine Learning Models using Intensive Care Unit Serology Hospital Data to Predict Survival outcomes of Hospital Surgery

Abstract

Introduction: Surgery and prolonged admission in intensive care units (ICU) can introduce complications such as sepsis, iatrogenic infections and other adverse events. Early recognition of serology laboratory markers by AI-assisted tools could identify high-risk patients to guide proactive decision-making. This study aimed to develop and evaluate no-code AI prediction models to estimate postoperative survival in ICUadmitted patients.

Materials and Methods: Open-source, de-identified ICU data from The Zigong Fourth People’s Hospital, China, available on Physio.net2 were analysed using risk-ratios to identify a basket of high-risk serological markers. These markers were then used as features to develop self-learning decision-tree models to predict patient survival outcomes based on their serological data.

Results: Multiple laboratory markers were associated with mortality outcomes. Our initial Decision-Tree model had a prediction accuracy of 56%, with similar results across Neural Networks (58.0%) and Random Forest models (57.9%). Performance across all models was consistent (~58.0%) against the Boston-SymileMIMIC-database.

Conclusion: This assessment of AI learning models demonstrated only modest accuracy (56-58%) in clinical prediction, limiting its clinical applicability. This study, however, did not leverage larger cloud architectures like Amazon Web Services or Google Cloud which could allow significantly larger feature selection. Predictive performance could be improved in future studies by scaling to large-scale cloud computing resources so that feature sets are not limited by on-site infrastructure constraints.