Towards interpretable, medically grounded, EMR-based risk prediction models.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support

Background:

• Machine learning (ML) models offer improved risk prediction accuracy over clinicians.
• Explainability and medical grounding are crucial for ML risk models to provide actionable insights.
• Current 'black-box' ML models often lack transparency, hindering clinical application.

Purpose of the Study:

• To demonstrate the development of explainable and medically grounded ML risk prediction models.
• To assess the utility of such models in predicting postoperative complications.
• To illustrate how model transparency can aid clinical decision-making.

Main Methods:

• Trained ML models using clinically relevant inputs from electronic medical records for pre- and postoperative risk assessment.
• Compared predictive performance against models using broader input ranges.
• Visualized model decision-making processes to explain input influence on predictions.

Main Results:

• Developed explainable ML models for postoperative complication risk prediction.
• Achieved predictive performance comparable to less transparent models.
• Demonstrated visualization techniques to clarify how input factors influence risk predictions.

Conclusions:

• Explainable and medically grounded ML models can achieve high predictive performance.
• Transparency in ML models facilitates the identification of actionable risk factors for clinical intervention.
• This approach enhances the utility of ML in clinical practice by providing interpretable risk assessments.