Improving clinical decision support through interpretable machine learning and error handling in electronic health...

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support Systems

Background:

• Electronic medical record (EMR) data often contains errors and biases, hindering the development of reliable machine learning (ML) models for clinical decision support.
• Existing ML algorithms struggle to incorporate complex physiological and biological constraints inherent in medical data, limiting their interpretability and performance.

Purpose of the Study:

• To develop a novel data processing tool, Trust-MAPS, that integrates clinical domain knowledge into EMR data to improve ML model error handling, bias mitigation, and interpretability.
• To enhance the predictive power and clinical relevance of ML models used in healthcare applications.

Main Methods:

• Developed Trust-MAPS, an algorithm translating clinical knowledge into mathematical models that capture physiological constraints.
• Projected EMR data onto a constrained space to identify and quantify deviations from healthy physiology using "trust-scores."
• Integrated trust-scores into the feature space for downstream ML tasks, demonstrating utility with a sepsis prediction model using XGBoost and SMOTE.

Main Results:

• The Trust-MAPS framework effectively handled data errors and improved predictive performance.
• Achieved an area under the receiver operating characteristic curve of 0.91 for predicting sepsis 6 hours prior to onset, a 15% improvement over baseline.
• Demonstrated bias reduction and enhanced interpretability of the ML model.

Conclusions:

• Trust-MAPS preprocessing significantly improves downstream classification performance and reduces bias in ML models.
• Trust-scores provide clinically meaningful features, enhancing predictive accuracy and interpretability for clinical decision support.
• This novel method translates clinical knowledge into mathematical constraints for improved high-dimensional medical data analysis.