Automatically explaining machine learning prediction results: a demonstration on type 2 diabetes risk prediction.

Area of Science:

• Healthcare Informatics
• Machine Learning
• Medical Predictive Modeling

Background:

• Machine learning (ML) models offer high accuracy in healthcare predictive modeling but lack interpretability, hindering clinical adoption.
• Interpretability is crucial for integrating ML into healthcare settings for trust and validation.
• Existing ML models often fail to provide explanations for their predictions.

Purpose of the Study:

• To present the first comprehensive method for automatically explaining ML predictive model results.
• To demonstrate the method's ability to maintain prediction accuracy while providing explanations.
• To address the critical need for interpretable ML in healthcare.

Main Methods:

• Developed and implemented a novel computational method for generating explanations for any ML predictive model.
• Utilized the Practice Fusion diabetes classification competition dataset, comprising electronic medical records from across the US.
• Applied the method to predict type 2 diabetes diagnosis within one year.

Main Results:

• The method successfully explained prediction results for 87.4% of patients accurately predicted to have type 2 diabetes.
• The champion ML model from the competition was utilized for demonstration.
• The explanation method was applied to a real-world healthcare dataset.

Conclusions:

• The developed method is feasible for automatically explaining ML predictions in healthcare.
• The approach successfully addressed the challenge of ML interpretability without accuracy degradation.
• This work paves the way for wider adoption of accurate ML models in clinical practice.