Patient similarity analytics for explainable clinical risk prediction

Area of Science:

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

Background:

• Clinical risk prediction models (CRPMs) are valuable tools but face limited adoption due to poor explainability and interpretability.
• Explainability refers to describing a model's prediction process, while interpretability concerns understanding its predictions.

Purpose of the Study:

• To demonstrate the utility of patient similarity analytics in developing explainable and interpretable CRPMs.
• To compare the performance of a patient similarity model against traditional machine learning models.

Main Methods:

• Utilized electronic medical records from patients with type-2 diabetes mellitus, hypertension, and dyslipidemia.
• Developed a modified K-nearest neighbour patient similarity model incorporating expert input.
• Validated the model on a separate dataset and compared its performance with logistic regression, random forest, and SVM models.

Main Results:

• The patient similarity model achieved an AUROC of 0.718, comparable to logistic regression (0.695), RF (0.764), and SVM (0.766).
• A prototype web application demonstrated the model's ability to provide quantitative and qualitative patient narratives.
• These narratives aided clinical decision-making, such as facilitating patient agreement for insulin therapy.

Conclusions:

• Patient similarity analytics provide a feasible method for creating explainable and interpretable CRPMs.
• This approach is generalizable and can be tailored to specific databases for locally relevant insights.
• The developed CRPMs can serve as effective clinical decision support tools, promoting shared decision-making.