An Approach for Combining Clinical Judgment with Machine Learning to Inform Medical Decision Making: Analysis of...

Area of Science:

• Medical Informatics
• Health Services Research
• Clinical Decision Support

Background:

• Machine learning (ML) offers potential for identifying treatment effect heterogeneity but faces barriers in interpretability and usability for clinical decision-making.
• Integrating clinical judgment with ML is crucial for developing transparent and trustworthy evidence for personalized medicine.

Purpose of the Study:

• To develop and illustrate a 4-stage approach combining clinical judgment with ML for generating comparative effectiveness evidence.
• To enhance the relevance and usefulness of ML-driven insights for clinical decision-making, particularly for patients with multiple long-term conditions.

Main Methods:

• A 4-stage approach was developed, incorporating clinical judgment to identify modifying factors of treatment effectiveness.
• Least Absolute Shrinkage and Selection Operator (LASSO) ML was applied to large-scale administrative data (N=24,312) to select covariates from over 500 possibilities.
• Estimates of comparative effectiveness were generated for relevant patient subgroups, with results presented via an interactive visualization tool.

Main Results:

• Non-emergency surgery (NES) strategies for acute appendicitis showed increased days alive and out-of-hospital compared to emergency surgery (ES) overall.
• Treatment effects varied significantly across subgroups, with notable differences observed in patients with chronic heart failure and kidney disease versus those with cancer and hypertension.
• An interactive tool was developed to visualize ML outputs, allowing customization for specific clinical decision-maker needs.

Conclusions:

• Combining clinical judgment with ML provides a principled approach to enhance the trust, relevance, and utility of comparative effectiveness evidence.
• The developed method effectively generates clinically relevant subgroup evidence, supporting personalized decision-making for patients with multiple long-term conditions.
• Interactive visualization tools can improve the accessibility and application of complex ML findings in clinical practice.