A machine learning case study to predict rare clinical event of interest: imbalanced data, interpretability, and...

Area of Science:

• Pharmaceutical industry
• Clinical trial methodology
• Machine learning applications

Background:

• Accurate prediction of rare clinical events is critical in drug development due to potential life-threatening health risks.
• Delayed detection of rare adverse events can significantly impact patient safety.
• Machine learning offers a powerful approach to address this challenge in pharmaceutical research.

Purpose of the Study:

• To define and solve the rare clinical event prediction problem using machine learning in a pharmaceutical industry context.
• To propose a six-step investigation framework for better communication and interpretation of model performance.
• To enhance patient screening processes for future clinical trials.

Main Methods:

• Adaptation of rare-event stratified split for data splitting, accounting for multiple patient records.
• Employment of cost-sensitive learning to handle imbalanced data by weighting the minority class.
• Utilizing precision and recall metrics for performance evaluation, instead of raw accuracy.
• Application of SHAP values for identifying key risk factors and improving model interpretability.

Main Results:

• Demonstration of a practical machine learning framework for rare clinical event prediction.
• A proposed six-step framework aids non-technical stakeholder communication and practical interpretation of results.
• Identification of important risk factors through SHAP values enhances model interpretability.

Conclusions:

• The developed machine learning approach and framework effectively address the challenge of rare clinical event prediction in drug development.
• This methodology can improve patient safety through enhanced prediction and early risk identification.
• The study provides a valuable tool for optimizing patient screening in clinical trials.