Prediction of Effectiveness and Toxicities of Immune Checkpoint Inhibitors Using Real-World Patient Data

Area of Science:

• Oncology
• Immunology
• Data Science

Background:

• Immune checkpoint inhibitors (ICIs) improve cancer outcomes but can cause severe toxicities.
• Predicting these risks is crucial for personalized treatment decisions and clinical trial design.

Purpose of the Study:

• Develop a machine learning (ML) framework using electronic health record (EHR) data.
• Predict risks of hepatitis, colitis, pneumonitis, and 1-year overall survival in patients receiving ICIs.

Main Methods:

• Utilized real-world EHR data from over 2,200 patients treated with ICIs.
• Engineered features by aggregating laboratory data over 60-365 days.
• Developed and validated random forest classifiers.

Main Results:

• Models achieved AUCs between 0.729 and 0.755 for predicting pneumonitis, hepatitis, colitis, and 1-year survival.
• Performance was robust across distinct outcome-specific feature sets.
• The cohort predominantly had melanoma, lung, or genitourinary cancers.

Conclusions:

• This is the first ML solution to assess individual ICI risk-benefit profiles using routine EHR data.
• The framework requires no additional data collection, facilitating clinical integration.
• Enables personalized risk assessment for improved therapeutic decision-making.