Machine Learning Risk Stratification for Older Breast Cancer Survivors: Clinical Care Implications.

Area of Science:

• Oncology
• Biostatistics
• Health Services Research

Background:

• Breast cancer survivorship presents complex health challenges.
• Identifying survivors at high risk for adverse outcomes is crucial for effective management.
• Existing prediction models may not fully capture the multifactorial nature of survivorship risks.

Purpose of the Study:

• To develop and validate a machine learning-based clinical risk prediction algorithm.
• To identify breast cancer survivors at high risk for adverse outcomes post-treatment.
• To predict the risk of all-cause death, cancer-specific death, recurrence, and other adverse health events.

Main Methods:

• Retrospective analysis using linked Surveillance and Epidemiology End Results (SEER) registry and Consumer Assessment of Healthcare Providers and Systems (CAHPS) survey data.
• Cross-validated random forest machine learning models were employed.
• Prediction models were developed for adverse outcomes within 3 and 5 years post-treatment completion.

Main Results:

• The algorithm demonstrated high out-of-sample accuracy for predicting adverse outcomes (e.g., 91.9% for overall adverse events, 97.5% for cancer-specific death).
• Key predictors included geographic region, age, frailty, comorbidity, time since diagnosis, and out-of-pocket costs.
• Significant rates of death, recurrence, and hospitalization for adverse events were observed within the follow-up period.

Conclusions:

• Machine learning models can accurately predict adverse survivorship outcomes in breast cancer patients.
• Non-cancer-specific factors play a significant role in driving these adverse outcomes.
• Risk stratification can inform personalized care, with high-risk survivors potentially benefiting from intensive care and low-risk survivors from primary care management.