Predicting Prostate Cancer Diagnosis Using Machine Learning Analysis of Healthcare Utilization Patterns

Area of Science:

• Computational oncology
• Health informatics
• Machine learning in healthcare

Background:

• Early prostate cancer detection is crucial for effective treatment and improved patient outcomes.
• Understanding pre-diagnostic healthcare utilization patterns can inform predictive modeling.
• The All of Us Research Program provides a rich dataset for studying disease prediction.

Purpose of the Study:

• To investigate healthcare utilization patterns preceding prostate cancer diagnosis.
• To develop and evaluate machine learning models for early prostate cancer prediction.
• To identify key clinical variables predictive of prostate cancer diagnosis.

Main Methods:

• Utilized data from the All of Us Research Program (1,276 cancer patients, 1,232 controls).
• Extracted features from procedure, measurement, and condition records, including Prostate-Specific Antigen (PSA) levels, comorbidity index, and symptoms.
• Trained and tested multiple machine learning models (e.g., XGBoost) to predict prostate cancer diagnosis at 3, 6, 9, and 12 months prior.

Main Results:

• The XGBoost model achieved the highest performance at 3 months (Accuracy=0.73, F1=0.73, AUC=0.82) and 6 months (Accuracy=0.71, F1=0.71, AUC=0.78).
• Predictive performance decreased with longer prediction time windows.
• Prostate-Specific Antigen (PSA) levels were the most significant predictor, followed by triglyceride and creatinine levels.

Conclusions:

• Machine learning models can effectively predict prostate cancer diagnosis using pre-diagnostic healthcare utilization data.
• Early prediction is feasible, particularly within a 3-6 month window prior to diagnosis.
• PSA levels are a critical factor in early prostate cancer prediction models.