Prostate Cancer Prediction Using Healthcare Utilization Patterns

Early cancer detection is crucial for improved patient outcomes, as evidenced by research on major cancer types emphasizing the impact of timely treatment initiation. This study focuses on identifying pre-diagnosis patterns in prostate cancer, utilizing supervised machine learning to build predictive models analyzing patients' medical activities one year before diagnosis. The dataset, sourced from the All of Us Research Program, specifically targets prostate cancer cases diagnosed between 2010 and 2019. By grouping CPT4 codes in clinically significant categories and employing the XGBoost model in machine learning, the study achieved superior performance with accuracy and area under the curve (AUC) of 0.94 for predicting cancer one month prior to diagnosis and 0.76 five months before diagnosis. In addition, the top important features derived from the model were surgical pathology procedure, cardiac stress tests, hospital inpatient, number of total visits, and diagnostic ultrasound of the head and neck. Despite a decline in accuracy when predicting 5 months and 1 year ahead, this research lays the groundwork for personalized and timely interventions, advancing cancer diagnostics and early intervention strategies.