OBJECTIVES: Although only 5% of patients with prostate cancer (PCA) have metastases at diagnosis, approximately 20% of those with early stage disease have recurrence with metastases, often years later. Changes in prostate specific antigen (PSA) levels can predict recurrence. A machine learning approach to serum PSA lab kinetics, after diagnosis of localized PCA, was used to forecast progression with metastases.

METHODS: PSA lab values for 4,654 (4,321 non-metastatic; 333 metastatic) PCA patients, absent of metastases at diagnosis, were evaluated. Labs were included beginning at diagnosis and up to an observation cutoff of 3.5 years in surviving patients. All subsequent lab visits were censored. Lead time in patients with metastases was established by withholding labs 1.5 years before the appearance of metastases. Patients with normal PSA at diagnosis were included. Patients without metastases were sampled before training for equally-weighted labeling. A binary classification with XGBoost was used with aggregated PSA levels, alongside clinical and laboratory attributes, as feature inputs in a 70:30 train-test split. Model selection was determined by tuning of observation period constraints.

RESULTS: Our classifier achieved an overall test accuracy of 76.9%, with 76.6% precision, and 75.0% sensitivity. Prediction with PSA kinetics precedes clinical detection of metastases by 1.5 years using a maximum 3.5 years of observed lab values.

CONCLUSIONS: Using this machine learning approach, longitudinal PSA kinetics may forecast, years in advance, the appearance of metastases in patients diagnosed with early stage PCA. This accuracy is comparable to other methods predicting PCA recurrence by PSA kinetics but with a clinically-relevant lead time. With further development of this approach, accurate prediction of disease progression may be applied to other malignancies.