Background: Active surveillance (AS) for prostate cancer (PCa) requires serial assessments and there is a need to optimize resource utilization by personalizing AS. Evaluating serial PSA values is fundamental in AS and a model predicting future PSA may aid decision-making on optimizing the timing of the subsequent assessment. In this study, we use a machine learning (ML) approach to predict future PSA in AS patients. Methods: We searched electronic health record data at two tertiary academic centers for patients with a diagnosis of PCa or reason for visit related to PCa (e.g. elevated PSA) between 1990 and 2020. Within this group, we identified patients diagnosed with PCa and managed with AS. We trained different ML models (Table 1) to predict the next PSA value given a sequence of historical PSAs and expected date of next PSA assessment, and compared this prediction to the ground truth (actual PSA measured). We augmented the model with clinical features (age, prostate volume, and BMI). We compared the ML models to a baseline method that considers a patient-specific constant slope based on prior PSA values. The test set consisted of AS patients only (10% of all AS patients), not used in training. The dataset was split patient-wise. The primary performance metric was root-mean-square error (RMSE). We compared training using AS patients only to using all patients, still testing on AS patients only. We evaluated feature importance to facilitate model interpretability. Results: We identified 4269 patients (totaling 33371 PSA values) with at least 3 PSA values and complete data for the clinical features. Of these, 1134 were on AS (totaling 10358 PSA values). On average, AS patients were 65 years old (± 7 [standard deviation]), with a mean PSA of 6.5 ng/mL (± 6.4), prostate volume of 48g (± 27) and BMI of 28 kg/m2 (± 5), at diagnosis. The best performing model was the Gradient Boosting Regression model that was trained in the heterogeneous sample of all patients (including those not on AS) (Table 1). Evaluation of feature importance found that increasing the number of previous PSA measurements as input to the model further reduced the RMSE. Conclusions: Our study used machine learning to predict future PSA in PCa patients on AS. We also demonstrate that for our best performing model, a heterogeneous training sample further improved performance. While our model requires further validation, a robust similar model may be useful as a decision-aid tool to personalize AS and optimize resource-utilization. We have deployed our model at https://psa-evolution.streamlit.app/.