Machine Learning for the Prediction of Distant Metastases Following Postprostatectomy Salvage Radiation Therapy

Abstract

<h3>Purpose/Objective(s)</h3> Salvage radiotherapy (SRT) is the only curative option for patients with biochemical recurrence (BCR) following radical prostatectomy (RP). Tendulkar's nomogram was developed to predict outcomes following SRT in patients with a median pre-SRT PSA of 0.5 ng/mL. We aim to externally validate Tendulkar's nomogram and provide an updated tool for the prediction of distant metastasis (DM) in a contemporary cohort of patients with lower pre-SRT PSA. <h3>Materials/Methods</h3> Patients were included from two academic institutions, treated between 1989 and 2019, with RP followed by SRT for either persistent or increasing PSA. A gradient boosted tree was trained on data from one institution (n=2,529) and tested on another (n=566). For baseline comparison, we also evaluated Tendulkar's nomogram. Models were evaluated in terms of their time-dependent area under the curve (AUC) of the receiver operator curve (ROC). <h3>Results</h3> Median age was 64 years. 1,422 (469%) patients were pT2, 1,047 (33.8%) pT3a, and 626 (20.2%) ≥pT3b. Most patients (78.48%) had a surgical Gleason score (GS) of 7 and 10.6% had GS ≥8. Almost one-third (30%) had positive surgical margins and 12.6% had persistent PSA post-RP. Median pre-SRT PSA was 0.27 ng/mL. 208 (6.7%) patients received androgen deprivation therapy with SRT. Median RT dose was 66.6 Gy. Median follow up from SRT was 4.6 years. Distant metastasis occurred in 9% of patients at 5 years, and 12.5% at 10 years. On the test dataset, our model showed AUCs of 0.72 (CI: 0.63 – 0.81) and 0.74 (0.65 – 0.83) for the prediction of DM development at 5-and 10-years compared to 0.57 (0.49 – 0.66) and 0.66 (0.55 – 0.78) for Tendulkar' nomogram. The improvement in the prediction of DM development at 5 years showed statistical significance (p <0.05), whereas this improvement was nonsignificant for the prediction of DM development at 10 years. <h3>Conclusion</h3> We provide an updated tool with improved prediction of distant metastases for patients with BCR after RP treated with SRT in modern practice.