Machine Learning for the Prediction of Lymph Node Metastasis in Patients with Prostate Cancer

Purpose/Objective(s)Many tools are used to predict the probability of lymph node involvement (LNI) in patients with prostate cancer before radical prostatectomy (RP) and to select patients for lymph node dissection. Our goal was to externally validate common prediction tools and to create a machine learning model for the prediction of LNI in patients with prostate cancer referred for RP.Materials/MethodsPatients treated at two academic institutions between 1990 and 2020 with RP followed by pelvic lymph node dissection were included. We trained a gradient boosted tree on data from one institution and the model was validated on the other dataset. Variables incorporated into the model are patient's age, percentage of cores positive on biopsy, Gleason scores, preoperative PSA, and clinical T-stage (cT). The area under curve of the receiver operator characteristic curve (AUC) served as the performance metric. We further compared our model's performance to that of established nomograms, namely the Roach formula, the online Memorial Sloan Kettering Center (MSKCC) nomogram, and the 2012 Briganti nomogram.ResultsThe combined dataset consisted of 19,090 patients with median age of 65, median PSA at diagnosis of 7.60 ng/mL, and median percentage of positive cores at biopsy of 44%. 14,252 patients (74.6%) had a Gleason score (GS) of 7, and 4,841 (25.4%) had a GS of ≥ 8. Most patients (18,570 or 97.3%) had cT-stage ≤cT2 disease, and 523 (2.7%) had cT3-4 disease. 3,058 patients (16%) were found to have lymph node involvement. The external validation dataset had 1,093 patients, with 10.7% having lymph node positive disease. After training and on external validation, the AUC of our model was 0.79 (CI: 0.7465-0.83), compared to 0.69 (0.64-0.7455), 0.73 (0.6813-0.7742), and 0.74 (0.6931 - 0.7914) when using the Roach formula, MSKCC nomogram, and Briganti 2012 nomogram, respectively. Furthermore, our model's improved performance was statistically significant, with p-value <0.05 when comparing its AUC to that of all other models.ConclusionOur machine learning model can be used to estimate an individual patient's risk of LNI and serves as an additional tool with improved performance compared to established prediction models.