A novel machine-learning model using nuclear features to predict upper tract urothelial carcinoma recurrence

Abstract Upper-tract urothelial carcinoma (UTUC) is rare, constitutes < 5% of urothelial cancers, and, given its insidious clinical presentation, is typically diagnosed at advanced stages; post-radical nephroureterectomy recurrence remains challenging. Postoperative management is primarily guided by pathological factors, including T-stage and lymph-node status, which cannot accurately identify patients who would benefit from treatment intensification, particularly considering concerns regarding cisplatin-associated nephrotoxicity. Using an artificial intelligence (AI)-informed pathology model, we predicted postoperative recurrence using UTUC’s quantitative nuclear features. Among 222 patients with UTUC, support vector machine (SVM) and random forest (RF) models were trained using pT3 cases ( n = 68) comprising sufficient recurrence events. Patient-level model performance was evaluated and validated using an independent test cohort ( n = 50; pT1 = 12, pT2 = 11, pT3 = 22, and pT4 = 5). RF and SVM models achieved patient-level accuracies of 77.3% and 63.6%, respectively, in pT3 cases and overall accuracies of 80% and 68%, respectively, in the independent pT1–pT4 cohort. Model-derived risk stratification (1 point per model with predicted recurrence probability ≥ 0.5 and classifying patients into low, intermediate, and high-risk groups [0, 1, and 2 points, respectively]) significantly discriminated recurrence-free survival in the independent test cohort (pT1–pT4), distinguishing risk groups (low-risk: n = 7, intermediate-risk: n = 8, high-risk: n = 35; log-rank p = 0.01); the low-risk group had no recurrence.