Prostatectomy incontinence model retains risk prediction over time

A perioperative model developed from routinely collected clinical information continued to separate patients likely to experience persistent urinary incontinence after radical prostatectomy when tested in a later cohort treated at the same hospital. The study retrospectively analyzed 726 patients treated from 2020 through 2023 and a non-overlapping temporal validation cohort of 465 patients treated from January 2024 through June 2025. Persistent post-prostatectomy urinary incontinence was defined as an International Consultation on Incontinence Questionnaire-Short Form score of at least 1 at 12 months after catheter removal.

The LASSO-penalized logistic regression model retained eight original predictors: age, preoperative International Prostate Symptom Score, transformed prostate-specific antigen level, operative time, postoperative hospital stay, smoking history, hypertension, and neurovascular bundle preservation. It was designed for risk estimation around catheter removal, meaning that it incorporates information from the operation and immediate postoperative period rather than serving as a purely preoperative tool. In the later cohort, the model’s ROC area under the curve was 0.836, with a 95% confidence interval of 0.763–0.900, compared with 0.913 during nested internal validation.

That temporal discrimination was useful but weaker than the development estimate. The average precision score was 0.409 in the later cohort, with a 95% confidence interval of 0.277–0.564, while the Brier score was 0.062, with an interval of 0.045–0.081. Average precision remained above the event-prevalence reference of 0.082, showing that the model identified more of the relatively uncommon outcomes than would be expected from chance alone. However, mean predicted risk was only 5.9%, compared with an observed 12-month event rate of 8.2%, indicating that the locked model underestimated absolute risk on average.

Calibration results reinforced that distinction between ranking patients and estimating their individual probabilities. The temporal calibration-in-the-large value was 0.445, with a 95% confidence interval of 0.012–0.808, and the calibration slope was 0.850, with an interval of 0.621–1.174. The interval around the slope included the ideal value of 1, but the average predicted risk remained too low. The authors noted that changes in case mix, unmeasured surgical or postoperative factors, outcome ascertainment, or relationships between predictors and outcomes may have contributed. The later period also included more robotic procedures and differences in estimated blood loss and nerve-sparing practice.

An exploratory binary cutoff derived from development predictions did not transport as reliably as the model’s continuous risk ranking. At a predicted probability of 0.128, temporal sensitivity was 0.500 and specificity was 0.913. The positive predictive value was 0.339, the negative predictive value was 0.954, and the F1 score was 0.404. In internal validation, sensitivity at the same fixed threshold had been 0.824, demonstrating that the cutoff was sensitive to changes over time. The threshold was therefore considered descriptive rather than a validated clinical triage rule, particularly because the later cohort contained only 38 outcome events.

During nested repeated cross-validation, LASSO numerically outperformed a deliberately constrained gradient boosting machine given the same 19 starting predictors. LASSO produced a ROC area under the curve of 0.913, compared with 0.896 for gradient boosting; its average precision was 0.542 rather than 0.488, and its Brier score was 0.046 rather than 0.048. The comparison supported choosing the sparse linear model for this dataset, but it did not establish universal superiority of one algorithm. With just 51 development events and 21 candidate predictor parameters before penalization, controlling model complexity was central to the analysis.

Repeated resampling also showed that the selected variables differed in their stability. Age, preoperative urinary symptom score, transformed PSA, and neurovascular bundle preservation appeared in all 1,000 bootstrap refits. Operative time was retained in 82.4%, postoperative hospital stay in 68.5%, smoking history in 65.7%, and hypertension in 65.4%. These frequencies described selection stability rather than determining the final predictor set. The penalized coefficients, including the apparently protective directions for longer operative time and postoperative stay, are components of a predictive equation rather than causal effects or conventional inferential odds ratios.

Decision-curve analysis suggested potential net benefit from using the model to select patients for intensified functional follow-up or pelvic-floor rehabilitation across examined risk thresholds of 2–30%, relative to treating everyone or no one. The analysis was retrospective and did not prospectively define an actionable threshold. Positive net benefit in this setting indicates possible decision value, not evidence that model-guided care improves continence, quality of life, adherence, resource use, or cost-effectiveness. Clinical thresholds would also need to account for the consequences of missing high-risk patients, the burden of additional support, patient preferences, and local rehabilitation capacity.

Several limitations constrain the findings. Both cohorts came from one Chinese clinical setting, and the model was not tested in another hospital, region, or surgical team. Selection bias cannot be excluded because only patients with ascertained 12-month outcomes and complete essential data were analyzed. The outcome threshold of an questionnaire score of at least 1 captured any residual self-reported symptom rather than a universally established diagnostic definition, and no parallel pad-based outcome was available for sensitivity analysis. A post hoc sample-adequacy assessment estimated that approximately 3,044 participants and 214 events would be needed in its primary conservative scenario, compared with the 726 participants and 51 events available for development.

The temporal cohort had also been examined during an earlier modeling stage for baseline descriptions, data quality, and preliminary performance, although those observations were not used to update or recalibrate the final model. The authors consequently classified the evaluation as same-center temporal validation rather than fully untouched prospective confirmation. Independent geographic validation, local calibration assessment, and prospective evaluation of clinical impact are needed before routine use. For now, the evidence supports treating the model as a graded risk aid for survivorship counseling and rehabilitation planning, not as a stand-alone gatekeeper to services or a replacement for clinical judgment.

Subject of Research: Temporal validation of a perioperative prediction model for persistent urinary incontinence after radical prostatectomy

Article Title: Risk prediction for 12-month post-prostatectomy urinary incontinence to support survivorship care: development and same-center temporal validation of a perioperative model

Article References: Tang, L., Hao, M., Luo, M., & Li, S. (2026). Risk prediction for 12-month post-prostatectomy urinary incontinence to support survivorship care: development and same-center temporal validation of a perioperative model. Frontiers in Public Health, 14, Article 1968281. https://doi.org/10.3389/fpubh.2026.1968281

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1968281

Keywords: prostatectomy, urinary incontinence, risk prediction, LASSO, temporal validation, calibration, survivorship care, pelvic-floor rehabilitation, prostate cancer, clinical prediction models, Risk, prediction