Trajectory-based prediction modeling for persistent high LARS after rectal cancer resection

Persistent high low anterior resection syndrome (LARS) severely impairs quality of life after sphincter-preserving surgery for rectal cancer, highlighting the urgent need for robust predictive tools. This prospective longitudinal study enrolled 208 patients who underwent rectal cancer resection. LARS scores were prospectively assessed at 1, 3, 6, and 12 months postoperatively. Latent growth mixture modeling (LGMM) was used to identify distinct LARS trajectories. The primary clinical objective was to identify patients with persistent High LARS; accordingly, trajectory groups were designated, and the 6-month Physical Component Summary (PCS) and Mental Component Summary (MCS) of the SF-36 were used to validate this classification. Ten machine learning algorithms and a clinical nomogram were developed to predict persistent High LARS. To identify patients with clinically meaningful persistent bowel dysfunction, we defined the High LARS Group as those with a High‑stable trajectory. The Low‑stable and Progressive improvement trajectories, which showed comparable HRQoL at 6 months, were combined into a Controlled LARS Group. This grouping was further validated by the significantly lower PCS and MCS scores in the High‑stable group versus the other two trajectories (all p < 0.001). Among the ten machine learning algorithms, Support Vector Machine (SVM) achieved the highest predictive performance (AUC = 0.811, 95% CI: 0.753–0.870). The logistic regression-based nomogram showed good discrimination (AUC = 0.766, 95% CI: 0.696–0.829) and clinical utility. The SVM achieved good predictive performance for persistent High LARS in this single‑center cohort, while the visually interpretable nomogram exhibited favorable interpretability and potential clinical practical value. Both tools show preliminary promise for supporting personalized risk stratification after sphincter‑preserving surgery, and their formal clinical utility requires confirmation in independent external cohorts.