Early Postoperative Trajectories for Six-Month Outcome Prediction after Endoscopic Sinus Surgery: A Leakage-Safe Comparison of Transformer and Conventional Machine-Learning Models
Year: 2026
Type: preprint
Abstract: Early recovery after endoscopic sinus surgery may contain prognostic information that is not available preoperatively. We investigated whether trajectories measured through one postoperative month predict six-month outcomes in chronic rhinosinusitis and whether a trajectory-specific Transformer outperforms conventional models. This secondary analysis used 95 participants from a prospective randomized study (49 control; 46 biodegradable steroid-eluting stent). Predictors comprised baseline covariates and endoscopy, five-item visual analogue scale (VAS), and 22-item Sinonasal Outcome Test (SNOT-22) measurements at baseline, two weeks, and one month. Six-month continuous outcomes were predicted using repeated nested cross-validation (10 repetitions of five folds). For six-month SNOT-22 prediction from one-month data, Extra Trees achieved a mean absolute error (MAE) of 2.044 (95% CI 1.934–2.156), compared with 2.055 (1.953–2.149) for a clinically restricted ridge model and 2.624 (2.482–2.777) for the proposed EPT-Former. EPT-Former was significantly worse than the horizon-specific best comparator at every horizon (Holm-adjusted p < 0.001). Adding early postoperative observations improved all principal models, but tree and regularized linear models retained an advantage. Early trajectories were informative, whereas increased architectural complexity did not improve prediction in this small structured cohort. External validation is required before clinical use.
Source: Research Square
Authors: Jia Liu, Shibo Jin, Zhenyu Deng, Xueliang Zhang
Institution: TianjinSino-German University of Applied Sciences
Language: English
Subfield: Otorhinolaryngology
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7236733211
