Development and internal validation of a nomogram and machine-learning models for postoperative recurrence in adult
Lin Song1,2, Xiaoli Zhang1,2, Mingliang Feng2,3
1Department of Clinical Laboratory, The Affiliated Yongchuan Hospital of Chongqing Medical University, Chongqing, China.
Background:
Practical preoperative tools for estimating postoperative recurrence in chronic rhinosinusitis with nasal polyps (CRSwNP) remain limited. We developed and internally validated a conventional nomogram and multiple machine-learning models and evaluated the systemic coagulation-inflammation index (SCI) as a deterministic nonlinear composite of routine laboratory measurements.
Methods:
The analysis included 245 adults. To preserve the integrity of the originally locked internal validation, the original outcome-stratified assignments were retained after age exclusions, leaving 170 patients in the training set and 75 in the same-center internal held-out set. Centered VIFs were calculated with an intercept, component-versus-composite SCI models were compared, and ten algorithms were tuned by five-fold stratified cross-validation in the training set. A Firth logistic sensitivity analysis forced hypertension and diabetes into the primary model.
Results:
Recurrence occurred in 95/245 patients (38.8%). SCI was the only initial candidate with VIF > 10 (10.259); after its structural exclusion, all remaining VIFs were ≤1.515. In the five-variable multivariable screening model, WBC was inversely associated with recurrence (OR 0.437, 95% CI 0.307-0.623), whereas PLT (OR 1.024, 95% CI 1.013-1.035) and FIB (OR 4.480, 95% CI 2.346-8.557) were positively associated (all p < 0.001). The final three-variable nomogram had a held-out AUC of 0.872 and Brier score of 0.149. Hypertension and diabetes data were available for 243/245 adults; forcing both comorbidities into a Firth sensitivity model did not materially change the three primary associations or improve held-out AUC (0.862 versus 0.872; DeLong p = 0.565). SCI did not show stable incremental held-out value beyond its components.
Conclusion:
A nomogram based on WBC, PLT, and FIB showed promising same-center internal performance in adults. SCI should be interpreted as a deterministic nonlinear composite rather than unique biological information. External validation in cohorts with standardized tissue histopathology and comprehensive comorbidity data is required before clinical use.