Work

A machine-learning-benchmarked, ridge-penalised nomogram excluding oral-frailty items for oral-frailty risk stratification in older inpatients with lung cancer: development and internal validation

Year: 2026
Type: preprint
Study design: Observational Study
Abstract: Background Oral frailty is common among older inpatients with lung cancer, but published prediction tools embed diagnostic oral-frailty items as predictors, inducing circular reasoning and outcome leakage. This single-centre study develops and internally validates a transparent ridge-penalised nomogram from routine pre-examination admission variables and benchmarks it against six machine-learning classifiers. Methods Consecutive inpatients aged ≥ 60 years with histologically confirmed lung cancer were retrospectively enrolled at Guangxi Medical University Cancer Hospital (September 2024–April 2025). Oral frailty was defined as an Oral Frailty Index-8 (OFI-8) score ≥ 4, and six overlapping oral variables were excluded a priori. The 477 patients were stratified 7:3 (random_state = 42) into training (n = 333; 236 events) and test (n = 144; 102 events) sets. Five-fold cross-validated LASSO over a prespecified 44-term P1 pool selected 11 predictors at λ ~ 1se ~ = 7.543; a ridge-penalised logistic model (C = 6.866; λ = 0.146) was the primary model, and six further classifiers were fitted without class weights. Discrimination (AUC; 2000 bootstrap replicates), calibration and decision-curve analysis were assessed, with SHAP for XGBoost; three prespecified comparisons used Holm correction, and three sensitivity analyses reintroduced oral, taste/drinking and education variables. Results Oral frailty was prevalent in 70.9% (338/477). The nomogram, built on pre-examination variables excluding diagnostic items, achieved test AUC 0.888 (95% CI 0.827–0.942; training 0.867), with calibration slope 1.065, intercept 0.295 and Brier 0.122. XGBoost attained the highest AUC (0.959, 95% CI 0.920–0.988), followed by random forest (0.945, 95% CI 0.902–0.978). Both outperformed the nomogram (ΔAUC + 0.071 and + 0.057; Holm P = 0.020 each) but did not differ from each other (ΔAUC + 0.014, P = 0.156) and were less well calibrated. Reintroducing the six oral variables left XGBoost unchanged (0.950; Δ≈−0.009, ≈−0.9%), whereas the symmetric ridge re-run on the 50-term pool raised ridge AUC to 0.968 (Δ=+0.080; +9.0%) and retained five oral items, quantifying the excluded circularity. Conclusions A printable, interpretable nomogram built from admission variables available before any oral examination shows good discrimination and a near-ideal calibration slope, supporting nurse-led risk stratification; the higher-AUC ensembles serve as performance benchmarks. Owing to its positive intercept, absolute-risk estimates require site-specific recalibration, and multicentre external and temporal validation is the priority before bedside deployment.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7225907217
Cites: 52
Cited by:
Related to: 10