Background Obstructive sleep apnea (OSA) is associated with excess mortality, and readily obtainable biomarkers may support pragmatic risk stratification in OSA, but their comparative and incremental value remains uncertain. We benchmarked inflammation-nutrition indices and TyG-related metabolic indices for mortality risk in adults with questionnaire-defined OSA and evaluated whether these markers improve 5-year all-cause mortality prediction beyond a basic clinical model. Methods We analysed 3,503 adults with questionnaire-defined OSA from the NHANES 2005–2008 and 2015–2018 derivation cohorts. Seven candidate biomarkers were evaluated: TyG, TyG-BMI, TyG-WC, TyG-WHtR, TG/HDL-C, advanced lung cancer inflammation index (ALI), and neutrophil percentage-to-albumin ratio (NPAR). Survey-weighted Cox models and restricted cubic splines were used to characterise mortality associations. Two 5-year all-cause mortality prediction strategies were developed: a tertile-based model (Model 1) and a continuous-scale model (Model 2). Internal validation used bootstrap optimism correction, calibration curves, Brier scores, and decision-curve analysis. Incremental value beyond a prespecified base model was assessed using likelihood-ratio testing, change in AUC, and net reclassification improvement (NRI). External validation was performed in an independent multicenter Chinese cohort of 200 patients from six hospitals. All included patients had at least 5 years of observation time from baseline assessment, allowing complete ascertainment of the binary 5-year all-cause mortality endpoint. The externally validated object was the final Base + Combine model, and its performance was assessed alongside the base model using ROC analysis, calibration plots, calibration intercept and slope, Brier score, and decision-curve analysis. Results During a median follow-up of 57.0 months (IQR 33.0–150.0), 293 all-cause deaths occurred in the derivation cohort. Among individual biomarkers, ALI showed the most robust mortality-related signal across analyses, whereas NPAR showed signal in selected single-marker and nonlinear analyses but was less consistent after full multivariable adjustment. TyG-related indices were also variably associated with mortality and contributed mainly within the combined prediction model. In Model 1, the full 7-marker composite model achieved an AUC of 0.735, a bootstrap-corrected AUC of 0.721, and a Brier score of 0.039. In Model 2, the best-performing combined model incorporated TyG-BMI, TyG-WC, TyG-WHtR, TG/HDL-C, and ALI, yielding an AUC of 0.765, a bootstrap-corrected AUC of 0.761, and a Brier score of 0.038. Internal calibration was acceptable for both derivation models, with Model 2 performing better. Decision-curve analysis showed positive net benefit over treat-all and treat-none strategies, with a wider clinically useful threshold range for Model 2. Compared with the base model, the combined biomarker model improved model fit (likelihood-ratio test p < 0.001) and reclassification (NRI p < 0.001), although the increase in AUC was modest. In the external validation cohort, the prespecified final Base + Combine model achieved an AUC of 0.697 (95% CI 0.581–0.813), compared with 0.672 (95% CI 0.553–0.791) for the base model. External calibration remained imperfect for both models. The base model showed a calibration intercept of −1.609, a calibration slope of 0.385, and a Brier score of 0.138, whereas the final Base + Combine model showed a calibration intercept of 3.782, a calibration slope of 0.402, and a Brier score of 0.089. In external decision-curve analysis, the base model provided greater net benefit at lower threshold probabilities, whereas the final Base + Combine model showed greater net benefit mainly within a narrower higher-threshold range (approximately 0.14–0.30). Given the limited number of external events, these findings should be interpreted as preliminary evidence of transportability rather than as definitive support for a clearly superior prediction tool. Conclusions In adults with questionnaire-defined OSA from NHANES, inflammation-nutrition markers, particularly ALI, showed stronger mortality-related signal than TyG-related indices at the single-marker level, while NPAR showed supportive signal in selected analyses. More importantly, a combined inflammatory-metabolic biomarker model provided modest incremental enrichment for 5-year all-cause mortality risk benchmarking beyond routine clinical variables. In an external multicenter cohort with PSG-confirmed OSA, the final Base + Combine model showed moderate but preliminary transportability, with slightly improved discrimination but poor calibration. Any incremental net benefit in external decision-curve analysis appeared to be confined to a relatively narrow higher-threshold range. These findings support pragmatic risk benchmarking rather than immediate use as a clearly superior clinical prediction tool.

Benchmarking inflammation-nutrition and TyG-related indices for 5-year mortality risk in adults with questionnaire-defined obstructive sleep apnea: a survey-weighted NHANES derivation cohort with multicenter external validation

期刊论文
出版模式：

开放获取
发表日期：

2026年7月10日

查看全文

Jian Liu,
Huiqian Yang,
Zheng Wang,
Shaban Ahmad,
Shuang Zhou,
Lingling Zhou,
Dan Lou,
Zhiqiang Guo,
Hui Li,
Xiaoming Li

摘要

Background

Obstructive sleep apnea (OSA) is associated with excess mortality, and readily obtainable biomarkers may support pragmatic risk stratification in OSA, but their comparative and incremental value remains uncertain. We benchmarked inflammation-nutrition indices and TyG-related metabolic indices for mortality risk in adults with questionnaire-defined OSA and evaluated whether these markers improve 5-year all-cause mortality prediction beyond a basic clinical model.

Methods

We analysed 3,503 adults with questionnaire-defined OSA from the NHANES 2005–2008 and 2015–2018 derivation cohorts. Seven candidate biomarkers were evaluated: TyG, TyG-BMI, TyG-WC, TyG-WHtR, TG/HDL-C, advanced lung cancer inflammation index (ALI), and neutrophil percentage-to-albumin ratio (NPAR). Survey-weighted Cox models and restricted cubic splines were used to characterise mortality associations. Two 5-year all-cause mortality prediction strategies were developed: a tertile-based model (Model 1) and a continuous-scale model (Model 2). Internal validation used bootstrap optimism correction, calibration curves, Brier scores, and decision-curve analysis. Incremental value beyond a prespecified base model was assessed using likelihood-ratio testing, change in AUC, and net reclassification improvement (NRI). External validation was performed in an independent multicenter Chinese cohort of 200 patients from six hospitals. All included patients had at least 5 years of observation time from baseline assessment, allowing complete ascertainment of the binary 5-year all-cause mortality endpoint. The externally validated object was the final Base + Combine model, and its performance was assessed alongside the base model using ROC analysis, calibration plots, calibration intercept and slope, Brier score, and decision-curve analysis.

Results

During a median follow-up of 57.0 months (IQR 33.0–150.0), 293 all-cause deaths occurred in the derivation cohort. Among individual biomarkers, ALI showed the most robust mortality-related signal across analyses, whereas NPAR showed signal in selected single-marker and nonlinear analyses but was less consistent after full multivariable adjustment. TyG-related indices were also variably associated with mortality and contributed mainly within the combined prediction model. In Model 1, the full 7-marker composite model achieved an AUC of 0.735, a bootstrap-corrected AUC of 0.721, and a Brier score of 0.039. In Model 2, the best-performing combined model incorporated TyG-BMI, TyG-WC, TyG-WHtR, TG/HDL-C, and ALI, yielding an AUC of 0.765, a bootstrap-corrected AUC of 0.761, and a Brier score of 0.038. Internal calibration was acceptable for both derivation models, with Model 2 performing better. Decision-curve analysis showed positive net benefit over treat-all and treat-none strategies, with a wider clinically useful threshold range for Model 2. Compared with the base model, the combined biomarker model improved model fit (likelihood-ratio test p < 0.001) and reclassification (NRI p < 0.001), although the increase in AUC was modest. In the external validation cohort, the prespecified final Base + Combine model achieved an AUC of 0.697 (95% CI 0.581–0.813), compared with 0.672 (95% CI 0.553–0.791) for the base model. External calibration remained imperfect for both models. The base model showed a calibration intercept of −1.609, a calibration slope of 0.385, and a Brier score of 0.138, whereas the final Base + Combine model showed a calibration intercept of 3.782, a calibration slope of 0.402, and a Brier score of 0.089. In external decision-curve analysis, the base model provided greater net benefit at lower threshold probabilities, whereas the final Base + Combine model showed greater net benefit mainly within a narrower higher-threshold range (approximately 0.14–0.30). Given the limited number of external events, these findings should be interpreted as preliminary evidence of transportability rather than as definitive support for a clearly superior prediction tool.

Conclusions

In adults with questionnaire-defined OSA from NHANES, inflammation-nutrition markers, particularly ALI, showed stronger mortality-related signal than TyG-related indices at the single-marker level, while NPAR showed supportive signal in selected analyses. More importantly, a combined inflammatory-metabolic biomarker model provided modest incremental enrichment for 5-year all-cause mortality risk benchmarking beyond routine clinical variables. In an external multicenter cohort with PSG-confirmed OSA, the final Base + Combine model showed moderate but preliminary transportability, with slightly improved discrimination but poor calibration. Any incremental net benefit in external decision-curve analysis appeared to be confined to a relatively narrow higher-threshold range. These findings support pragmatic risk benchmarking rather than immediate use as a clearly superior clinical prediction tool.