The predictive value of malnutrition on the prognosis of severe respiratory failure in elderly patients: a multicenter retrospective study based on interpretable machine learning.

BACKGROUND: Elderly patients (≥65 years) with respiratory failure in the ICU have high mortality. Malnutrition worsens outcomes but is often overlooked. This study investigates nutritional status as a predictor of 28-day mortality and develops an interpretable machine learning-based model for early risk stratification. MATERIALS AND METHODS: This multicenter retrospective study enrolled elderly patients (≥65 years) with respiratory failure. The internal cohort came from the MIMIC-IV database, and the external validation cohort from Binzhou Medical University Hospital. Multivariable Cox regression analyzed the associations of the Prognostic Nutritional Index (PNI), Hemoglobin-Albumin-Lymphocyte-Platelet (HALP) score, and Geriatric Nutritional Risk Index (GNRI) with 28-day all-cause mortality. Restricted cubic spline and Kaplan-Meier curves explored dose-response relationships and survival differences. The internal cohort was randomly split into training (70%) and testing (30%) sets. The Boruta algorithm combined with LASSO regression selected prognostic features from laboratory tests, vital signs, and demographics. Six machine learning algorithms were built and validated by five-fold cross-validation. Model performance was assessed using calibration curves and decision curve analysis. SHAP was used for interpretability. RESULT: Among 1,385 patients, the 28-day mortality was 29.03%. In fully adjusted Cox models, all three indices as continuous variables were significantly inversely associated with mortality risk (PNI: HR = 0.966, 95% CI: 0.952-0.979, p < 0.001; HALP: HR = 0.994, 95% CI: 0.990-0.998, p = 0.004; GNRI: HR = 0.992, 95% CI: 0.986-0.998, p = 0.009). Severe malnutrition (vs. normal) was associated with 33-49% higher risk (all p < 0.05). Linear dose-response relationships were confirmed (overall p < 0.05, nonlinearity p > 0.05). The random forest model achieved the best performance, with AUCs of 0.807 (training), 0.754 (internal test), and 0.734 (external validation). SHAP identified low PNI, elevated BUN, LDH, and RDW as key predictors. CONCLUSION: PNI, HALP, and GNRI are independent linear predictors of 28-day mortality in elderly respiratory failure patients. The interpretable machine learning model provides a proof-of-concept tool for risk stratification. However, its low sensitivity limits its current clinical utility, and further optimization is required before clinical implementation.