Explainable Machine Learning Model Based on Routine Admission Laboratory Tests for Predicting New-Onset...

Malnutrition in older patients hospitalized for acute exacerbation of chronic obstructive pulmonary disease (AECOPD) is frequently overlooked. New-onset hypoalbuminemia during hospitalization is an important signal of worsening nutritional and inflammatory burden and is closely associated with adverse outcomes such as prolonged mechanical ventilation, readmission, and mortality. Early proactive warning tools based on admission data are lacking.

This dual-center retrospective cohort study included patients aged ≥65 years who were hospitalized for AECOPD between January 2023 and December 2025 and had normal serum albumin at admission (≥35 g/L). The Zigong cohort (n=1,502) was used for model development and internal validation, whereas the Qiannan cohort (n=1,086) served as the external test cohort. Among the Zigong cohort, 1052 patients were assigned to the training set and 450 to the internal validation set. The primary outcome was defined as new-onset hypoalbuminemia recorded in the discharge diagnosis during the same hospitalization among patients with normal admission albumin. Because the retrospective source database did not retain a standardized schedule for inpatient albumin re-testing, albumin measurement frequency and the median time from admission to outcome ascertainment could not be evaluated. To minimize information leakage, all feature-selection procedures were performed exclusively in the training set. Missing data were handled using multiple imputation. Candidate predictors were demographics and routine laboratory tests completed within 24 hours of admission. Core features were selected using the intersection of univariable screening, least absolute shrinkage and selection operator (LASSO) regression, and the Boruta algorithm. Logistic regression, decision tree, random forest, XGBoost, LightGBM, support vector machine, and artificial neural network models were developed with five-fold cross-validation and grid-search tuning. Discrimination, calibration, and decision-curve analysis were evaluated in an internal validation set and an external test set. Model interpretability was assessed with SHapley Additive exPlanations (SHAP), and the optimal model was deployed as an online risk calculator.

Among 1502 older AECOPD inpatients, 335 (22.3%) met the study definition of new-onset hypoalbuminemia by discharge. Eight core predictors available at admission were retained: cholinesterase (CHE), high-sensitivity C-reactive protein (hs-CRP), hematocrit (HCT), anion gap (AG), serum magnesium (Mg), alanine aminotransferase (ALT), age, and international normalized ratio (INR). XGBoost achieved AUCs of 0.85 in the internal validation set and 0.83 in the external test set, compared with 0.84 and 0.82, respectively, for logistic regression, indicating only a modest performance advantage. SHAP indicated that lower CHE, HCT, AG, and Mg and higher hs-CRP, age, and INR were associated with higher risk.

An interpretable model derived from routine admission laboratory tests may support early risk stratification for new-onset hypoalbuminemia in older hospitalized patients with AECOPD. Nevertheless, further prospective validation is required to confirm its clinical utility and generalizability.