Machine learning prediction models and nomogram for identifying early post-COVID interstitial lung abnormalities in older adults

Post-COVID-19 interstitial lung abnormalities (PC-ILA) have emerged as an important component of post-COVID pulmonary sequelae in older adults, though their associated factors remain poorly defined. In this retrospective study, 349 older adults hospitalized with COVID-19 between December 2022 and June 2024 were followed after discharge, and 265 were included in the final analysis. Patients were categorized into a PC-ILA group ( n = 96) and a comparison group ( n = 169). Laboratory data and Comprehensive Geriatric Assessment (CGA) results obtained at admission were collected. The cohort was first divided into training and validation datasets, and candidate predictors were identified exclusively within the training dataset using baseline comparisons and LASSO regression. Three machine learning models (logistic regression, random forest, and XGBoost) were established and visualized using nomogram or SHapley Additive exPlanations (SHAP) analysis, with performance evaluated via assessment metrics, Receiver Operating Characteristic (ROC) curve analysis with Area Under the Curve (AUC) calculation, calibration curve plotting, and clinical decision curve analysis. In the final analytic cohort, 96 of 265 patients (36.2%) met the study definition of PC-ILA. Eight candidate predictors were identified through differential analysis and LASSO regression, including C-reactive protein (CRP), Interleukin-6 (IL-6), IL-1β, Tumor Necrosis Factor-α (TNF-α), lactate dehydrogenase (LDH), the Mini Nutritional Assessment Scale-Short Form (MNA-SF) score, the Activities of Daily Living (ADL) score and the Fried Frailty Scale score. The logistic regression model showed the most favorable overall predictive performance in the validation dataset, with the highest AUC (0.747, 95% confidence interval: 0.639–0.856) and a relatively favorable balance across accuracy, sensitivity, and F1 score. LDH, frailty, and inflammatory markers were identified as important predictors of PC-ILA. We developed an exploratory prediction framework integrating laboratory and CGA results to assess PC-ILA risk in older COVID-19 patients. The models may support early risk stratification and individualized follow-up.