Multidimensional characteristics and associated factors of anorexia among older adults with chronic heart failure: an explainable machine learning approach

Anorexia is a common nutrition-related symptom in older adults and is associated with malnutrition, functional decline and adverse clinical outcomes. Older patients with chronic heart failure (CHF) are especially likely to experience anorexia, given the combined burden of disease, symptoms, biological abnormalities and psychological factors. The multidimensional characteristics and associated factors of anorexia in this population, however, remain incompletely described. To characterize the multidimensional features of anorexia and to identify factors associated with it among older adults with CHF, using an explainable machine learning approach informed by the Symptom Science Model 2.0 (SSM 2.0). This single-center cross-sectional study enrolled 400 hospitalized older adults with CHF; 280 (70.0%) screened positive for anorexia on the Simplified Nutritional Appetite Questionnaire (SNAQ ≤ 14). Participants were divided 7:3 into a training set ( n = 285) and a test set ( n = 115). Ten-fold cross-validated LASSO with minimum binomial deviance selected 11 of 66 candidate variables; because the selection of sleep disturbance rested on a single undocumented response code, coding-sensitivity analyses were also run and retained 10 variables. Six algorithms were compared using these 11 variables, with all preprocessing estimated within resampling folds, and repeated nested cross-validation in the training set identified logistic regression as the primary model. The test set, which had been examined during earlier development, was used only for exploratory evaluation at fixed training-derived cutoffs. Analyses were performed in R 4.3.3. The training-set LASSO selected 11 variables, but sleep disturbance entered through a single undefined response category; recoding or excluding that record reduced the set to 10. In repeated nested cross-validation within the training set, logistic regression (LR) had the highest mean AUC (0.868), followed by random forest (RF; 0.866), a difference too small to indicate superiority. In the test set, LR reached an AUC of 0.844 (95% CI 0.760–0.929) and RF 0.869 (0.806–0.932). At the training-derived LR cutoff of 0.618, sensitivity was 0.855, specificity 0.688, precision 0.877 and negative predictive value 0.647; the corresponding RF values at its cutoff of 0.677 were 0.771, 0.750, 0.889 and 0.558. In the primary LR model, only kinesiophobia and alcohol consumption history had confidence intervals excluding 1; these are post-selection estimates and should not be read as confirmed associations (Supplementary Table S6a). In exploratory RF SHAP analysis, kinesiophobia, self-compassion, heart failure symptom burden, fatigue and creatinine showed the largest mean absolute contributions. In this single-center sample of hospitalized older adults with CHF, anorexia screening positivity was common and co-occurred with psychological characteristics, symptom burden and selected laboratory markers, and these features contributed most in the exploratory RF model. Because the analysis had to be reconstructed from source data after the original script was lost, the test set had been examined previously, and the selection of sleep disturbance rested on one undocumented response code, these associations are hypothesis-generating and require independent prospective validation before clinical use. Chinese Clinical Trial Registry (ChiCTR2600119442) the registration was retrospective, completed on 27 February 2026, after the recruitment period had ended.