Prediction of depression for inpatients with chronic heart failure

Abstract

Depression is a common and prognostically important comorbidity of chronic heart failure (CHF).
Because datasets from cohort studies in this research area are often small and strongly imbalanced, and a ready-to-apply training and evaluation procedure for binary classification in this setting is still lacking, we provide a rigorous methodological overview and elaborate it into a straightforward protocol: leave-one-out cross-validation (LOOCV) with fixed seeds and default library hyperparameters, and a full panel of confusion matrix metrics, including imbalance-adjusted summaries. As an illustration of the chosen methodology, we reformulated the 9-item Patient Health Questionnaire (PHQ-9) screening as a binary classification task and compared three deliberately chosen standard methods---ridge-regularized logistic regression, decision tree, and RUSBoost---on a public cohort published by Cheng and co-authors in 2023. These methods are evaluated at each learner's default operating point over the range of PHQ-9 binarization thresholds $\theta\in\{1,\ldots,15\}$.

RUSBoost obtained the best results at the standard diagnostic PHQ-9 cut-off $\theta = 10$---sensitivity 0.641, specificity 0.619, accuracy 0.621 and a sensitivity--specificity gap of $0.022$. As the only learner with built-in imbalance handling, it also consistently outperforms the two observation-weighted baselines on these metrics across the entire range of thresholds. However, absolute sensitivity and specificity remain below clinically ambitious levels for all three models. We attribute part of this performance ceiling to selection bias introduced during cohort acquisition and eligibility filtering.