Answers are grounded in this paper and cite their sources.
Low-cost and accessible tools for preliminary pneumonia screening remain particularly important in resource-limited settings. This study aimed to develop and evaluate a machine-learning model with post-hoc SHAP-based explanations using routine hematological markers to distinguish patients with pneumonia from healthy controls. We retrospectively analyzed data from 4,389 participants, including 1,671 patients with pneumonia and 2,718 healthy controls, from two medical centers. The cohort was partitioned at the patient level into training, validation, and held-out testing sets. Least absolute shrinkage and selection operator regression was used for initial feature reduction, after which 15 candidate machine-learning base learners were developed, tuned, and evaluated. Four class-balancing strategies—synthetic minority oversampling technique, random oversampling, random undersampling, and random over-sampling examples—were compared within the training workflow in terms of feature-distribution preservation and downstream predictive performance. Entropy-weighted TOPSIS was used to rank the base learners, and TOPSIS-weighted averaging, stacking, and Super Learner strategies were compared across progressively expanded combinations of top-ranked learners. An outer five-fold cross-validation analysis was performed to assess robustness. Model performance was further evaluated using calibration and decision curve analyses, while post-hoc model explanations were generated using SHapley Additive exPlanations and a SHAP-informed surrogate nomogram. Synthetic minority oversampling technique showed the strongest overall downstream performance and was retained for the final modelling workflow. Among the candidate ensemble strategies, the Super Learner with a random forest meta-learner achieved the highest validation AUC of 0.9749 and was selected as the final model. On the held-out testing set, the final model achieved an AUC of 0.9822 (95% CI, 0.9744–0.9891), accuracy of 0.9431 (95% CI, 0.9282–0.9579), sensitivity of 0.9401 (95% CI, 0.9143–0.9641), specificity of 0.9449 (95% CI, 0.9253–0.9631), positive predictive value of 0.9128, negative predictive value of 0.9625, F1-score of 0.9263, and Brier score of 0.0450 (95% CI, 0.0360–0.0549) at the Youden-index-derived threshold of 0.50. In outer five-fold cross-validation, the prespecified final workflow achieved a mean AUC of 0.9612 ± 0.0077 and an accuracy of 0.9071 ± 0.0172. SHAP analysis identified mean corpuscular hemoglobin concentration, lymphocyte percentage, lymphocyte-to-monocyte ratio, platelet distribution width, and mean platelet volume as the most influential predictors. The Super Learner demonstrated strong discrimination and calibration for distinguishing patients with pneumonia from healthy controls using routinely available hematological markers. The model may provide a low-cost tool for preliminary pneumonia screening and risk stratification in resource-limited settings. However, because the comparator group consisted of healthy controls, prospective validation in symptomatic patients with clinically relevant non-pneumonia respiratory or inflammatory conditions is required before the model can be applied to differential diagnosis.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: