Characteristics, cost/effect consideration of clinical examinations, and construction of machine learning models of restrictive cardiomyopathy: insights from Peking Union Medical College Hospital

We read with great interest the article by Chen et al, “Characteristics, cost/effect consideration of clinical examinations, and construction of machine learning models of restrictive cardiomyopathy: insights from Peking Union Medical College Hospital,” which combines clinical phenotyping, interpretable machine learning (ML), and a pragmatic laboratory-cost analysis to improve identification of amyloidosis among patients with restrictive cardiomyopathy (RCM)[1]. The authors compared eight ML algorithms and reported strong discrimination using leave-one-out cross-validation, with the random forest model achieving an AUC of 0.977 and SHAP-based explanations to support interpretability[1]. They further estimated that avoiding routine low-yield laboratory panels in patients without corresponding etiologies could reduce expenditure (up to 2990 yuan per patient), while prioritizing amyloidosis-focused testing[1]. As AI-supported clinical reporting expands, transparent disclosure of model development, validation, and intended use helps readers judge reproducibility and safety[2]. Here, we propose workflow-aligned clarifications and a staged testing framework to strengthen safe deployment and interpretation of the “cost/effect” message. First, the impressive apparent performance should be interpreted cautiously in light of limited-sample validation and potential information leakage. In small datasets, cross-validation can yield optimistically biased estimates if imputation, feature selection, or hyperparameter tuning are informed by the full dataset rather than contained within each resampling iteration[3]. Because the pipeline includes k-nearest-neighbor imputation and LASSO-based feature selection, it would strengthen confidence to state explicitly that every preprocessing and selection step was executed within each training split, ideally with nested tuning. We would appreciate clarification on whether KNN imputation, variable screening, and LASSO selection were performed fold-wise within each LOOCV training split, with tuning nested inside the split. Reporting calibration (intercept/slope and calibration plots) alongside discrimination and providing optimism-corrected estimates (e.g., bootstrap) would further support deployment decisions[4]. Second, outcome definition and reference standards shape both validity and transportability. The authors note that pathology – the gold standard – was not available for all patients, and amyloidosis subtype information was incomplete, with confirmed cases diagnosed as AL[1]. In rare-disease retrospective cohorts, partial verification and spectrum effects may materially influence both estimated discrimination and real-world consequences of misclassification. Prospective, adjudicated reference cohorts aligned with contemporary diagnostic pathways would help quantify generalizability across amyloidosis phenotypes[5]. As an interim step, reporting performance in patients with available pathology or predefined non-biopsy criteria, and conducting a sensitivity analysis that excludes uncertain labels, would be informative. Third, model inputs should align with the stewardship goal of reducing unnecessary testing. Notably, urine immunofixation electrophoresis appears among the most influential predictors in the final model[1]. However, immunofixation is itself part of the amyloidosis work-up that many centers reserve for higher-suspicion cases; if specialized tests are required to run the model, the opportunity to reduce those same tests is inherently constrained. A pragmatic staged strategy may help: an initial “pretest-probability” model using variables available before specialty testing (clinical features and routine imaging/laboratory data), followed – only above a prespecified risk threshold – by targeted monoclonal-protein testing and confirmatory pathways, including validated non-biopsy criteria where appropriate[6]. For triage, a threshold prioritizing sensitivity (to minimize missed amyloidosis) may be appropriate, with a second-stage confirmatory pathway to recover specificity. Fourth, the cost/effect component would be strengthened by a decision-analytic perspective that incorporates downstream consequences. We interpret the current results primarily as a cost-description and stewardship signal rather than a full cost-effectiveness evaluation. Immediate savings from deferring broad panels should be balanced against costs of diagnostic delay (false negatives), unnecessary follow-up (false positives), repeat admissions, and time to disease-modifying therapy. Extending the analysis into a formal economic evaluation with clearly stated perspective, time horizon, included resources, and outcomes – reported in line with CHEERS 2022 – would make the “cost/effect” message more robust and comparable across settings[7]. Pairing this with decision-curve analysis could quantify net benefit across clinically plausible thresholds, linking model output to patient-centered and system-level utility[8]. Overall, Chen et al provide an important proof-of-concept for interpretable ML-enabled diagnostic stewardship in RCM[1]. These additions would not alter the main message but would improve reproducibility, threshold-based actionability, and economic interpretability – thereby accelerating responsible clinical integration.