Interpretable multi-task prediction in rhabdomyolysis: promise and caveats

We read with great interest the recent article by Liu et al., who reported the development of interpretable multi-task machine learning (ML) models for patients with rhabdomyolysis (RM).Their approach simultaneously predicted acute kidney injury (AKI), disease severity, renal-replacement therapy (RRT), and inhospital mortality, derived from MIMIC-IV/eICU and externally validated in a four-hospital Chinese cohort.The authors further implemented compact five-variable models with web and Android prototypes and demonstrated SHAP-based interpretability, representing significant steps toward clinical translation. 1However, several methodological issues may compromise validity, robustness, and readiness for real-world application.First, the estimation of baseline serum creatinine (SCr) relied on back-calculation from an assumed eGFR of 75 mL/min/1.73m 2 .This practice, not formally recommended by KDIGO or the Acute Dialysis Quality Initiative, may misclassify both the onset and severity of AKI. 2 Such mislabeling directly affects supervised learning, potentially leading to distorted risk estimates and "optimistic" AUCs.As SCr was an important predictor across all four models, clarification of the proportion of patients with recalculated values and sensitivity analyses excluding these cases would be crucial.Second, the definition of "disease severity" is problematic.According to the numbers reported, patients meeting both AKI and RRT (n = 173) or death (n = 140) should total 313, whereas Table 1 lists 524 severe cases, suggesting inconsistency.More importantly, the definition of "disease severity" lacks clinical consensus, overlapping with AKI, RRT, and death.Without temporal restrictions, predictor variables collected after early AKI may inadvertently inform AKI prediction, causing label leakage.Survival or multi-state models would more appropriately capture disease trajectories. 3hird, missing data were handled through median imputation.While simple, this ignores informative missingness and underestimates uncertainty.