Acute infectious diseases represent a persistent public health burden that exerts sustained pressure on hospital bed capacity, treatment resources, and the allocation of the healthcare workforce. Strengthening hospital-level preparedness and resource planning requires reliable early-risk stratification tools that remain robust to real-world temporal shifts. However, many existing clinical prediction studies simplify inherently ordered outcomes into binary categories and rely on random data splits, limiting their relevance for real-world health system decision-making. In this study, we developed and evaluated an ordinal machine learning framework using clinical data from 5066 patients hospitalized with acute infectious diseases between 2022 and 2024. Recovery trajectories were modeled as an ordinal outcome, reflecting changes in status between admission and discharge. Models were trained on 2022–2023 data and externally evaluated on a fully isolated 2024 cohort to assess temporal generalizability under realistic deployment conditions. Performance was evaluated using order-aware metrics, including Quadratic Weighted Kappa, Macro-F1, Balanced Accuracy, and ordinal mean absolute error, with explicit analysis of clinically meaningful error structures. Although predictive performance under future holdout validation was modest, misclassifications were predominantly concentrated between adjacent recovery levels, and no clinically critical extreme errors were observed. Model reliability was further assessed through calibration analysis, bootstrap-based uncertainty estimation, and temporal stability of explanatory patterns. Finally, ordinal predictions were translated into structured risk stratification categories aligned with hospital bed management, treatment prioritization, and workforce allocation logic. These findings demonstrate the methodological potential of temporally validated ordinal modeling as a proof-of-concept framework. Given the modest predictive performance and the absence of key clinical variables, the current model should not be regarded as a ready-made clinical decision-support tool, but rather as a foundation for further development with richer data in future research. monitoring prioritization. In practical terms, this framework demonstrates how ordinal predictions could, in principle, be structured for use at admission points. However, given the modest predictive performance observed, further development with richer clinical data is required before deployment.
Sodnomdavaa et al. (Tue,) studied this question.