Risk stratification and determinant identification of high-need, high-cost ICU patients using machine learning: a large-scale retrospective study from a multi-specialty ICU in a tertiary hospital.

Background: Intensive care units (ICU) account for a disproportionate share of hospital costs. Retrospectively identifying high-need, high-cost (HNHC) ICU patients using machine learning (ML) may inform structured cost auditing and more efficient healthcare resource allocation. Methods: This retrospective study included adult patients with ICU admission (≥24 h) from multiple specialty ICUs in a Chinese tertiary hospital (2018-2024). HNHC patients were defined as the top 5% of annual ICU costs. Clinical, laboratory, and resource-use variables were extracted and preprocessed. Six ML models were developed using feature selection, class balancing, cross-validation, and Bayesian optimization, with performance evaluated by area under the receiver operating characteristic curve (AUC) and related metrics. Shapley Additive Explanations (SHAP) was applied for model interpretability. Results: Among 51,056 ICU patients, 2,556 (5.0%) were classified as HNHC. HNHC patients had longer ICU stays, greater disease complexity, and substantially higher use and duration of intensive therapies, resulting in markedly increased total and ICU-related costs. After correlation and Boruta feature selection, 19 variables were retained for model development. Among six ML models, the random forest (RF) achieved the highest discriminative performance in the independent test set, with an AUC of 0.942 (95% CI 0.931-0.952), followed by Extra Trees and LightGBM. The random forest model showed a favorable balance between sensitivity and F1 score in this highly imbalanced population. Decision curve analysis demonstrated stable net benefit across clinically relevant threshold probabilities. SHAP interpretation identified ICU length of stay and mechanical ventilation duration as the strongest contributors, revealing pronounced nonlinear effects, while diagnosis-related group (DRG) reform did not substantially alter the contribution patterns of key features. Stratified analyses confirmed that model performance and feature contribution patterns remained stable across the pre- and post-DRG reform periods. Conclusion: High ICU costs were primarily associated with intensive resource use rather than demographics alone. A RF model reliably classified HNHC patients and remained stable across DRG reform, supporting its use as a tool for retrospective risk stratification, identification of cost drivers, and more efficient allocation of critical care resources.