From Species Identification to Empirical Therapy: A Machine Learning and Rule-Based Decision Support Framework for Antifungal Resistance Prediction in ICU Candida Infections

Objectives: When a Candida species is identified in an ICU patient, susceptibility results are typically available in 24–72 h. In this study, we built a machine learning model using four variables available at identification to estimate resistance probability in real time. Methods: We analysed 747 fungal isolates from 725 ICU patients (January 2021–March 2026). We trained and compared a Random Forest and a Logistic Regression model, evaluating both with temporal cross-validation, permutation feature importance, three-category (S/I/R) prediction, and calibration analysis. Results: Multidrug resistance doubled from 24.5% (2021) to 51.1% (2025), and Candida auris grew eight-fold in three years. Random Forest reached AUC 0.885 on the held-out test set and 0.848 on prospective 2024–2025 data (Brier score 0.093). Species identity and drug choice together explained 87% of predictive signal. Local C. albicans fluconazole resistance (~16%) far exceeded the ECMM European figure of 0%, and C. krusei was four times more prevalent than the continental average. Conclusions: A four-variable model may provide calibrated resistance estimates during the critical gap before susceptibility results return, though performance reflects predominantly deterministic species–drug patterns rather than complex learned biology. Overall performance was comparable to a rule-based lookup table, confirming that the majority of predictive signal derives from established species–drug susceptibility patterns. Meaningful added value is limited to temporal trend tracking and improved prediction where resistance is acquired rather than intrinsic (C. albicans, C. tropicalis hard-subset AUC 0.929 vs. rule-based 0.899). The model complements a local antifungal testing; it does not replace one.