Clinical predictors, species distribution, and machine learning–based risk stratification of invasive candidiasis and in-hospital mortality: a retrospective episode-level cohort analysis

Invasive candidiasis carries disproportionate mortality in hospitalised patients, yet the clinical determinants distinguishing invasive from non-invasive disease remain inconsistently characterised across healthcare settings. This study aimed to identify independent predictors of invasive candidiasis and in-hospital mortality, describe species distribution and antifungal resistance patterns, and evaluate machine learning models for risk stratification. A retrospective episode-level analysis was conducted on 206 laboratory-confirmed Candida -positive microbiological episodes at a single secondary care centre. Binary logistic regression univariable and multivariable was used to identify predictors of invasive candidiasis and all-cause mortality. Four machine learning classifiers (Random Forest, Gradient Boosting, Extra Trees, Logistic Regression) were developed and internally validated using five-fold stratified cross-validation. Permutation feature importance was applied for model explainability. Of 206 episodes, 37 (18.0%) were invasive. Overall, in-hospital mortality among invasive cases was 73.0%. Elderly age (≥ 65 years; aOR 53.22) and male sex (aOR 4.38) independently predicted invasive disease. Elderly age (aOR 5.12), male sex (aOR 4.96), and diabetes mellitus (aOR 2.60) were independent mortality predictors. Non-albicans species predominated among invasive isolates (73.0%), with Candida glabrata and Candida parapsilosis most over-represented. Fluconazole and caspofungin resistance were detected in 3/20 (15.0%) and 2/21 (9.5%) tested invasive isolates, respectively. The Random Forest classifier achieved AUROC values of 0.987 and 0.776 for invasive candidiasis and mortality prediction, respectively; the near-perfect invasive-candidiasis performance should be interpreted as exploratory and largely driven by marked age-based separation. Elderly age, male sex, and diabetes mellitus consistently predicted invasive disease and/or mortality in this episode-level cohort. Machine learning identified these routinely available variables as dominant discriminative features, but the models remain exploratory and require patient-level and external validation before clinical deployment.