Rapid antifungal phenotypic detection of Candida tropicalis based on laser-scattering technology combined with machine learning.

BACKGROUND: Invasive fungal infections have become a significant cause of mortality among immunocompromised and critically patients. Candida tropicalis (C. tropicalis) has exhibited a marked increase inprevalence and a high rate of azole resistance as one of the major non C. albicans pathogens associated with bloodstream and deep infections. Therefore, the necessity for rapid and accurate susceptibility testing is paramount for the precise administration of treatment. The objective of this study was to develop a rapid prediction method for antifungal susceptibility based on early phenotypic signals (1-8 h) acquired through laser-scattering rapid culture technology. METHODS: Antifungal susceptibility testing was performed on 85 clinical isolates of C. tropicalis using Sensititre Yeast one YO10 panels to determine the minimum inhibitory concentrations (MICs). The assay for the detection of antifungal susceptibility for C. tropicalis clinical isolates was designed with laser-scattering technology. The model input features comprised laser-scattering intensity values (scattering units) and growth-curve parameters (slope, growth rate, maximal scattering intensity, area under the curve, etc.) collected from 1-8 h of drug exposure. Machine-learning models targeting fluconazole (FLZ), voriconazole (VOR) were constructed using random forest (RF), support vector machine (SVM), logistic regression (Logit), and stacking ensemble algorithms. The performance of the model was evaluated through five-fold cross-validation, with the area under the curve (AUC), sensitivity, specificity, accuracy, and F1-score being used as the evaluation metrics. RESULTS: The proportion of isolates demonstrating resistance to FLZ was 28.2% (24/85), while to VOR was 21.2% (18/85). With regard to the antifungal susceptibility prediction models, it was found that all machine-learning models achieved effective prediction within 6-8 hours of drug exposure. The FLZ and VOR models demonstrated the highest discriminative ability, with the Logit model for FLZ attaining an AUC of 0.923 and the SVM model for VOR achieving an AUC of 0.900. Notably, both models exhibited F1-scores that surpassed 0.88. Feature-importance analysis revealed that phenotypic signals from 5-8 h contributed most to prediction accuracy, reflecting the time-dependent nature of antifungal effects. CONCLUSIONS: Machine-learning models based on early phenotypic data have been shown to be capable of accurately predicting antifungal susceptibility within 8 hours.This represents a significant improvement over conventional methods and provides a promising technical framework for the development of rapid antifungal susceptibility testing and precision anti-infective therapy.