Multimodal Deep Learning for Differentiating Bacterial and Fungal Keratitis Using Prospective Representative Data.

Abstract

Objective: This study develops and evaluates multimodal machine learning models for differentiating bacterial and fungal keratitis using a prospective representative dataset from South India.

Design: Machine learning classifier training and validation study.

Participants: Five hundred ninety-nine subjects diagnosed with acute infectious keratitis at Aravind Eye Hospital in Madurai, India.

Methods: We developed and compared 3 prediction models to distinguish bacterial and fungal keratitis using a prospective, consecutively-collected, representative dataset gathered over a full calendar year (the MADURAI dataset). These models included a clinical data model, a computer vision model using the EfficientNet architecture, and a multimodal model combining both imaging and clinical data. We partitioned the MADURAI dataset into 70% train/validation and 30% test sets. Model training was performed with fivefold cross-validation. We also compared the performance of the MADURAI-trained computer vision model against a model with identical architecture but trained on a preexisting dataset collated from multiple prior bacterial and fungal keratitis randomized clinical trials (RCTs) (the RCT-trained computer vision model).

Main outcome measures: The primary evaluation metric was the area under the precision-recall curve (AUPRC). Secondary metrics included area under the receiver operating characteristic curve (AUROC), accuracy, and F1 score.

Results: The MADURAI-trained computer vision model outperformed the clinical data model and the RCT-trained computer vision model on the hold-out test set, with an AUPRC 0.94 (95% confidence interval: 0.92-0.96), AUROC 0.81 (0.76-0.85), accuracy 77%, and F1 score 0.85. The multimodal model did not substantially improve performance compared with the computer vision model.

Conclusions: The best-performing machine learning classifier for infectious keratitis was a computer vision model trained using the MADURAI dataset. These findings suggest that image-based deep learning could significantly enhance diagnostic capabilities for infectious keratitis and emphasize the importance of using prospective, consecutively-collected, representative data for machine learning model training and evaluation.

Financial disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.