Predicting sex from retinal fundus photographs using automated deep learning

Area of Science:

• Ophthalmology
• Medical Informatics
• Artificial Intelligence

Background:

• Deep learning models for healthcare often require specialized coding expertise, limiting clinician involvement.
• Automated machine learning (AutoML) offers a potential solution for enabling clinician-driven model development.

Purpose of the Study:

• To develop and validate a clinician-created, code-free deep learning model for predicting reported sex from retinal fundus photographs.
• To assess the model's performance and identify factors influencing its accuracy, such as foveal pathology.

Main Methods:

• A deep learning model was trained using the UK Biobank dataset (84,743 retinal fundus photos).
• The model's performance was evaluated through internal validation and external validation on data from a tertiary ophthalmic referral center (252 photos).
• Model explainability was considered due to the lack of known distinct retinal features between sexes.

Main Results:

• The code-free deep learning (CFDL) model achieved an area under the receiver operating characteristic curve (AUROC) of 0.93 for internal validation.
• Internal validation metrics included 88.8% sensitivity, 83.6% specificity, 87.3% positive predictive value (PPV), and 86.5% accuracy (ACC).
• External validation yielded 83.9% sensitivity, 72.2% specificity, 78.2% PPV, and 78.6% ACC. Performance decreased significantly with foveal pathology (ACC: 69.4% vs. 85.4% in healthy eyes).

Conclusions:

• Automated machine learning enables clinicians to develop deep learning models without coding, facilitating the discovery of novel insights.
• The CFDL model demonstrates potential for predicting sex from retinal images, though foveal pathology impacts performance.
• Model explainability is crucial for understanding predictions, especially when underlying biological differences are not well-established.