Explainable machine learning predictions of perceptual sensitivity for retinal prostheses.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Ophthalmology

Background:

• Retinal prostheses restore vision by stimulating retinal cells, but require extensive system fitting due to variable perceptual thresholds.
• Accurate prediction of these thresholds is crucial for effective device calibration but remains a challenge.

Purpose of the Study:

• To develop machine learning models for predicting individual electrode thresholds and deactivation in retinal prosthesis users.
• To utilize explainable artificial intelligence (XAI) to identify key predictors of perceptual sensitivity.

Main Methods:

• Fitted machine learning models to a large longitudinal dataset, incorporating stimulus, electrode, and clinical parameters.
• Employed XAI techniques to analyze predictor importance for threshold prediction and electrode deactivation.

Main Results:

• Models explained up to 76% of perceptual threshold variance.
• Achieved F1 scores of 0.732 and AUC of 0.911 for predicting electrode deactivation.
• Identified subject age, time since blindness onset, and electrode-fovea distance as novel predictors.

Conclusions:

• Machine learning and XAI can significantly improve the prediction of visual outcomes in retinal prosthesis users.
• Routinely collected clinical data and system fitting may suffice for an effective XAI-based prediction strategy.
• This approach has the potential to transform clinical practice for retinal prosthesis calibration and patient care.