Explainable Machine Learning Predictions of Perceptual Sensitivity for Retinal Prostheses.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Ophthalmology

Background:

• Retinal prostheses require precise calibration to individual perceptual thresholds for effective visual stimulation.
• Existing models lack accuracy in predicting these thresholds, which vary significantly across subjects, electrodes, and time.
• Factors like electrode-retina distance and impedance are known influences, but a comprehensive predictive model is needed.

Approach:

• Developed and applied machine learning (ML) models to a large longitudinal dataset to predict individual electrode thresholds and deactivation.
• Utilized explainable artificial intelligence (XAI) to identify the most influential predictors of perceptual sensitivity.
• Incorporated stimulus, electrode, and clinical parameters as predictors in the ML models.

Key Points:

• ML models explained up to 77% of the variance in perceptual threshold responses.
• Achieved high predictive performance for electrode deactivation (F1 score up to 0.740, AUC up to 0.913).
• Identified novel predictors including subject age, time since blindness onset, and electrode-fovea distance.

Conclusions:

• Routinely collected clinical data and a single system fitting session can inform an XAI-based threshold prediction strategy.
• This approach has the potential to significantly transform clinical practice in predicting visual outcomes for retinal prosthesis users.
• Personalized prediction models can optimize stimulation levels, leading to improved functional vision restoration.