# Leveraging large scale deep learning models for diagnosis and visual outcome prediction in retinitis pigmentosa

## TL;DR

This study uses deep learning models trained on retinal images to improve diagnosis and predict visual outcomes in retinitis pigmentosa patients.

## Contribution

A novel hybrid model combining deep learning image features and clinical data for RP prognosis, showing superior performance in female patients.

## Key findings

- EfficientNetB4 achieved an AUC of 0.94 for RP diagnosis, especially effective in patients with good vision.
- A hybrid model combining imaging and clinical data outperformed single-modality models in visual prognosis.
- Prognostic performance was highest between 500 and 1400 days post-examination, with distinct features for diagnosis and prognosis.

## Abstract

Retinitis pigmentosa (RP) is an inherited progressive retinal degeneration that shows symptoms of night blindness, visual field loss, declining of vision and eventually, blindness. Currently, gene therapy and retinal prosthesis are available, but the indication for these treatments is limited. In this study, we report on the development of a diagnostic and prognostic model for RP based on large-scale deep learning (DL) models pre-trained with fundus images. The EfficientNetB4 model performed best in diagnosing RP with an AUC of 0.94. The diagnosis of RP with this model is superior in cases with good vision. For visual prognosis, we applied machine learning survival analysis to DL-derived image features and clinical metadata, using a strict patient-level split to avoid data leakage. The hybrid model combining imaging and clinical data outperformed models based on either modality alone, especially in female patients. Time-dependent AUC analysis showed that prognostic performance was highest between 500 and 1400 days after examination. SHAP-based interpretability analysis revealed that the features contributing to RP diagnosis and those associated with prognosis were distinct. While our findings demonstrate the added value of fundus images in visual outcome prediction, further validation using external and multi-center datasets is necessary for clinical translation.