An interpretable model predicts visual outcomes of no light perception eyes after open globe injury

BACKGROUND: The visual outcome of open globe injury (OGI)-no light perception (NLP) eyes is unpredictable traditionally. This study aimed to develop a model to predict the visual outcomes of vitrectomy surgery in OGI-NLP eyes using a machine learning algorithm and to provide an interpretable system for the prediction results. METHODS: Clinical data of 459 OGI-NLP eyes were retrospectively collected from 19 medical centres across China to establish a training data set for developing a model, called 'VisionGo', which can predict the visual outcome of the patients involved and compare with the Ocular Trauma Score (OTS). Another 72 cases were retrospectively collected and used for human-machine comparison, and an additional 27 cases were prospectively collected for real-world validation of the model. The SHapley Additive exPlanations method was applied to analyse feature contribution to the model. An online platform was built for real-world application. RESULTS: The area under the receiver operating characteristic curve (AUC) of VisionGo was 0.75 and 0.90 in previtrectomy and intravitrectomy application scenarios, which was much higher than the OTS (AUC=0.49). VisionGo showed better performance than ophthalmologists in both previtrectomy and intravitrectomy application scenarios (AUC=0.73 vs 0.57 and 0.87 vs 0.64). In real-world validation, VisionGo achieved an AUC of 0.60 and 0.91 in previtrectomy and intravitrectomy application scenarios. Feature contribution analysis indicated that wound length-related indicators, vitreous status and retina-related indicators contributed highly to visual outcomes. CONCLUSIONS: VisionGo has achieved an accurate and reliable prediction in visual outcome after vitrectomy for OGI-NLP eyes.