Abstract
Purpose :
Open Globe Injuries (OGI) are a predominant cause of vision loss worldwide. There are still persisting challenges of predicting visual outcomes despite advancements in identification of prognostic factors and treatment protocols. This study investigates the efficacy of machine learning techniques, such as Random Forest (RF), Decision Tree (DT), Gradient Boosting (GB), and Gaussian Naive Bayes (Gaussian NB), in predicting six-month visual outcomes post-OGI and determine critical factors influencing these outcomes.
Methods :
This retrospective study reviewed cases diagnosed with 'S05 – Injury of eye and orbit' or 'H5.5x – Retained (old) foreign body', undergoing surgical eye procedures between 2015 and 2021. Data collected included demographics, injury details, exam results, surgical interventions, antibiotic usage, and follow-up data up to six months. Machine learning models (RF, DT, GB, Gaussian NB) were developed to predict visual outcomes and were evaluated and compared using Area Under the Curve (AUC) benchmarks. A 70/30 split was used for training and testing data. Feature importance was analyzed in the most effective model.
Results :
From 199 identified patients, 179 were included in the study, predominantly male (77.1%) with an average age of 44.6 ± 23.7 years. Sixty-nine patients (39%) had comprehensive six-month follow-up data. The RF model outperformed others, achieving an AUC of 0.81, followed by GB (0.71), DT (0.68), and Gaussian NB (0.65). Notable features impacting visual outcomes were visual acuity at the first post-discharge follow-up (0.17), total antibiotic doses while inpatient (0.10), initial visual acuity (0.08), wound size (0.07), and time from ophthalmologic consult to antibiotics (0.06).
Conclusions :
Our results highlight the considerable potential of machine learning techniques, especially Random Forest (RF), in predicting visual outcomes for OGI patients. The identification of key predictors can significantly enhance vision-saving decision-making and treatment planning. When integrated into clinical practice, these predictive models present a substantial opportunity to improve patient care by establishing evidence-based expectations for visual recovery. This approach marks a significant advancement in the application of data-driven methods to refine and optimize treatment strategies for open globe injuries.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.