Automated Machine Learning Prediction of Visual Acuity from Preoperative OCT Images After Macular Hole Surgery

Abstract

Purpose : Analysis of baseline optical coherence tomography (OCT) in patients with macular hole (MH) can provide insight into surgical success and postoperative visual acuity (VA). Automated machine learning (AutoML) is a new area in artificial intelligence research. We designed a machine learning model using AutoML to predict the visual outcome from preoperative OCT images in eyes with successful MH closure following vitrectomy.

Methods : We developed a single-label classifier using AutoML (Google Cloud AutoML Vision) to predict the visual outcome (“<70 letters” vs. “≥70 letters”; Snellen equivalent: 20/40) at 6 months in eyes with successful primary surgery for idiopathic MH. We used retrospective data obtained from consecutive eyes with successful primary surgery for idiopathic MH between 2014 and 2018 at the Centre Hospitalier Universitaire de Québec – Université Laval (Canada). We included a single eye per patient and excluded eyes with ocular comorbidities with a potential detrimental effect on VA. Baseline horizontal fovea-centered high-definition 30-degree OCT scans were obtained in all patients using the Cirrus HD-OCT 5000 machine (Carl Zeiss Meditec, Germany). VA at 6 months was measured in Snellen and converted to ETDRS letters for analysis. Eyes were divided into two groups based on VA at 6 months: <70 letters and ≥70 letters. During model development, 80% of images were used for training, 10% for the validation process, and 10% were used for evaluating the model. Precision, recall and area under the precision-recall curve (AuPRC) were used to evaluate the performance of the model.

Results : The dataset included 383 patients (383 eyes), 69% (263/383) of which were women. The mean age was 68 ± 8 years and 24% (91/383) of eyes were pseudophakic. Baseline VA was 51 ± 14 letters and 49% (187/383) of eyes had VA of 70 letters or more at 6 months. The model correctly classified postoperative VA 68% of the time for the “<70 letters” category and 79% of the time for the “≥70 letters” category. The model had a precision (positive predictive value) of 73.68%, a recall (sensitivity) of 73.68%, and an AuPRC of 0.799.

Conclusions : The development of a deep learning model using AutoML for the prediction of postoperative VA from baseline OCT images in eyes with MH is feasible. The AutoML model displayed good discriminative performance, but more research is needed to determine external validity.

This is a 2021 ARVO Annual Meeting abstract.