Background: 285 million people are visually impaired around the world and An estimated 153 million people had visual impairment (VI) from Uncorrected Refractive Error (URE) in 2004, and 8 million of them were blind. The L V Prasad Eye Institute is a 176 centre network spread across 4 states of India. eyeSmart EMR is a digital system implemented across the network and has clocked over 4 million consults in 7 years.
Purpose: Refractive error, is the most common human eye disorder in children and especially myopia has become a global health problem due to recent steep increase in its prevalence and its associated risk of visually-disabling pathologic myopia. Contrastingly, hyperopia although its prevalence is relatively less, in severe cases it is associated with of visual motor and sensory problems such as squint and amblyopia leading to visual impairment. However, there is still exists no tool to identify kids who are at risk of progressing to higher refractive error. We aim to develop a machine learning model for prediction of refractive error in children to reduce avoidable blindness.
Methods: A prediction model was developed using the Azure Machine Learning platform. The model was based on a two-point linear regression where the algorithm shows the impact on the predicting variable (refraction status i.e. sphere and cylinder) on changing
the explanatory variables such as sphere, cylinder, axis, age, gender, age of onset of refractive error, uncorrected and best corrected visual acuity, at the first and subsequent follow-up visits, and time duration between the visits. The eye and the ocular diagnosis also was taken into consideration as it also has a bearing in the pathogenesis of refractive error.
Results: The model was trained using the dataset obtained from about 50000 patients who at least visited twice across the L V Prasad Eye Institute network. The number of consults were 2.4 million. The age range of patients was 0 to 25 years including both males and females. Any patients where refractive error could not be recorded were not included in the training model. This model provides an L1 error of 0.4 and 0.34; and an R value of 0.92 and 0.78 for prediction of sphere and cylinder, r
espectively. Today, the prediction of the progression of the refractive error is displayed in the electronic medical record system in the hospital which enables us to identify children at risk for further management and treatment.
Conclusion: A novel robust model for refractive error prediction was developed and is currently in the validation phase. Considering high myopia as potential to cause blindness for which there exists no effective restorative treatment currently, this machine learning model that predicts the refractive error will be critical in identifying the ones at risk of developing high myopia, when integrated into clinical practice. Thus the proposed method would be beneficial for monitoring of refractive error, assessment of response to interventions and prognostication of risk for high myopia and can be used to develop and apply appropriate anti-myopia strategies in timely matter.