Explainable Machine Learning for Vitamin A Deficiency Classification in Schoolchildren

Abstract

Vitamin A deficiency is one of the leading causes of visual impairment globally. While blood tests are common approaches in developed countries, various socioeconomic and public perspectives render this a challenge in developing countries. In Africa and Southeast Asia, the alarming rise of preventable childhood blindness and delayed growth rates has been dubbed as an “epidemic”. With the proliferation of machine learning in clinical support systems and the relative availability of electronic health records, there is the potential promise of early detection, and curbing ocular complication progression. In this work, different machine learning methods are applied to a sparse dataset of ocular symptomatology and diagnoses acquired from Maradi, Nigeria collected during routine eye examinations conducted within a school setting. The goal is to develop a screening system for Vitamin A deficiency in children without requiring retinol serum blood tests, but rather by utilizing existing health records. The SVC model achieved the best scores of accuracy: 75.7%, sensitivity:83.7%, and specificity: 74.9%. Additionally, Shapley values are employed to provide post-hoc clinical explainability (XAI) in terms of relative feature contributions with each classification decision. This is a vital step towards augmenting domain expert reasoning, and ensuring clinical consistency of shallow machine learning models.