Applications of Machine Learning towards Global Screening, Diagnosis, and Treatment for Child Blindness
Author: Duccio Rocca
Faculty Supervisor: Anagha Kulkarni
Department: Computer Science
This project applies Machine Learning (ML) approaches to provide clinical diagnostic tools for Cerebral Visual Impairment (CVI) in children 4-18 years old. Cerebral visual impairment is the commonest cause of bilateral visual impairment in children, accounting for 40% of cases is often treatable. Standard medical examination for CVI is prohibitively expensive at scale and there is no current standard for screening. Questionnaires, such as the Higher Visual Function Question Inventory (HVFQI), have been shown to be inexpensive, non-invasive, and scalable diagnostic tools that can be remotely administered to detect CVI-related Higher Visual Functional Deficits (HVFDs). In this work, ML CVI-evidence classifier models are trained on anonymous HVFD profiles from 212 participants. Feature importance ranking identifies a subset of questions to suggest viable screeners. We developed and implemented a pipeline to select, train, and evaluate a series of models on recall (sensitivity), specificity, precision, and F2 score for CVI evidence prediction using the full question inventory and abbreviated screeners. Our top-performing screening models demonstrate significant efficacy across all metrics, achieving 94% sensitivity, and a 92.2% F2 score. Accurate remote screening and diagnostic tools fulfill a critical role towards providing global equitable treatment for child blindness.