Commentary: How useful is a deep learning smartphone application for screening for amblyogenic risk factors?

As elsewhere artificial intelligence (AI) seems to be replacing human endeavor. Recall the recent docking of the Space-X Dragon module with the International Space Station, Tesla's self-driving cars, or the targeted advertisements encroaching your view while browsing: all examples of AI at work.

AI is often spoken with machine learning (ML) and deep learning (DL): the former uses algorithms applied to huge data sets, to analyze and learn patterns to make informed predictions. In case of errors, human experts step in. DL, in contrast, uses a layered algorithmic structure, appropriately labelled an artificial neural network, which senses the inaccuracy if any and auto-course corrects.

As Ting recently pointed out, the advanced mathematical models along with access to big data, has permitted AI to enlarge its foot print into healthcare.[1] DL has demonstrated its capability in image, speech, and motion recognition, understandably impacting medical specialties like radiology, dermatology, and pathology. In ophthalmology, AI-based equipment has successfully shown its functionality while evaluating fundus images for glaucoma, and macular degeneration and diabetic retinopathy.[2] Li et al. developed and evaluated an OCT trained DL technique – OCTD-Net, to detect early DR.[3] They reported meaningful accuracy, sensitivity, and specificity of 0.92, 0.90, and 0.95 for grade 1, though not for 0. Importantly in comparison studies, for predicting glaucomatous optic neuropathy, Jamal et al. pitted a DL trained with RNFL-thickness parameters from SD-OCT against two glaucoma specialists: DL performed significantly better on Spearman's correlations with standard automated perimetry: roh of 0.54 Vs 0.48, at P < 0.001; and on partial AUC, for predicting GON: 0.529 vs 0.411, P = 0.016.[4]

In this context, the authors need to be commended in demonstrating the use of DL for screening for amblyogenic risk factors (ARFs) using an android based smartphone.[5] Yet it needs to stand upto ordinary digital cameras,[6] smartphones,[7] and even the retinoscope all using the Bruckner's reflex.[8] Moreover the study has no comparator group, and has perhaps merely demonstrated the capability of using this approach on 18–23 year-old optometry students: it needs to be whetted on 4–7 year olds, when anti-amblyopia measures can be effective. Interestingly as an exercise, if we draw up a 2 × 2 table, and plug in the sensitivity (88.2%) and specificity (75.6%) values, and imagine screening a thousand children, it yields a false positive rate of around 83%: 228 of 272 who would test positive. The magic of the paper lies in the novel approach using AI to demonstrate functionality: How well it will perform in a realistic environment remains a moot question.