Predicting ROP Severity by Artificial Intelligence: Pragmatic Versus Knowledge-Based Approach.

Abstract

Retinopathy of prematurity (ROP) blindness is largely preventable, with many countries having already established programs for its timely identification. However, ∼93% of infants screened do not develop ROP severe enough to require treatment.1 As reported in this issue of Pediatrics by Coyner et al.2 artificial intelligence (AI) is innovative methodology to optimize the timing of ROP examinations. For example, it can predict whether an eye will not require treatment, reducing the number of examinations needed while also predicting which eyes need closer surveillance. This is an important development because it would make ROP programs more effective and efficient, thus reducing the burden on infants, families, and the health care team.From the earliest descriptions, congestion and tortuosity of the posterior retinal vessels were recognized as prominent features of serious ROP.3 In the first international classification of ROP in 1984,4 the term “plus disease” was included to describe progressive vascular incompetence. Two decades later, after the recommendations of the Early Treatment for Retinopathy of Prematurity trial,5 plus disease became the major driver for treatment. With focus now directed to the timely identification of plus disease, the historical dichotomous categorization of normal versus plus disease was deemed inadequate and led to the introduction of an intermediate level, “preplus disease,” in 20056 and the subsequent critical acknowledgment in the 2021 classification7 that the vascular changes associated with ROP do not progress stepwise but are a continuous spectrum.However, categorizing the plus spectrum is subjective and subject to significant variation.8–10 In the late 1990s, the introduction of digital imaging into the NICU provided the opportunity to objectively measure what had long been seen, and this development spawned several methods of retinal vessel measurement.11,12 Despite revealing great early promise, these methods have yet to be incorporated into routine ROP screening worldwide.AI has recently been used to help interpret the complex images captured to monitor for ROP and other patient-specific risk factors. Coyner et al1 report the use of this technology to explore the spectrum of ROP-related vascular changes, and using deep learning (a type of machine learning that can be used to develop predictive rules), the Oregon team developed i-ROP (an algorithm to detect plus disease),13 from which a vascular severity score (VSS) was derived from zone I (the area closest to the optic disc).13,14 Coyner et al1 have devised a risk model to predict ROP requiring treatment using 2 components: gestational age and VSS. Based on images obtained at 32 to 33 weeks’ postmenstrual age, the model had 100% sensitivity and 80.8% negative predictive value to predict (on average, >1 month in advance of the diagnosis) an eye which would later require treatment. This is an exciting and important development. As a result, instead of a stepwise progression through normal, preplus, and plus disease, VSS permits a sliding scale and detailed analysis of the entire plus spectrum, which hopefully will permit fine tuning of the model in the future.To paraphrase Lord Kelvin (1883), when you can measure what you are speaking about, and express it in numbers, you know something about it; but when you cannot measure it, your knowledge is of a meager and unsatisfactory kind. Herein lies a dilemma: Terms such as AI and deep learning are now in common parlance, so much so that few of us are brave enough to express ignorance. By expressing it (VSS) in numbers, it is not clear that, currently, we fully understand what is being measured or if it has advanced our scientific understanding. Hopefully, this will become apparent in the future. Pragmatically, however, there is no doubt that AI-related techniques, including deep learning–derived VSS, are effective and will likely be important in ROP screening. For this to be incorporated into routine clinical practice, some challenges need to be overcome: a short processing time will be essential preferably within the time span of the NICU ward rounds. In addition, this research is North America based and needs to be validated in other countries in which neonatal care may well be more variable and larger infants develop sight-threatening ROP over a shorter time period. Coyner et al1 have introduced us to a new and exciting era of efficient ROP identification.