AI in cataract care stays stuck at internal testing, scoping review finds

Artificial intelligence applications now span the cataract care pathway, according to the authors of a new systematic scoping review. Yet according to a systematic scoping review published in BMC Ophthalmology, the vast majority of these tools have never been tested beyond the datasets they were built on, and the field still lacks a clinically validated model that can predict how complex an individual patient’s cataract surgery will be before the operation begins.

The review, led by Ludwig Melino Tjokrovonco and colleagues, set out to map where artificial intelligence applications actually sit along the cataract care pathway and to assess how mature their validation really is. Cataracts remain a leading cause of reversible blindness worldwide, and the authors note that AI applications now span the entire care continuum. What has been unclear until now, they argue, is how those applications are distributed across the pathway and how far each has progressed toward real-world clinical use.

To answer those questions, the team followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews, known as PRISMA-ScR. They searched PubMed, Scopus, the Cochrane Library, WorldCat, and the American Academy of Ophthalmology Meeting Archive, with the final search completed on 20 July 2025. Eligible studies were English-language primary research published from 1 January 2020 through 20 July 2025. Two reviewers independently carried out study selection, data extraction, endpoint adjudication, and validation-maturity assessment, and performance results were summarized descriptively rather than pooled statistically.

A distinctive feature of the review is a new evidence-mapping framework called CAICC, short for the cataract artificial intelligence clinical continuum, which the authors developed during data charting and synthesis. The framework organizes AI research into six stages of care: detection, characterization, personalized surgical planning, intelligent surgery, recovery, and patient engagement. This structure allowed the reviewers to place each of the 185 included studies at a specific point along the patient journey rather than treating all cataract AI research as a single undifferentiated field.

The resulting map shows a striking imbalance. Detection accounted for 52 studies and characterization for 50, meaning that more than half of the published work concentrates on identifying and describing cataract rather than on supporting treatment decisions. Personalized surgical planning drew 27 studies and intelligent surgery 28, while recovery attracted 24 and patient engagement just 4.

The validation picture is even more sobering. Among the 181 included studies for which validation assessment was applicable, internal testing was the most frequent highest-maturity category, accounting for 123 studies, or 68.0 percent. External testing reached 33 studies, or 18.2 percent. Prospective single-center evaluation appeared in only 6 studies, or 3.3 percent, and prospective multicenter evaluation in just 2 studies, or 1.1 percent. A further 17 studies, or 9.4 percent, had no independent testing or an unclear testing status.

The review also identified a specific and consequential gap. Among the retrieved and included studies, the authors found no comprehensive clinically validated multimodal preoperative cataract surgical-complexity model. The review’s title names this as the surgical complexity prediction gap, and the authors call it out as a priority for future research alongside independent external testing and prospective multicenter evaluation of locked models.

The reviewers are careful about what their validation-maturity categories mean. They characterize evaluation independence and clinical stage, not risk of bias or overall methodological quality. In other words, a study classified at a higher maturity level has simply been evaluated on more independent data or in a more realistic clinical setting; the framework does not certify that the model is safe, accurate, or free from methodological flaws. This distinction matters for readers interpreting the map, because a tool that reached prospective testing is not automatically better built, only better tested.

Another important limitation concerns retrieval. Of 361 reports sought for retrieval, 122, or 33.8 percent, were not retrieved, and 113 of those 122, or 92.6 percent, were considered plausibly eligible from available metadata. Because so many potentially relevant reports could not be obtained and most of those seemed likely to qualify, the authors state that retrieval bias may affect the comprehensiveness of the evidence map.

The review was registered on the Open Science Framework, and the article was published on 27 October 2026 as an open-access systematic review in BMC Ophthalmology. Its central message is measured but pointed: most included studies remained at internal validation stages, prospective multicenter evaluation was rare, and the field’s next steps should prioritize multimodal preoperative surgical-complexity prediction, independent external testing, and prospective multicenter evaluation of locked models.

Subject of Research: A scoping review mapping artificial intelligence applications and validation maturity across the cataract care continuum

Article Title: Artificial intelligence across the cataract care continuum: a systematic scoping review introducing the CAICC framework and the surgical complexity prediction gap

Article References: Tjokrovonco, L. M., Kartasasmita, A. S., Kartiwa, R. A., Virgana, R., Prabaswara, C., & Azuraa, A. N. (2026). Artificial intelligence across the cataract care continuum: a systematic scoping review introducing the CAICC framework and the surgical complexity prediction gap. BMC Ophthalmology. https://doi.org/10.1186/s12886-026-05483-4

Image Credits: AI Generated

DOI: 10.1186/s12886-026-05483-4

Keywords: artificial intelligence, cataract, scoping review, CAICC framework, surgical complexity prediction, validation maturity, BMC Ophthalmology, ophthalmology, prospective multicenter evaluation, PRISMA-ScR, Artificial, intelligence