Artificial intelligence applications facilitate decision-making on cataract surgery for high myopic patients

Kaimeng Su¹Wenwen He¹Haifeng Jiang²Keke Zhang¹Jiao Qi¹Jiaqi Meng¹Yu Du¹Kaiwen Cheng¹Xiaoxin Hu¹Dongling Guo¹Haike Guo^3*Yong Wang^2*Yi Lu^1*Xiangjia Zhu^1*

• ¹Eye and Ent Hospital, Fudan University, Shanghai, China
• ²Wuhan Aier Eye Hospital, Wuhan, Hebei Province, China
• ³Shanghai Heping Eye Hospital, Shanghai, China

Background: Surgical decision-making for highly myopic cataracts demands high expertise from doctors. We therefore aimed to develop a preliminary AI model for surgical decision-making in highly myopic cataracts based on previous deep-learning models. Materials and Methods: We first established a highly myopic cataract decision-making AI model by integrating cataract grading and postoperative visual acuity prediction models of highly myopic eyes we developed before with the surgical decision logic. The outcomes of the surgical decision-making were classified into four categories: surgery not advised, cataract surgery recommended, retinal surgery recommended, and combined cataract-retinal surgery recommended. The gold standard for the surgical decision is defined as the decision made by 2 professional ophthalmologists together (X.Z. and Y.W.). If the decision-makings on the highly myopic cataract surgery are not fully consistent, a final judgment will be made by the third expert (Y.L.). Subsequently, we evaluated AI model's surgical decision-making accuracy with gold standard and doctors in different levels in both internal (107 highly myopic eyes from Eye & ENT Hospital, Fudan University) and external (55 highly myopic eyes from Wuhan Aier Eye Hospital) test datasets. Results: In the internal and external datasets, according to the Lens Opacities Classification System (LOCS) III international standards of cataract grading, 99.07% and 87.27% of automatic nuclear grading, and 88.79% and 61.82% of automatic cortical grading, respectively had an absolute prediction error ≤1.0 with gold standard. Then, the mean postoperative visual acuity prediction error was 0.1560 and 0.3057 logMAR in the internal and external datasets, respectively. Finally, the consistency of AI model's and the gold standard's surgical decisions on highly myopic cataract patients in the internal and external dataset was 96.26% and 81.82% respectively. The AI demonstrated substantial agreement with the gold standard (Kappa value = 0.811, 0.556 in the internal, external dataset respectively). Conclusion: The AI decision-making model for highly myopic cataracts based on two deep learning models demonstrated good performance, which is helpful for doctors in complex highly myopic cataract surgical decision-making.