C é dric Schweitzer, MD, reported at the annual meeting of the French Association for implant and refractive surgery (Safir) that machine learning algorithms show promise in predicting the success of cataract surgery and may be helpful in identifying patients at risk of obtaining unsatisfactory results.
"The algorithm has shown good performance in determining the overall success of cataract surgery, but it still needs external validation and further improvement of some parameters before it can be used clinically," he said.
Dr. Schweitzer explained that the algorithm was developed in cooperation with the National Institute of digital science and Technology (INRIA) of the University of Bordeaux, using data from the femcat study, a two-year French multicenter trial comparing femtosecond cataract surgery with standard phacoemulsification.
"We want to see if we can use artificial intelligence to predict the success of cataract surgery based on the data in femcat. Cataract surgery has a high level of anatomy and visual reproducibility. In recent years, driven by the needs of patients and the development of high-quality IOL, refractive cataract surgery has shown a clear trend. It is obviously necessary to identify those patients who can expect the full benefits of modern cataract surgery," Dr. Schweitzer said.
The initial femcat study defined success as a combination of four factors: no preoperative or postoperative complications three months after surgery, best corrected visual acuity (BCVA) of 0.0 LogMAR at three months, refractive error of less than 0.75 d at three months, and corneal astigmatism of less than 0.5 d at three months.
The algorithm developed by INRIA uses random forest AI, which is an integrated learning method for classification, regression and other tasks, and can construct a large number of decision trees during training.
"We used 80% of the relevant femcat data as the training data set and 20% for the actual test," Dr. Schweitzer said.
The algorithm gives preoperative and intraoperative data extracted from 1497 eyes of 909 patients. A total of 29 parameters were selected, including demographic, biometric, anatomical, visual and refractive preoperative data and intraoperative total data.
From the results, the parameters that had the greatest impact on the overall success rate included age, cataract grade, IOL power, intraoperative complications, surgeon factors, and corneal measurements.
"The algorithm performs well in determining the overall success of cataract surgery, but this may be improved in the future. The sensitivity of individual criteria is in the range of 92.5% to 100%, which is very good, but for the overall criteria, it is about 60.3%. We plan to improve the algorithm by optimizing the selected parameters and externally validating it, as well as expanding the database to form a more heterogeneous population," Dr. Schweitzer said.
He concluded that when properly tested and validated, the algorithm can be used to help identify patients who may not be able to achieve the expected target results - especially those with advanced toric or multifocal lenses, where expectations are often higher.