Researchers in South Korea say that physicians may be able to better predict a patient’s risk of a cardiovascular event by using a deep-learning program to evaluate the patient’s fundus.

In the retrospective cohort study, the surgeons trained a deep-learning model using 15,408 images from the Health Promotion Center of the Seoul National University Hospital. They trained the program to predict carotid artery atherosclerosis from the images, and called the result the deep-learning fundoscopic atherosclerosis score (DL-FAS). They then constructed a retrospective cohort of patients between 30 and 80 years of age who had completed elective health exams at the hospital.

For predicting carotid artery atherosclerosis among subjects, the model achieved an area under receiver operating curve of 0.713, and area under the precision-recall curve of 0.569. The accuracy was 0.583, the sensitivity was 0.891 and the specificity was 0.404. The positive and negative predictive values were 0.465 and 0.865, respectively. 

In the cohort, which consisted of 32,227 participants, there were 78 cardiovascular disease (CVD) deaths, and follow-up visits every 7.6-years (median). Those with DL-FAS greater than 0.66 had an increased risk of CVD deaths compared to those with DL-FAS <0.33 (hazard ratio: 8.33; 95% confidence interval [CI], 3.16-24.7). Risk association was significant among intermediate and high Framingham risk score subgroups. The researchers found that the DL-FAS improved the concordance by 0.0266 (95% CI, 0.0043-0.0489) over the FRS-only model. 

Ultimately, the physicians say that the deep-learning program’s DL-FAS score was an independent predictor of CVD deaths when adjusted for FRS and even had an added predictive value beyond the FRS.

Amer J Ophthalmol 2020;217:121-130. 
Chang J, Ko A, Park SM, et al. 


Accuracy of Next-gen IOL Formulas in Vitrectomized Eyes
Physicians from Guangdong, China, analyzed the latest intraocular lens formulas’ performance in eyes that had undergone vitrectomies. 

The study was a retrospective, consecutive-case-series review of 111 eyes of 111 patients that underwent uneventful phacoemulsification and IOL implantation after vitrectomy. The surgeons divided the patients into four groups according to whether the posterior chamber was filled with silicone oil. They then evaluated the performance of several IOL formulas, with and without lens-constant optimization.

The researchers say that, before lens-constant optimization, the mean prediction errors (MEs) of all formulas were statistically different from zero (0.14 to 0.46 D) in vitrectomized eyes, except for the Kane formula. The Barrett Universal II, Emmetropia Verifying Optical, Kane and Haigis formulas all had relatively lower mean absolute error (MAE) and median absolute error (MedAE) with optimized constants. The doctors found no significant systemic bias in the new formulas for vitrectomized eyes with axial lengths greater than 26 mm (p>0.05). 

The Hoffer Q and Holladay 1 displayed significant hyperopic shift (0.39 and 0.51 D) for long eyes, which was corrected by the Wang-Koch axial length adjustment. The researchers report that there were no significant differences in the prediction accuracy of all formulas among the four subgroups (p>0.05).

The physicians say that the BUII, EVO, Kane and Haigis displayed comparable performance in vitrectomized eyes with optimized constants. In vitrectomized highly myopic eyes, the new formulas and traditional formulas with WK adjustment exhibited “satisfactory prediction accuracy.” Silicone oil tamponade didn’t affect the prediction accuracy of formulas when using the IOLMaster 700.  

Amer J Ophthalmol 2020;217:81-90.
Tan X, Zhang J, Zhu Y, et al.