Abstract
Purpose :
Intraocular lens (IOL) power calculations for cataract surgery remain an imperfect science with known shortcomings in eyes with extreme biometric measurements. The purpose of this study was to investigate whether an augmented intelligence (AI)-based clinical decision aid integrated into the electronic medical record (EMR) improved prediction accuracy of the SRK/T formula in eyes undergoing cataract surgery.
Methods :
This was a retrospective consecutive case series compared SRK/T and AI-modified SRK/T predictions on eyes in which monofocal IOLs were implanted in adult subjects. The testing dataset included 116 eyes of 79 patents, implanted with an MX60E IOL. Exclusion criteria included corneal or intraocular opacities of any kind other than cataract, history of other intraocular surgery, post-operative refraction recorded outside study period (21-90 days following surgery), or post-operative best-corrected vision worse than 20/30. A tool was built into the EMR that analyzes biometric data and makes recommendations from the peer-reviewed literature on how to adjust refractive targets for eyes with short or long axial lengths (AI-AL), flat or steep keratometries (AL-K), or shallow or deep anterior chamber depths (AI-ACD). Primary endpoints included standard deviation (SD) of the prediction error (PE) for both SRK/T and AI-modified SRK/T predictions. Secondary endpoints include mean absolute prediction error (MAE), and proportion of eyes with post-operative spherical equivalent (SE) within 0.25 and 0.50 diopters (D) of predicted.
Results :
SRK/T + AI-K improved upon the original SRK/T formula based upon several metrics. The SDs of the signed PEs were 0.657 and 0.623 D for the SRK/T and SRK/T + AI-K, respectively. MAEs for the same were 0.455 and 0.420 D, respectively (p < 0.0003). The proportion of eyes with post-operative SE within 0.25 and 0.50 D of predicted were 39.7% and 69.8% vs. 44.0% and 72.4%, respectively (p < 0.0002 for both). Other adjustments of the SRK/T formula based on AL and ACD did not result in meaningful improvements.
Conclusions :
An AI-based clinical decision aid integrated into an EMR may help to refine prediction accuracy of SRK/T. Further validation is needed in more eyes and consideration of multiple biometry measurements simultaneously.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.