ESCRS - FPT01.04 - Predictive Factors Of Cataract Surgery Success Using Machine Learning Methods: Results Of The Femcat Study (Impact Médico-Economique De La Chirurgie De La Cataracte Au Laser Femtoseconde)

Predictive Factors Of Cataract Surgery Success Using Machine Learning Methods: Results Of The Femcat Study (Impact Médico-Economique De La Chirurgie De La Cataracte Au Laser Femtoseconde)

Published 2022 - 40th Congress of the ESCRS

Reference: FPT01.04 | Type: Free paper | DOI: 10.82333/r8s7-7709

Authors: Cedric - Schweitzer* 1 , Beatrice Cochener 2 , Antoine Brezin 3 , Dominique Monnet 3 , Nathalie Hayes 1 , Christine Germain 4 , Philippe Denis 5 , Pierre-Jean Pisella 6 , Nejib ZemZemi 7

1Ophthalmology,Bordeaux University Hospital,Bordeaux,France, 2Brest University Hospital,Brest,France, 3Ophthalmology,AP-HP,Paris,France, 4Biostatistics,ISPED- Bordeaux University,Bordeaux,France, 5Ophthalmology,Lyon University Hospital,Lyon,France, 6Ophthalmology,Tours University Hospital,Tours,France, 7National Institute for Research in Digital Science and Technology,IHU LYRIC- Bordeaux University,Bordeaux,France

Purpose

To analyse predictive factors associated with cataract surgery success using machine learning methods

Setting

University hospital of Bordeaux (France), and Lyon, Tours, Paris, Brest University Hospitals (France)

Methods

The FEMCAT study is a prospective multicentre randomised clinical trial comparing femtosecond laser and phacoemulsification cataract surgery in two parallel patient group. Consecutive patients eligible for uni- or bilateral cataract surgery were included. Success was defined at 3 months as a combination of 4 outcome measures: absence of severe peri- or postoperative complications within the 3 months period, best corrected visual acuity of 0.0 LogMAR (VA), absolute refractive error ≤0.75 Dioptres (RE) and no change in postoperative corneal astigmatism (CA). Predictive factors of success were performed on complete data using a random forest classification model. Training and validation sets were composed of 80% and 20% of the eyes.

Results

1497 eyes of 909 patients were included and a total of 29 parameters were used in the model. The best diagnostic performance of the training set was observed for the absence of severe complications outcome (Area under the receiver operating curve characteristics (AUROC): 1.0, sensitivity: 100%), followed by RE (AUROC: 0.983, sensitivity: 100%), VA (AUROC: 0.939, sensitivity: 99.4%) and CA outcomes (AUROC: 0.801, sensitivity: 99.0%).The best diagnostic performance of the validation set was observed for CA outcome (AUROC: 0.736, sensitivity: 76.9%), followed by RE (AUROC: 0.651, sensitivity: 89.6%), absence of severe complications (AUROC: 0.609, sensitivity: 94.2%) and VA outcomes (AUROC: 0.515, sensitivity: 83.7%).

Conclusions

Machine learning methods showed good diagnostic performance to predict cataract surgery success in our study. Such an algorithm could help determine patient characteristics for postoperative anatomical, visual and refractive success to optimise the indications of premium intraocular lens for correction of astigmatism and presbyopia.