Oftalʹmologiâ (Oct 2022)

Machine Learning Methods in the Comparative Evaluation of Various Approaches to the Surgical Treatment of Primary Angle Closure

  • N. I. Kurysheva,
  • A. L. Pomerantsev,
  • O. Ye. Rodionova,
  • G. A. Sharova

DOI
https://doi.org/10.18008/1816-5095-2022-3-549-556
Journal volume & issue
Vol. 19, no. 3
pp. 549 – 556

Abstract

Read online

Purpose. To evaluate the application of the principal component analysis (PCA) and DD-SIMCA in a comparative analysis of the surgical treatment of primary angle closure.Material and methods. The prospective study included 90 patients. Group 1 — 30 patients with primary angle closure (PAC) with planned laser peripheral iridotomy (LPI). Group 2 — 30 patients with PAC, with planned phacoemulsification with intraocular lens implantation (PE+IOL). Group 3 — 30 eyes without ophthalmic pathology. All subjects underwent SS-OCT. Thirty-seven parameters were analyzed, including intraocular pressure, choroidal thickness in the macula, anterior chamber depth, lens vault, iris curvature and thickness, angle opening distance, and iridotrabecular space at 500 µm and 750 µm from the scleral spur. Since all these parameters correlate with each other, machine learning methods were used: PCA and the DD-SIMCA one-class classification method. For this purpose graphs of scores and loads in the PCA model for groups 1 and 2 were plotted. In the score plot, patients with PAC with average and extreme eye parameters were identified, and in the loading plot, relationships between the parameters of patients with PM were used to analyze correlations in the future. In the DD-SIMCA method, group 1 is taken as representatives of the target class.Results. A classification model based on 2 principal components with a given type I error α = 0.01 demonstrated a sensitivity of 100 % for patients in its own group and a sensitivity of 93 % for patients in group 2. These results confirm similarity of group 1 and group 2. The specificity for the control group was 100 %, and this group located far from the target group.Conclusion. Machine learning methods make it possible to compare groups with multivariate and correlated parameters. PCA allows the identification of patients with extreme parameters and the evaluation of correlations between multiple parameters. DDSIMCA confirms the validity of comparing the results of treatment with LPI and FE + IOL.

Keywords