The neural network for measuring IOP by Maklakov method: comparison between neuronal net and experts

Purpose. To evaluate the possibility of clinical application of the AITonometry neural network in determining IOP by Maklakov method and identifying patients with increased intraocular pressure. Material and methods. A prospective study, 697 impressions of IOP measurement by Maklakov were selected. Each impression was evaluated using the neural network (group I), and by 3 experts with using the measuring ruler of Prof. B.L. Polyak (groups II1 , II2 , II3 ), the experts' data were averaged in order to create a «comparison standard» (group IIM ) for analysis. Results. The data obtained are presented in the form of M±σ(Me [Q25%; Q75%]), where M is the mean value, ±σ is the standard deviation, Me is the median ([Q25%; Q75%]) are quartiles: group I – 22,32±4,18 (22 [20; 24]); II1 group – 19.95±3.58 (19 [18; 21]); II2 group – 20,35±3,65 (20 [18; 22]); II3 group – 20,41±3,58 (20 [18; 22]); IIM – 20,23±3,53 (19,66 [18,33; 21,66]). A statistically significant difference was found between all data groups. The accuracy of AI-Tonometry: the mean absolute error (MAE) is 2.5 mmHg, the mean squared error (MSE) is 8.76, the root mean squared error (RMSE) is 2.96. Performance: Sensitivity – 91.94%; Specificity – 93.7%; Accuracy – 58.76%; Accuracy – 93.54%; F1 Score – 0.7170, AUC-ROC – 0.987. Conclusion. The AI-Tonometry neural network has high characteristics of accuracy and performance in combination with the convenience of using the service, the speed of interpretation of impressions made by Maklakov method, a small discrepancy in the results of determining the presence or absence of increased IOP. This service can be considered as an alternative to the usual approach with using ruler of Prof. B.L. Polyak and can be used in clinical practice. Keywords: neural network, intraocular pressure, artificial intelligence, tonometry, Maklakov applanation tonometry