Approach to glaucoma diagnosis and prediction based on multiparameter neural network

Area of Science:

• Ophthalmology
• Medical Imaging
• Data Science

Background:

• Glaucoma diagnosis relies on various clinical indicators.
• Integrating advanced parameters may enhance diagnostic accuracy.
• Novel analytical methods are needed for complex glaucoma assessment.

Purpose of the Study:

• To evaluate the impact of comprehensive factor analysis on glaucoma assessment.
• To investigate the combined effect of trans-laminar cribrosa pressure difference (TLCPD) and fractional pressure reserve (FPR) in glaucoma prediction.
• To compare the efficacy of different parameters in neural network-based glaucoma classification.

Main Methods:

• Utilized clinical data from 1029 patients (Beijing Tongren Hospital) and 600 cases (Beijing Eye Research).
• Employed doc2vec for data vectorization and multivariate imputation by chained equations (MICE) for data interpolation.
• Trained neural networks (VGG16 and autoencoder) using original, TLCPD-combined, FPR-combined, and non-combined parameters for glaucoma prediction.

Main Results:

• Achieved accuracy rates exceeding 0.90 and precision rates of 0.70-0.80 for glaucoma classification using original data.
• Autoencoder models incorporating TLCPD and FPR demonstrated accuracy rates near 1.0 and precision rates of 0.90 and 0.70.
• Combined parameters significantly improved classification performance compared to individual parameters.

Conclusions:

• The integration of FPR and TLCPD parameters substantially enhances glaucoma diagnosis and prediction.
• Fractional pressure reserve (FPR) shows greater suitability than TLCPD for improving neural network performance in glaucoma classification.
• Comprehensive factor analysis with advanced parameters offers a promising approach for more accurate glaucoma detection.