Patient-Specific Visual Field Forecasting With Hybrid Deep Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional visual field (VF) testing for glaucoma is tedious, time-consuming, and prone to high test-retest variability, requiring long wait times and multiple tests to forecast future VF, with existing deep learning methods lacking patient-specific analysis and accuracy deteriorating with multiple inputs.
Innovation Solution
A hybrid deep learning framework combining 2-D convolutional neural networks (CNN) with transformers and recurrent neural networks (RNNs) for spatial and temporal modeling, utilizing a few prior VF tests to forecast future VFs with improved accuracy and reduced wait times.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional VF testing is performed repeatedly to detect disease progression, then measurement precision improves, but loss of time increases significantly
Solution Approach 1:
The system performs preliminary VF forecasting using deep learning models to predict future visual field outcomes before actual progression occurs. By analyzing patterns from one or more prior VF tests, the system generates early warnings of impending vision loss, enabling clinicians to take preventive action before significant damage occurs, thereby reducing the need for frequent in-person visits.
Solution Approach 2:
The system creates a digital twin or virtual representation of the patient's visual field progression by training deep learning models on historical VF data. This virtual model can simulate future VF outcomes without requiring actual physical testing, allowing clinicians to monitor progression continuously through computational predictions rather than repeated physical measurements.
2Measurement precision
If multiple prior VF tests are used for conventional regression modeling, then forecasting accuracy improves, but loss of time increases due to extended follow-up periods
Solution Approach 1:
The system transforms the forecasting approach by changing key parameters: instead of requiring multiple VF tests (n≥3) over extended periods (≥1.5 years), the deep learning model achieves comparable or superior accuracy using only one prior VF test. The model learns complex non-linear relationships and patterns from the single test that traditional linear regression cannot capture, fundamentally altering the time-accuracy tradeoff.
Solution Approach 2:
The system replaces the mechanical approach of collecting multiple physical VF measurements over time with an intelligent system that uses deep learning algorithms to extract predictive information from minimal data. The neural network processes VF test data, patient demographics, and disease characteristics to generate forecasts, substituting computational intelligence for brute-force data collection.
3Loss of time
If deep learning methods use only one prior VF test for forecasting, then loss of time decreases, but measurement precision deteriorates compared to conventional methods
Solution Approach 1:
The system adds new dimensions to the forecasting problem by incorporating multiple data types beyond just VF test values, including patient demographics, disease characteristics, and temporal patterns. The deep learning model processes this multi-dimensional information space to compensate for using fewer VF tests, extracting richer predictive signals from each individual test through advanced feature representation and pattern recognition.
4Quantity of substance
If existing deep learning methods take multiple VFs as input, then more data is available for analysis, but forecasting performance deteriorates
Solution Approach 1:
The system applies partial action by intentionally using only one prior VF test as input, recognizing that adding more tests beyond this point provides diminishing or negative returns for existing deep learning architectures. The model is specifically designed to extract maximum predictive information from minimal input, avoiding the performance degradation that occurs when forcing multiple VF inputs into poorly suited network architectures.
Data Source
AI summary
Methods and systems for forecasting a patient's future pointwise visual field (VF) based on one or more visual field tests are disclosed. A hybrid deep learning framework is employed to combine the strengths of recurrent neural networks (RNN), convolutional neural networks (CNN), and transformers. Specific embodiments incorporate self-attention as part of a hybrid CNN and transformer architecture. The disclosed deep learning framework may also be used to generate an estimate of a patient's VF based on 2D or 3D optical coherence tomography (OCT) retinal image data provided as input.