Recognizing patterns of visual field loss using unsupervised machine learning

Using ChatGPT-4 in visual field test assessment

CLINICAL RELEVANCE

Visual field testing is essential in the diagnosis and management of various ophthalmic diseases, particularly glaucoma. Integrating ChatGPT-4 into the interpretation of these tests has the potential to aid clinical decision making and improve efficiency and accessibility in clinical practice.

BACKGROUND

This study aims to evaluate the capability of ChatGPT-4 in interpreting visual field tests.

METHOD

A total of 30 patient visual field printouts, either with or without defects, were included in this study. The performance of ChatGPT-4 in identifying test name, pattern, reliability indices, total deviation map, pattern deviation map and greyscale map was evaluated and compared with that of 2 experienced glaucoma consultants. The study also focused on the ability of ChatGPT to categorise tests as 'normal' or suggest diagnosis by interpreting tests accurately.

RESULTS

The results showed that ChatGPT-4 was highly accurate in identifying test names (100%), patterns (90%) and global visual field indices (96.7%). It also accurately classified tests as reliable or unreliable (93.3%).The model provided 66.7% and 30% accurate and adequate answers in interpreting deviation and greyscale maps, respectively. In addition, in 33.3% of tests, it was able to accurately interpret the visual field test and classify it as 'normal' or suggest a diagnosis.

CONCLUSION

The study highlights the potential of large language models like ChatGPT-4 in assessing visual field tests. ChatGPT-4 could interpret numeric data on tests accurately. However, it was inadequate in interpreting deviation and greyscale maps and suggesting a diagnosis according to the defects.

A Consensus Statement on the Terminology for Automated Visual Field Abnormalities

BACKGROUND

A multitude of terms have been used to describe automated visual field abnormalities. To date, there is no universally accepted system of definitions or guidelines. Variability among clinicians creates the risk of miscommunication and the compromise of patient care. The purposes of this study were to 1) assess the degree of consistency among a group of neuro-ophthalmologists in the description of visual field abnormalities and 2) to create a consensus statement with standardized terminology and definitions.

METHODS

In phase one of the study, all neuro-ophthalmologists in Israel were asked to complete a survey in which they described the abnormalities in 10 selected automated visual field tests. In phase 2 of the study, the authors created a national consensus statement on the terminology and definitions for visual field abnormalities using a modified Delphi method. In phase 3, the neuro-ophthalmologists were asked to repeat the initial survey of the 10 visual fields using the consensus statement to formulate their answers.

RESULTS

Twenty-six neuro-ophthalmologists participated in the initial survey. On average, there were 7.5 unique descriptions for each of the visual fields (SD 3.17), a description of only the location in 24.6% (SD 0.19), and an undecided response in 6.15% (SD 4.13). Twenty-two neuro-ophthalmologists participated in the creation of a consensus statement which included 24 types of abnormalities with specific definitions. Twenty-three neuro-ophthalmologists repeated the survey using the consensus statement. On average, in the repeated survey, there were 5.9 unique descriptions for each of the visual fields (SD 1.79), a description of only the location in 0.004% (SD 0.01), and an undecided response in 3.07% (SD 2.11%). Relative to the first survey, there was a significant improvement in the use of specific and decisive terminology.

CONCLUSIONS

The study confirmed a great degree of variability in the use of terminology to describe automated visual field abnormalities. The creation of a consensus statement was associated with improved use of specific terminology. Future efforts may be warranted to further standardize terminology and definitions.

How Do Patients Feel About Visual Field Testing? Analysis of Subjective Perception of Standard Automated Perimetry

PURPOSE

A high rate of unreliability is an issue in visual field (VF) testing, especially in elderly patients, and warrants patient education. We assessed whether subjective perception of the visual field test (VFT) is a good predictor of its reliability in different age groups and examined age differences in patients' awareness of VF damage.

METHODS

This cross-sectional study investigated 107 VFT results of 54 patients with glaucoma or ocular hypertension. Subjective perceptions were compared to reliability indices for cooperation analysis and to mean deviation results for VF damage analysis, and an age-segregated sub-analysis was performed.

RESULTS

Kappa coefficients showed poor agreement between subjective and objective parameters. Nevertheless, there were age differences. Younger patients had a higher positive predictive value and sensitivity in cooperation analysis and a higher negative predictive value in VF damage analysis.

CONCLUSIONS

Patients' perception of cooperation in VFT is a poor predictor of its reliability. Although young cooperative patients may be aware of their good cooperation, even the youngest are unaware of their poor performance. This emphasizes the importance of giving proper directions to all patients during VFT to obtain better reliability indices. Younger, healthy patients are more aware of their health status, than those with a damaged VF, regardless of age. Therefore, illness education is crucial in all glaucoma patients.

Unsupervised learning for large-scale corneal topography clustering

Machine learning algorithms have recently shown their precision and potential in many different use cases and fields of medicine. Most of the algorithms used are supervised and need a large quantity of labeled data to achieve high accuracy. Also, most applications of machine learning in medicine are attempts to mimic or exceed human diagnostic capabilities but little work has been done to show the power of these algorithms to help collect and pre-process a large amount of data. In this study we show how unsupervised learning can extract and sort usable data from large unlabeled datasets with minimal human intervention. Our digital examination tools used in clinical practice store such databases and are largely under-exploited. We applied unsupervised algorithms to corneal topography examinations which remains the gold standard test for diagnosis and follow-up of many corneal diseases and refractive surgery screening. We could extract 7019 usable examinations which were automatically sorted in 3 common diagnoses (Normal, Keratoconus and History of Refractive Surgery) from an unlabeled database with an overall accuracy of 96.5%. Similar methods could be used on any form of digital examination database and greatly speed up the data collection process and yield to the elaboration of stronger supervised models.

Convex Representations Using Deep Archetypal Analysis for Predicting Glaucoma

Goal: The purpose of this study was to identify clinically relevant patterns of glaucomatous vision loss through convex representation to predict glaucoma several years prior to disease onset. Methods: We developed a deep archetypal analysis to identify patterns of glaucomatous vision loss, and then projected visual fields over the identified patterns. Projections provided a representation that was more accurate in detecting glaucomatous vision loss, thus, more appropriate for recognizing preclinical signs of glaucoma prior to disease development. To overcome the class imbalance in prediction, we implemented a class-balanced bagging with neural networks. Results: Using original visual field as features of the class-balanced bagging classification provided an area under the receiver-operating characteristic curve (AUC) of 0.55 for predicting glaucoma approximately four years prior to disease development. Using convex representation of the visual fields as input features provided an AUC of 0.61 while using deep convex representation as input features improved the AUC to 0.71. Relevance vector machine (RVM) achieved an AUC of 0.64. Conclusion: Deep archetypal analysis representation of visual functional features with balanced bagging classification could serve as an automated tool for predicting glaucoma. Significance: Glaucoma is the second leading cause of worldwide blindness. Most people with glaucoma have no early symptoms or pain, delaying diagnosis in many patients until they reach late irreversible vision loss stages. In fact, about 50% of people with glaucoma are unaware they have the disease. Deep archetypal analysis models may impact clinical practice in effectively identifying at-risk glaucoma patients well prior to disease development.

Monitoring Glaucomatous Functional Loss Using an Artificial Intelligence-Enabled Dashboard

PURPOSE

To develop an artificial intelligence (AI) dashboard for monitoring glaucomatous functional loss.

DESIGN

Retrospective, cross-sectional, longitudinal cohort study.

PARTICIPANTS

Of 31 591 visual fields (VFs) on 8077 subjects, 13 231 VFs from the most recent visit of each patient were included to develop the AI dashboard. Longitudinal VFs from 287 eyes with glaucoma were used to validate the models.

METHOD

We entered VF data from the most recent visit of glaucomatous and nonglaucomatous patients into a "pipeline" that included principal component analysis (PCA), manifold learning, and unsupervised clustering to identify eyes with similar global, hemifield, and local patterns of VF loss. We visualized the results on a map, which we refer to as an "AI-enabled glaucoma dashboard." We used density-based clustering and the VF decomposition method called "archetypal analysis" to annotate the dashboard. Finally, we used 2 separate benchmark datasets-one representing "likely nonprogression" and the other representing "likely progression"-to validate the dashboard and assess its ability to portray functional change over time in glaucoma.

MAIN OUTCOME MEASURES

The severity and extent of functional loss and characteristic patterns of VF loss in patients with glaucoma.

RESULTS

After building the dashboard, we identified 32 nonoverlapping clusters. Each cluster on the dashboard corresponded to a particular global functional severity, an extent of VF loss into different hemifields, and characteristic local patterns of VF loss. By using 2 independent benchmark datasets and a definition of stability as trajectories not passing through over 2 clusters in a left or downward direction, the specificity for detecting "likely nonprogression" was 94% and the sensitivity for detecting "likely progression" was 77%.

CONCLUSIONS

The AI-enabled glaucoma dashboard, developed using a large VF dataset containing a broad spectrum of visual deficit types, has the potential to provide clinicians with a user-friendly tool for determination of the severity of glaucomatous vision deficit, the spatial extent of the damage, and a means for monitoring the disease progression.

Retinal Nerve Fiber Layer Features Identified by Unsupervised Machine Learning on Optical Coherence Tomography Scans Predict Glaucoma Progression

Purpose

To apply computational techniques to wide-angle swept-source optical coherence tomography (SS-OCT) images to identify novel, glaucoma-related structural features and improve detection of glaucoma and prediction of future glaucomatous progression.

Methods

Wide-angle SS-OCT, OCT circumpapillary retinal nerve fiber layer (cpRNFL) circle scans spectral-domain (SD)-OCT, standard automated perimetry (SAP), and frequency doubling technology (FDT) visual field tests were completed every 3 months for 2 years from a cohort of 28 healthy participants (56 eyes) and 93 glaucoma participants (179 eyes). RNFL thickness maps were extracted from segmented SS-OCT images and an unsupervised machine learning approach based on principal component analysis (PCA) was used to identify novel structural features. Area under the receiver operating characteristic curve (AUC) was used to assess diagnostic accuracy of RNFL PCA for detecting glaucoma and progression compared to SAP, FDT, and cpRNFL measures.

Results

The RNFL PCA features were significantly associated with mean deviation (MD) in both SAP (R² = 0.49, P < 0.0001) and FDT visual field testing (R² = 0.48, P < 0.0001), and with mean circumpapillary RNFL thickness (cpRNFLt) from SD-OCT (R² = 0.58, P < 0.0001). The identified features outperformed each of these measures in detecting glaucoma with an AUC of 0.95 for RNFL PCA compared to an 0.90 for mean cpRNFLt (P = 0.09), 0.86 for SAP MD (P = 0.034), and 0.83 for FDT MD (P = 0.021). Accuracy in predicting progression was also significantly higher for RNFL PCA compared to SAP MD, FDT MD, and mean cpRNFLt (P = 0.046, P = 0.007, and P = 0.044, respectively).

Conclusions

A computational approach can identify structural features that improve glaucoma detection and progression prediction.

Development of a new algorithm based on FDT Matrix perimetry and SD-OCT to improve early glaucoma detection in primary care

Purpose

The purpose of this study was to develop an objective algorithm to discriminate the earliest stages of glaucoma using frequency doubling technology (FDT) Matrix perimetry and spectral domain-optical coherence tomography (OCT) technology to improve primary care detection.

Materials and methods

Three hundred six eyes (mean age 58.67±15.12) from 161 patients were included and classified in the following three groups: 101 nonglaucoma (GI-NG), 100 glaucoma suspect (GII-SG), and 105 open-angle glaucoma (GIII-OAG). All participants underwent a visual field exploration using the Humphrey Matrix visual field instrument and retinal nerve fiber layer evaluation using the Topcon 3D OCT-2000. Pattern deviation plot was divided into 19 areas and five aggrupation or quadrants and ranked with a value between 0 and 4 according to its likelihood of normality, and differences among three groups were analyzed. Principal component analysis (PCA) was also used to extract the most notable features of FDT and OCT, and a logistic regression analysis was applied to obtain the classification rules.

Results

Only area numbers 7 and 12 and the central zone of FDT Matrix showed statistical differences (P<0.05) between GI-NG and GII-SG. The classification rules were estimated by the four PCA obtained from FDT Matrix and 3D OCT-2000 in a separate and combined use. Area under the receiver operating characteristic curve was 78.88% with FDT-PCA, 82.09% with OCT-PCA, and 94.27% with combined use of FDT and OCT-PCA to discriminate GI-NG and GII-SG.

Conclusion

The predictive rules based on FDT-PCA or OCT-PCA provide a high sensitivity and specificity to detect the earliest stages of glaucoma and even better in combined use. These predictive rules may help the future development of software for FDT Matrix perimetry and 3D OCT-2000, which will greatly improve their diagnostic ability, making them useful in daily practice in a primary care setting.

Unsupervised Gaussian Mixture-Model With Expectation Maximization for Detecting Glaucomatous Progression in Standard Automated Perimetry Visual Fields

Purpose

To validate Gaussian mixture-model with expectation maximization (GEM) and variational Bayesian independent component analysis mixture-models (VIM) for detecting glaucomatous progression along visual field (VF) defect patterns (GEM–progression of patterns (POP) and VIM-POP). To compare GEM-POP and VIM-POP with other methods.

Methods

GEM and VIM models separated cross-sectional abnormal VFs from 859 eyes and normal VFs from 1117 eyes into abnormal and normal clusters. Clusters were decomposed into independent axes. The confidence limit (CL) of stability was established for each axis with a set of 84 stable eyes. Sensitivity for detecting progression was assessed in a sample of 83 eyes with known progressive glaucomatous optic neuropathy (PGON). Eyes were classified as progressed if any defect pattern progressed beyond the CL of stability. Performance of GEM-POP and VIM-POP was compared to point-wise linear regression (PLR), permutation analysis of PLR (PoPLR), and linear regression (LR) of mean deviation (MD), and visual field index (VFI).

Results

Sensitivity and specificity for detecting glaucomatous VFs were 89.9% and 93.8%, respectively, for GEM and 93.0% and 97.0%, respectively, for VIM. Receiver operating characteristic (ROC) curve areas for classifying progressed eyes were 0.82 for VIM-POP, 0.86 for GEM-POP, 0.81 for PoPLR, 0.69 for LR of MD, and 0.76 for LR of VFI.

Conclusions

GEM-POP was significantly more sensitive to PGON than PoPLR and linear regression of MD and VFI in our sample, while providing localized progression information.

Translational Relevance

Detection of glaucomatous progression can be improved by assessing longitudinal changes in localized patterns of glaucomatous defect identified by unsupervised machine learning.

Detecting glaucomatous change in visual fields: Analysis with an optimization framework

Detecting glaucomatous progression is an important aspect of glaucoma management. The assessment of longitudinal series of visual fields, measured using Standard Automated Perimetry (SAP), is considered the reference standard for this effort. We seek efficient techniques for determining progression from longitudinal visual fields by formulating the problem as an optimization framework, learned from a population of glaucoma data. The longitudinal data from each patient's eye were used in a convex optimization framework to find a vector that is representative of the progression direction of the sample population, as a whole. Post-hoc analysis of longitudinal visual fields across the derived vector led to optimal progression (change) detection. The proposed method was compared to recently described progression detection methods and to linear regression of instrument-defined global indices, and showed slightly higher sensitivities at the highest specificities than other methods (a clinically desirable result). The proposed approach is simpler, faster, and more efficient for detecting glaucomatous changes, compared to our previously proposed machine learning-based methods, although it provides somewhat less information. This approach has potential application in glaucoma clinics for patient monitoring and in research centers for classification of study participants.