Reducing variability of visual field sensitivities in glaucoma through spatial filtering

Abstract

 

Purpose

 

To determine if spatial filtering that uses functional and structural test data reduces variability while assessing glaucomatous visual field progression.

 

Methods

 

Humphrey visual field (VF) data (SITA Std, 24-2) and confocal scanning laser ophthalmoscopy (HRT) data from 112 eyes of 62 participants with high-risk ocular hypertension or mild to moderate glaucoma were included. All eyes had a minimum of 5 visits with mean follow up 5.3 years (range: 1.9-11.5 years). For each of the 52 non-blind spot VF locations, the filtered value (Ã_j ) at location j was Ã_j =0.5(A_j +Â_j ), where A_j and Â_j are the observed and predicted sensitivities. The predicted value was Â_j = ∑_i=1³⁷ č_ix_i , where predictors x_i included sensitivities at the other 25 locations in the same hemi-field and the logarithm of rim area at twelve 30° optic disc sectors. An independent dataset containing VF and HRT data from 1057 eyes of 637 clinic patients was used to estimate the č_is. Least Absolute Shrinkage And Selection Operator was used to estimate č_is, limited to 8 non-zero positive č_is to simplify the filter and avoid over-fitting. At each location per eye, the trend over time was modeled by a linear model (LM), and a non-linear model (NLM) of the form A=a+ce^t , using filtered or unfiltered data from visits 1 to (n-1). The standard deviations (SD) of residuals from the trends, and prediction errors (PE; deviation at visit n between predicted and observed sensitivities) were compared between  and A. The analyses were repeated after truncating VF data so that thresholds <15dB were set to equal 15dB to reduce noise.

 

Results

 

SD of the residuals about the trend were reduced by filtering at all 52 VF locations (p<0.001) for all analyses. PE² were reduced by filtering at 44 and 48 VF locations (p<0.05) for LM analyses on observed and truncated data respectively, and at all 52 VF locations (p<0.05) for both NLM analyses. Truncating data before filtering reduced the variability (p<0.01) at 38 and 40 VF locations for LM and NLM analyses.

 

Conclusions

 

Filtering which uses functional and structural test data can reduce variability in predicting pointwise sensitivities in longitudinal sequences of VF data, and improve the accuracy of predictions. Truncating data before filtering can further reduce variability.

 

 

Histogram and density plot of SD of residuals from the linear model using unfiltered and filtered data.

 

Keywords: 758 visual fields • 627 optic disc • 642 perimetry