Purpose
The purpose was to perform comparative evaluation of the several machine learning algorism to predict quality of visual life (QoVL) in glaucomatous patients. Furthermore, we aimed to identify important visual field (VF) test points for QoVL considering the inter-relationship among visual acuity (VA) and VF test points.
Methods
QoVL score was surveyed in 164 glaucoma patients using the ‘Sumi questionnaire’ (Ophthalmology 2003;110:332-339). The relationship between VAs of better/worse eye, total deviation (TD) values of all of the test points on integrated binocular visual field (IVF), and the general and each QoVL score (letters/sentences, walking, going out, dining) were investigated using four machine learning algorism: the Random Forest (RF), Gradient Boosting machine (Boost), Support vector machine (SVM) and Feed forward neural network (NNET).For comparison, multiple regression (MR) method was also investigated. The cross validation was carried out using the one leave out method and the prediction errors of those four methods were compared. Then the contributing IVF test points for the general and each QoVL scores were determined using the algorism with tightest prediction error.
Results
The prediction errors from RF, Boost, NNET and SVM were 1.97, 1.96, 2.59 and 2.23, respectively, which were significantly smaller compared to that from MR (3.38). Thus, we adopted RF for the following analyses. The important VF test points for general QoVL score existed widely around the horizontal line and some points existed at peripheral area (Figure 1). Specific test points were chosen for each QoVL task (Figure 2); VF test points along the horizontal line were chosen for the letters/sentences, and walking. In addition, peripheral points in the left hemi-field were chosen for letters/sentences, and peripheral inferior hemi-field was chosen for the walking. For going out, test points just beneath the horizontal line in the left hemifield and also those in the peripheral superior hemi-field was chosen. For dining, the selected test points were scattered widely in peripheral areas.
Conclusions
Accurate prediction of QoVL was obtained by analyzing TDs on IVF and VA simultaneously using the RF method and also important VF test points were identified using the prediction system. It would be beneficial to pay attention to these VF locations for clinicians and patients of glaucoma.
Keywords: 669 quality of life •
758 visual fields