Vascular physiology-informed machine learning to identify similar subgroups of glaucoma patients across studies: Indianapolis Glaucoma Progression Study, Thessaloniki Eye Study, and Singapore Epidemiology of Eye Disease Study

Abstract

Purpose : Intraocular pressure (IOP) and blood pressure (BP) are two risk factors in the development and progression of open angle glaucoma (OAG). However, various population-based studies differ for the role that the IOP-BP association has in OAG. To address this, we present a novel approach based on a physiology-informed mathematical model and machine learning (ML) applied to 3 studies conducted in different continents.

Methods : We considered: the Indianapolis Glaucoma Progression Study (IGPS) of 115 OAG eyes; the Singapore Epidemiology of Eye Disease Study (SEED) of 19,625 eyes (283 OAG); and the Thessaloniki Eye Study (TES) of 3,136 eyes (140 OAG). For each study, a validated mathematical model (Guidoboni et al 2014, IOVS) was used to generate 8 hemodynamic variables based on measurements of 4 patient-specific inputs (systolic and diastolic BP, heart rate, IOP). These combined variables form a 12-dimensional physiology-enhanced dataset. An unsupervised ML clustering algorithm, Fuzzy C-Means (FCM), was applied to the 3 enhanced datasets, partitioning eyes into groups with similar hemodynamics.

Results : The FCM algorithm produced 3 clusters on the 12-D data. Fig.1 shows the clustering result from the IGPS study (only OAG eyes) projected onto the IOP-MAP plane (MAP=mean arterial pressure=(2/3)DBP+(1/3)SBP). A wedge-like shape with a tilted boundary between cluster 1 (green) and cluster 3 (blue) is observed, indicating a non-trivial interplay between IOP and MAP. Clustering only the OAG eyes in SEED yielded similar wedge-shaped clusters (Fig. 2a), but with all eyes considered (healthy+OAG) the resulting clusters were vertically stacked (Fig. 2b), indicating they differ only by MAP. A similar phenomenon occurred in TES (Figs. 2c and 2d).

Conclusions : The physiology-informed ML approach revealed a wedge-shaped pattern among OAG eyes that is consistent across studies and continents. When OAG eyes are pooled together with healthy eyes in population-based studies, such as TES and SEED, this pattern is masked by the overwhelming presence of healthy eyes. The pattern identifies OAG eyes with similar disease biomarkers and may enable individualized OAG management and treatment across populations.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Clusters from the IGPS study

 

Clusters of all eyes (b,d) vs. OAG eyes (a,c)