Physiology-informed machine learning to enable precision medical approaches of intraocular pressure and blood pressure management in glaucoma

Abstract

Purpose : Vision loss in many open-angle glaucoma (OAG) patients continues despite successful treatment lowering intraocular pressure (IOP). The reasons remain elusive. Here, we show that mathematical modeling and machine learning (ML) can help identify subgroups of OAG patients whose disease process entails different relative contributions of IOP and blood pressure (BP).

Methods : 115 OAG patients were assessed every 6 months over a 7-year period for IOP, systolic and diastolic blood pressures (SBP, DBP), heart rate (HR), structural and hemodynamic evaluations via ocular coherence tomography (OCT), Heidelberg Retinal Tomography (HRT), Heidelberg Retinal Flowmetry (HRF), and Color Doppler Imaging (CDI). Fuzzy c-means (FCM) clustering was applied to the dataset comprising: (i) IOP, SBP, DBP, HR measured at the first visit for each patient in the IGPS study; and (ii) individualized estimates of vascular pressures and resistances obtained via a validated mathematical model (Guidoboni et al, IOVS, 2014). FCM is part of ML and the mathematical model is based on physiology, leading to physiology-informed ML. Follow-up visits and data from OCT, HRT, HRF and CDI were not used for clustering.

Results : While the data for IOP and mean arterial pressure (MAP) for the first visit of each IGPS patient do not exhibit any particular pattern (Fig.1a), the physiology-informed ML method revealed 3 distinct clusters. The slanted lines separating the clusters result from the nonlinear interplay that IOP and BP have on ocular hemodynamics captured by the mathematical model. Fig. 2 shows that the cluster membership based on the first visit is associated with different clinical outcomes after 4 years (p values obtained with the 2-sample paired Wilcoxon signed rank test for medians; p <= 0.05 in bold). Cluster 1 shows minimal progression, whereas Cluster 2 shows marked structural progression accompanied by significant hemodynamic changes. Cluster 3 exhibits significant changes only in HRT and HRF markers, but not in OCT and CDI.

Conclusions : This study suggests that the proposed physiology-informed ML approach can identify and quantify the relative contributions of IOP and BP on the OAG risk for patient subgroups. Thus, this approach may enable precision medical approaches of IOP and BP management in OAG.

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

 

MAP-IOP plots before (a) and after (b) clustering

 

Clinical markers at Years 1 and 4