Clarifying the roles of high and low blood pressure in glaucoma via physiology-informed machine learning

Abstract

Purpose : Both high and low blood pressure (BP), together with intraocular pressure (IOP), are risk factors for glaucoma. The underlying physiological mechanisms for both high and low BP as a risk factor remain elusive. Here a novel physiology-informed machine learning (ML) approach combines physiology-based mathematical modeling and unsupervised clustering to discern the impact that abnormal BP may have on glaucoma.

Methods : 115 open angle glaucoma (OAG) patients were assessed for: IOP, systolic and diastolic BP (SBP, DBP), heart rate (HR), and color Doppler imaging (CDI) of peak-systolic and end-diastolic velocity (PSV, EDV) of the ophthalmic artery (OA) and central retinal artery (CRA). For each patient, IOP, SBP, DBP, and HR were fed to a validated mathematical model (Guidoboni et al, IOVS, 2014) which estimates hemodynamic variables unavailable from instrumentation. The enhanced dataset and the model-estimated variables are fed to a Fuzzy c-means (FCM) clustering algorithm to identify patient clusters with similar ocular hemodynamics.

Results : The FCM clustering algorithm revealed 3 clusters plotted in the IOP-MAP plane in Fig.1 (left) (MAP=mean arterial pressure = 2/3DBP+1/3SBP). Fig.1 (right) shows that ocular perfusion pressure (OPP = 2/3MAP-IOP) is high in cluster 2 and low in cluster 3. The PSV medians in the CRA are similar for all clusters (<10%), while the PSV median in the OA of cluster 2 was ~22% higher than in clusters 1 and 3 (see Fig.2 (top)). On the other hand, cluster 3 exhibits higher vascular resistance in the venules and central retinal vein (Fig.2 (bottom)).

Conclusions : High and low BP may contribute to glaucoma in different ways, depending also on the IOP level. The bigger discrepancy in PSV medians in OA with respect to CRA may indicate that patients in cluster 2 need a stronger autoregulation engagement to maintain homeostasis, thereby limiting their capacity to accommodate physiological BP fluctuations. High venous resistance in cluster 3 might render those vessels susceptible to venous collapse. The technology for measuring venous resistances is currently unavailable and our results indicate that its development might be an important goal for the study of glaucoma.

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

 

 

Velocities (as measured by CDI) and model-estimated resistances. We report the median and (25,75)-percentile.