Abstract
Purpose :
This study aims to identify a subtype of patients with ocular hypertension that benefited the most from topical medication using unsupervised and supervised machine learning models.
Methods :
We applied unsupervised machine learning such as density-based clustering and t-distributed stochastic neighbor embedding (t-SNE) to identify different groups of patients that participated the ocular hypertension treatment study (OHTS) with similar patterns of visual field (VF) loss at the baseline. A total of 1636 patients (817 in the active and 819 in the observation group) at the baseline of OHTS study were included in the analysis. The rate of mean deviation (MD) progression based on seven years of follow up was used as the outcome. To determine which subgroup of patients responded to the medications the most, MD worsening was compared between the active and observation groups within each cluster of patients separately.
Results :
We identified 13 clusters with similar patterns of VF loss at baseline. Eyes in one cluster responded to ocular hypertensive medication better than others. More specifically, in this cluster, MD progression rate of the active group was significantly lower than the observation group (MD rate of +0.04 dB/year versus -0.09 dB/year; P=0.03). While the patients in this subtype were stable, other patients taking medicine had a significantly worse progression rate (-0.07 dB/year, P≤0.001). Compared to other patients taking medicine, the patients in this subtype were older (64.3 ±9.5 versus 54.7 ±8.94 years, P≤0.001) and had a worse baseline MD (-2.2±0.1 dB versus +0.47±1.2 dB, P≤0.001). Additionally, the patients in this subgroup had a significantly lower family history of glaucoma (P≤0.001), higher history of heart disease (P≤0.001), and experienced high blood pressure (P≤0.001).
Conclusions :
We identified a subtype of patients with ocular hypertension that better responded to topical medications. This may lead to the personalized glaucoma care which assign patients with shared clinical characteristics to different treatment plans that benefits the patients the most.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.