Machine Learning-Based Prediction of Long-Term Intraocular Pressure Fluctuations in Open-Angle Glaucoma.

OBJECTIVE: To investigate the predictability of long-term intraocular pressure (IOP) fluctuations in open-angle glaucoma eyes implanted with a telemetric IOP sensor. DESIGN: A prospective, open-label, single-arm, multicenter study. SUBJECTS: Twenty-four patients were enrolled, including 20 with primary open-angle glaucoma, 2 with pseudoexfoliative, 1 with pigmentary, and 1 with uveitic glaucoma. Mean age was 65.2 ± 10.2 years. METHODS: Telemetric IOP measurements were aggregated into nyctohemeral means. The first 90 postoperative days were excluded. A rolling reference framework was applied, in which temporally paired observations were generated by comparing each eligible day with a future time point at fixed prediction horizons. The relationship between short-term (7, 14, or 28 days) and long-term (273 or 364 days) fluctuations was assessed using Pearson correlation. Multivariate regression was applied to predict long-term fluctuations based on short-term data. In addition, supervised machine learning with a Random Forest Classifier was used to predict long-term fluctuations from clinical, demographic, and IOP-derived features. For each horizon (273 or 364 days) and threshold (+2.0, +3.0, and +4.0 mmHg), changes in mean nyctohemeral IOP were calculated. Outcomes were labeled as "1" if the increase met or exceeded the threshold and "0" otherwise. MAIN OUTCOME MEASURES: Predictability of long-term IOP fluctuations at 273 and 364 days. RESULTS: Short-term fluctuations correlated only weakly with long-term variability (Pearson r ≤ 0.33) and explained at most 15.2% in regression analysis. Across 1224 Random Forest Classifier models, 47 met inclusion criteria of area under the receiver operating characteristic curve (AUROC) >0.8 and sensitivity and specificity >0.7 (27 for 364 days, 20 for 273 days). On average, 5563 ± 116 valid pairings from 9.2 ± 0.8 patients were used per configuration. Five final configurations were selected for each threshold-horizon combination based on the highest F1 values. Performance metrics included AUROC 0.81 to 0.86, cross-validated AUROC 0.78 to 0.83, accuracy 0.72 to 0.81, sensitivity 0.72 to 0.78, specificity 0.70 to 0.82, precision 0.32 to 0.44, and F1 value 0.44 to 0.56. All models included 7, 14, and 28-day fluctuations, mean nyctohemeral IOP, ocular pulse amplitude, age, body mass index, and central corneal thickness as predictors, with mean nyctohemeral IOP contributing most (38%-55%). CONCLUSIONS: Long-term IOP fluctuations can be predicted from baseline clinical and demographic data combined with IOP-related features. Telemetric devices and remote IOP monitoring, combined with predictive modeling, could reduce the burden of time-intensive procedures and health care costs while supporting individualized care in the face of rising demand. FINANCIAL DISCLOSURES: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.