Using a Machine Learning Technique Called Kalman Filtering to Forecast Conversion from Ocular Hypertension to Primary Open Angle Glaucoma

Abstract

Purpose : To use a machine learning technique called Kalman Filtering (KF) to create an algorithm that forecasts which patients with ocular hypertension (OHTN) will progress to open-angle glaucoma (OAG). The model dynamically updates its forecast each time new measurements are obtained.

Methods : A KF model (KF-OHTN) was parameterized and trained using 2806 eyes of 1407 patients with OHTN from the OHTS trial and compared to several other forecasting algorithms including a previously created KF model based on patients with OAG (KF-HTG) from AGIS and CIGTS, a Personalized Mean (PM) model which forecasts each patient’s future readings as the mean of their past readings, and 2 linear regression models. Each model forecasted MD, PSD, and IOP at 12 and 60 months into the future for all patients with OHTN in OHTS. Model performance was measured by assessing forecasting error, i.e., the difference between the model’s forecast and the actual values obtained in OHTS. We determined the proportion of MD forecasting errors within clinically relevant thresholds of the actual values (0.5, 1.0, 2.5 dB) and the root mean squared error (RMSE) for MD, PSD, and IOP forecasts at 12 and 60 months in the future. Lower RMSE values indicate lower magnitudes of forecasting error.

Results : 2537 (90.4%) eyes and 2124 (75.5%) eyes had sufficient measurements to forecast 12 and 60 months into the future, respectively. Forecasting 12 months into the future, KF-OHTN and the PM model exhibited similar distributions of MD forecasting errors with 42.0-42.1% of eyes within 0.5 dB, 70.5-70.7% within 1.0 dB, and 95.7-96.1% within 2.5 dB of the actual value from the OHTS trial. When forecasting MD 60 months into the future, 32.8% of eyes were within 0.5 dB, 61.0% of eyes were within 1.0 dB, and 93.2% of the eyes were within 2.5 dB of the actual value using the KF-OHTN model. For the subset of patients with OHTN who progressed to OAG, the 2 KF models had the lowest RMSE for MD while KF-OHTN had the lowest RMSE for PSD and IOP (Tables 1 and 2).

Conclusions : The KF algorithm we developed was able to accurately forecast MD, PSD, and IOP for the majority of patients with OHTN up to 5 years into the future. Among the subset of patients with OHTN who progressed to OAG, our KF models outperformed existing forecasting models.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Distribution of MD Forecasting Errors

 

Magnitude of Forecasting Errors