Preventive screening of open‐angle glaucoma: an innovative machine learning risk assessment tool based on health insurance claims data

Abstract

AbstractPurposeDespite progress in management of open‐angle glaucoma, early diagnosis remains difficult due to delayed ophthalmologist consultation. Early glaucoma treatment avoids vision loss, thus developing automated detection methods is of importance for public health. We used supervised machine learning models on medical insurance claims data to identify at‐risk patient profiles that would benefit from undergoing glaucoma diagnosis, leading to early detection.MethodsUsing medical claims records from a Swiss health insurance company, we identified a “glaucoma group” composed by 2700 patients (18–65 yo) with at least 3 years of historical data in the database and at least one year of medications for lowering intraocular pressure. To complete the training dataset, we randomly selected a “non‐glaucoma control group” (N = 2700) with same age range and medical criteria. Random Forest models were then trained using patient medical consumption data (e.g. number of drugs, number of consultations, ocular tests) in order to estimate individual’s probability of having glaucoma based on medical claims records. Population was split into training set (70%) and test set (30%) to estimate the model prediction performance.ResultsThe mean age of population with glaucoma was 55.3 (SD: 11.3), and had an average of 1.96 (SD: 2.8) eye visits within the 3 years period prior to their first glaucoma prescription. t‐test showed significant association between consumption of eye products (e.g. artificial tears) during the two years prior to glaucoma medication. The Random Forest classifier accuracy computed on the test set (N = 1600) showed a positive predictive value of 78.5%, a negative predictive value of 79.2%. Sensitivity and specificity were of 79.4% and 77.2% respectively.ConclusionsTo our knowledge, this study is the first to attempt estimating glaucoma probabilities from medical consumption invoice data. Our preliminary results show that using supervised machine learning algorithms could identify high‐risk patient profiles of glaucoma with an accuracy of almost 80%. These patients may benefit from an early ocular examination. Automated screening based on medical claims data could thus allow improving open‐angle glaucoma‐related health outcomes, minimizing serious vision loss, and reducing health costs.