Machine Learning-Based Prediction of Need for Disease-Modifying Antirheumatic Drugs in the Treatment of Autoimmune Uveitis

Theme: Uveitis
What: Uveitis
Part of: Uveitis II: Advancements in Uveitis Diagnosis and Management / Uvéite II: Avancées dans le diagnostic et la prise en charge de l’uvéite
When: 5/31/2024, 02:00 PM - 03:30 PM
Where: Room | Salle 714 B

Abstract

Purpose: Autoimmune uveitis is a prevalent yet growing ocular health condition. Its treatment often involves challenging immunosuppression regimens, necessitating extensive investigations and coordination with allied services. However, determining which patients will require such care can be a complex task, typically established later in the course of the disease. Fortunately, advances in machine learning have enabled enhanced predictive capabilities through computational modeling. In this study, we aim to develop a neural network capable of predicting the need for immunosuppression in the management of autoimmune uveitis.

Study Design: Retrospective analysis of prospectively collected data.

Methods: We utilized a dataset of 54 patients with autoimmune uveitis, sourced from The National Eye Institute and The University Medical Center Utrecht, which includes patients with anterior uveitis, intermediate uveitis, and Behcet's uveitis. Patient-specific information, including age, biological sex, diagnosis, and the use of disease-modifying antirheumatic drugs (DMARD), was made available. A neural network with a single layer and 15 nodes was constructed using Matlab. The network was trained with the Scaled Conjugate Gradient algorithm for classification tasks. Data splitting was performed with a 75/10/15 split for training, validation, and test sets, respectively. Input factors included diagnosis, age, and biological sex, with DMARD use as the output. Sensitivity, specificity, total accuracy, and area under the curve (AUC) were calculated for each dataset.

Results: Using the described methodology, the neural network completed training in 47 epochs with a gradient of descent of 0.063. The performance graph displayed favorable cross-entropy across training epochs. The network exhibited an accuracy of 82.9% in the training set, 100% in the validation set, and 87.5% in the test set, resulting in a total accuracy of 85.2%. Importantly, the network maintained strong specificity, with a specificity of 83.3% in the training set and 100% in both the validation and test sets. The total network specificity was 88.9%. The AUC in the training set was 0.898, in the validation set was 1.00, in the test set was 0.805, and the overall AUC was 0.905.

Conclusion: Despite a relatively small sample size, our neural network demonstrated high specificity in predicting DMARD use in patients with autoimmune uveitis. This predictive model employed a surprisingly small number of input factors while still achieving a respectable AUC. This study offers a proof of concept, indicating that machine learning applications have the potential to guide treatment decisions and improve the management of autoimmune uveitis in this complex patient population.

Presenter(s)

Presenting Author: Raheem Remtulla

Additional Author(s):

Karen Oliver, McGill University

---