![SonomaEye-full-color-vector[2].png](https://static.wixstatic.com/media/bed00b_da20b9cc704044b0b048ac8ef9190470~mv2_d_3024_2195_s_2.png/v1/fill/w_177,h_123,al_c,q_85,usm_0.66_1.00_0.01,enc_avif,quality_auto/SonomaEye-full-color-vector%5B2%5D.png)
Thank you for your submission!
Prediction of Treatment Response Using Artificial Intelligence in Birdshot Chorioretinopathy
Lingling Huang, MD, PhD
Presenter:
Lingling Huang1, Cheolhong An2, Tate Valerio3, Yasmin Massoudi3, Dirk-Uwe Bartsch2, Peter Y. Chang3, C. Stephen Foster3 Stephen D. Anesi 3,
Authors:
Affiliation:
1. The Permanente Medical Group, Sacramento, Northern California
2. University of California San Diego
3. Massachusetts Eye Research and Surgery Institution
Purpose: Birdshot chorioretinopathy (BSCR) is a bilateral posterior uveitis that can cause blindness. Treatment often involves a stepped approach, but predicting which patients need aggressive therapies (eg. intravenous infusion or intravitreal corticosteroid implants) versus milder options (eg. oral antimetabolites or self-injectable biologics) is difficult. This study aims to use artificial intelligence to analyze ocular images and predict the need for aggressive therapies to prevent treatment delays and vision loss.
Methods: Using deep convolutional neural network approach to classify BSCR patients as treatment responsive vs resistant via fluorescence angiographic (FA) images and via optical coherence tomography (OCT) images.
Results: Collaborating with Massachusetts Eye Research and Surgery Institution, we retrospectively screened 429 BSCR patients, and 141 patients met inclusion criteria. They were classified as Responsive (72 patients) or Resistant (69 patients) based on clinical course. We collected images from “Before” period when disease was active, and images from “After” period when disease was inactive. Each subject has 20-40 FA images and 122 OCT scans. The dataset was divided for Training (70%), Validating (10%) and Testing (20%). AI models were trained with different combinations of Before/After images to compare the performance.
In Resistant group, there are 29 patients with Before FA and 69 with After FA. In Responsive group, there are 21 patients with Before FA and 63 with After FA. Training with Before FA alone, AI model achieved 91.6% accuracy with 83.3% sensitivity and 100% specificity. However, training with After FA alone or mixed with Before FA, AI model achieved variable 54-66% accuracy only.
We are currently working on OCT-based AI model, however, the performance is significantly limited by small sample size and high data dimension.
Conclusion: Deep learning AI model can predict Resistant vs Responsive for BSCR patients using FA. Before images achieve excellent performance. However, our current data is limited by small sample size. We are expanding our network and inviting other institutions to joint this project. National or international colloboration will greatly enhance the ability of AI development in the field of rare diseases, such as uveitis.