Development of a Neural Network Model for Predicting Keratoconus in Leber Congenital Amaurosis and AIPL1 Mutations

Theme: Cornea, external disease & refractive surgery
What: Cornea, external disease & refractive surgery
Part of: Cornea III: Refractive surgery and keratoconus / Cornée III: Refractive surgery and keratoconus
When: 5/31/2024, 04:15 PM - 05:45 PM
Where: Room | Salle 713 AB

Abstract

Purpose Leber Congenital Amaurosis (LCA) is a rare genetic retinal disorder resulting in severe visual impairment from birth. AIPL1 gene mutations are among the causes of LCA and exhibit a highly variable clinical phenotype, often accompanied by symptoms like nystagmus, photophobia, and keratoconus. The exact cause of keratoconus in LCA remains uncertain but is suspected to be related to the oculo-digital phenomenon. Timely diagnosis of keratoconus is crucial for optimizing treatment outcomes. To address this knowledge gap, we have developed a neural network to predict the presence of keratoconus in LCA patients with AIPL1 mutations, utilizing clinical features and allele mutations as input data.

Study Design Machine learning applications of prospectively collected data

Methods Data collected globally in 2004 included nineteen patients diagnosed with Leber Congenital Amaurosis (LCA) possessing AIPL1 mutations and undergoing ocular examinations to assess keratoconus. A neural network model was developed using MATLAB, utilizing various clinical factors such as age, geographic origin, the presence of oculo-digital phenomenon, night blindness, photoaversion, photoattraction, pigmentary retinopathy, optic nerve pallor, maculopathy, and the presence or absence of the stop mutation W278Stop in one or both AIPL1 alleles as input parameters. The model's output data represented the presence or absence of keratoconus. The dataset was randomly divided into training, validation, and testing sets in a 60%, 20%, and 20% ratio, respectively. The neural network model's performance was assessed using accuracy, sensitivity, and specificity measures.

Results The average accuracy across all neural networks was 92.7±20.0% in the training set, 92.5±12.1% in the validation set and 82.5±26.5% in the test set. The average total sensitivity across all 10 neural networks was 88.3±11.3%, the average total specificity was 91.5±24.2% and the average total accuracy across all neural networks was 90.5±17.5%, Furthermore, 4 out of the 10 neural networks made no errors across the training, validation, or test sets, indicating a high level of consistency in their predictions. Additionally, 8 of the 10 neural networks had a 100% specificity in the test sets.

Conclusion We developed a neural network model specifically for keratoconus in blindness due to retinal degeneration, which shows promising results in predicting the presence of keratoconus in LCA patients with AIPL1 mutations. We were able to predict based on the clinical features and the AIPL1 mutations as input data. The trained networks consistently demonstrated high specificity. The results demonstrate proof of concept for the use of neural network models to predict the presence of keratoconus in LCA patients with AIPL1 mutations.

Presenter(s)

Presenting Author: Raheem Remtulla

Additional Author(s):

Robert Koenekoop, McGill Universtiy

---