A Joint Multitask Learning Model for Cross-sectional and Longitudinal Predictions of Visual Field Using OCT

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Accurate prediction of visual field (VF) progression is crucial for managing glaucoma.
• Current methods for predicting VF loss may not fully leverage multimodal data like optical coherence tomography (OCT).
• Multitask learning offers a promising approach to integrate information from different but related tasks.

Purpose of the Study:

• To evaluate the prediction accuracy of a novel multitask learning model, latent space linear regression and deep learning (LSLR-DL).
• To assess LSLR-DL's performance in both cross-sectional VF prediction using OCT and longitudinal VF progression prediction.
• To validate the model's efficacy on an independent dataset.

Main Methods:

• A cohort study was conducted using data from healthy individuals and open-angle glaucoma (OAG) patients.
• The LSLR-DL model was developed to jointly perform cross-sectional (10° VF) and longitudinal (30° VF) predictions by sharing a deep learning component.
• Root mean square error (RMSE) was used to quantify prediction accuracy, comparing LSLR-DL against its individual components and ordinary linear regression.

Main Results:

• LSLR-DL achieved a mean RMSE of 6.4 dB for cross-sectional VF prediction.
• For longitudinal prediction, LSLR-DL yielded RMSE values between 4.4 dB and 3.7 dB, depending on the number of prior VF tests used.
• The LSLR-DL model demonstrated significantly superior performance compared to traditional methods.

Conclusions:

• The LSLR-DL model is effective for both cross-sectional and longitudinal VF prediction in glaucoma.
• This multitask learning approach enhances prediction accuracy by leveraging auxiliary information from related tasks.
• LSLR-DL shows potential as a valuable tool for monitoring glaucoma progression.