A Novel Interpretable Transfer Learning Framework for Analyzing High-Dimensional Longitudinal Ophthalmic Data

Abstract

Purpose : Various transfer learning (TL) approaches have been developed in deep learning model schemes to address data inequality and discrepancies between datasets and domains. While TL has recently been introduced to enhance health equity in healthcare, particularly ophthalmology and vision health, persistent challenges include adequately modeling complex dependency structures such as longitudinal and multi-modal data, along with ensuring the interpretability. This study aims to develop a novel TL machine learning framework for analyzing high-dimensional longitudinal ophthalmic data, prioritizing model interpretability and accounting for temporal and inter-eye correlations.

Methods : The TL framework adopts a generalized linear mixed-effects regression model to investigate the association of longitudinally measured covariates, efficiently handling complex data dependencies such as repeated measurements. A two-step transfer algorithm is implemented to leverage information from source domain for enhancing the prediction task in the target domain. L₁-regularization techniques are used to address high dimensionality. We conducted a simulation study with a target sample size of 500, and source sample size ranging from 200 to 1200 with 5 temporal repeated measurements for each sample. We compared the proposed TL method with Naïve lasso regression model fitted using only the target domain, evaluating model prediction accuracy and parameter estimation through mean square errors (MSE) and L₂-estimators of estimated coefficients, respectively.

Results : The simulation study demonstrates that the TL method outperforms the naïve lasso method, evidenced by lower MSE and L₂-estimators. Furthermore, L₂-estimators of the proposed TL method exhibit a clear trend of decreasing as the sample size of the source domain increases, indicating the benefit of borrowing information from the source domain (Figure 1).

Conclusions : The proposed TL framework enhances model prediction accuracy compared with the naïve method. The selected features from the TL framework, along with corresponding estimated coefficients, offer a comprehensive interpretation of each feature’s contribution to the outcome of interest. This TL framework provides an efficient and flexible solution to address sample size constraints in the target domain, mitigate data imbalance issues, and handle complex data structures in clinical research.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.