Tabular-To-Image Transformation for Transfer Learning on Heterogeneous Health Data

Tabular health records are central to health informatics, yet variations across patients, institutions, and measurement protocols lead to misaligned feature spaces that hinder generalizable deep learning and transfer learning. To address this challenge, we propose Tabular-to-Image Transfer Learning (T2I-TL), a framework that converts heterogeneous tabular records into semantically structured 2D images by grouping related features into spatial regions. This representation enables pretrained convolutional neural networks (CNNs) to serve as transferable feature extractors without requiring explicit feature alignment. The extracted representations are then used with conventional classifiers for individualized risk prediction. We evaluate T2I-TL on two asthma patient cohorts with partially overlapping feature schemas. T2I-TL consistently outperforms traditional classifiers and TL baseline models, achieving up to 30.5% and 13.75% improvement in sensitivity over a conventional classifier and a CNN-based TL baseline, respectively. These results demonstrate its effectiveness for personalized prediction under structural heterogeneity and limited data.