Introduction
Electronic Health Records (EHR) provide a rich source of data for predictive healthcare, yet they are often siloed into structured visit sequences (medical codes) and unstructured clinical notes (free-text observations). Synthesizing these heterogeneous data types is challenging due to "heterogeneous semantic biases"—where discrete codes and descriptive text do not naturally align in a shared vector space. Furthermore, clinical notes are often noisy, and existing models frequently overlook the informative value of dissimilar patient trajectories.
Methodology
The VecoCare framework addresses these challenges through a three-stage joint learning approach designed to align structured and unstructured medical data.
1. GWD-based Contrastive Learning
To bridge the semantic gap, VecoCare employs Gromov-Wasserstein Distance (GWD). Unlike standard contrastive learning that forces different data types into identical distributions, GWD compares the relational distances within the structured code space against those within the text space. This allows the model to align the two modalities by preserving their internal relational structures.
2. Adaptive Masked Language Model (AMLM)
The core of the model is a "Global-Local Fusing Encoder" that integrates Transformers for long-range dependencies and CNNs for local textual patterns. The AMLM task involves predicting masked codes and words using an adaptive weighting strategy, preventing the model from becoming over-reliant on a single data modality.
3. Dual-channel Retrieval Module
During prediction, VecoCare utilizes a "Patient Memory Bank" with a unique dual-channel mechanism. While the positive channel identifies similar patients, the negative channel identifies dissimilar patients. This allows the model to learn from what a patient is not, providing a more comprehensive representation for final diagnosis.
Experimental Results
VecoCare was evaluated on the public MIMIC-III dataset and a private Real-World Hospital (RWH) dataset. It consistently outperformed state-of-the-art baselines, including CGL, Chet, and MetaCare++, across all metrics. On MIMIC-III, VecoCare achieved a ma-AUPRC of 0.2646, significantly higher than Chet (0.2328). On the RWH dataset, it reached 0.3107 compared to Chet's 0.2803.
A key strength of VecoCare is its robustness to data scarcity. As shown in the performance analysis, even when the training set was reduced to 10%, VecoCare maintained a higher ma-AUPRC than competing models trained on significantly more data. This suggests that the pre-training tasks effectively capture fundamental medical "common sense."
Conclusion
VecoCare demonstrates that joint learning from visit sequences and clinical notes, combined with a dual-channel retrieval mechanism, significantly improves diagnosis prediction. While current constraints limit note length to 512 words, future work aims to incorporate additional modalities such as medical imaging and continuous waveforms to create an even more comprehensive EHR representation.
Summary updated