Multi-modal fusion model for Time-Varying medical Data: Addressing Long-Term dependencies and memory challenges in sequence fusion.

BACKGROUND

Multi-modal time-varying data continuously generated during a patient's hospitalization reflects the patient's disease progression. Certain patient conditions may be associated with long-term states, which is a weakness of current medical multi-modal time-varying data fusion models. Daily ward round notes, as time-series long texts, are often neglected by models.

OBJECTIVE

This study aims to develop an effective medical multi-modal time-varying data fusion model capable of extracting features from long sequences and long texts while capturing long-term dependencies.

METHODS

We proposed a model called medical multi-modal fusion for long-term dependencies (MMF-LD) that fuses time-varying and time-invariant, tabular, and textual data. A progressive multi-modal fusion (PMF) strategy was introduced to address information loss in multi-modal time series fusion, particularly for long time-varying texts. With the integration of the attention mechanism, the long short-term storage memory (LSTsM) gained enhanced capacity to extract long-term dependencies. In conjunction with the temporal convolutional network (TCN), it extracted long-term features from time-varying sequences without neglecting the local contextual information of the time series. Model performance was evaluated on acute myocardial infarction (AMI) and stroke datasets for in-hospital mortality risk prediction and long length-of-stay prediction. area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), and F1 score were used as evaluation metrics for model performance.

RESULTS

The MMF-LD model demonstrated superior performance compared to other multi-modal time-varying data fusion models in model comparison experiments (AUROC: 0.947 and 0.918 in the AMI dataset, and 0.965 and 0.868 in the stroke dataset; AUPRC: 0.410 and 0.675, and 0.467 and 0.533; F1 score: 0.658 and 0.513, and 0.684 and 0.401). Ablation experiments confirmed that the proposed PMF strategy, LSTsM, and TCN modules all contributed to performance improvements as intended.

CONCLUSIONS

The proposed medical multi-modal time-varying data fusion architecture addresses the challenge of forgetting time-varying long textual information in time series fusion. It exhibits stable performance across multiple datasets and tasks. It exhibits strength in capturing long-term dependencies and shows stable performance across multiple datasets and tasks.