A fine-grained transformer combined with multimodal data for predicting hospital length of stay in acute coronary syndrome

The prediction of hospital length of stay (LOS) is of great significance for hospitals to rationally allocate medical resources and provide timely treatment for patients, especially for acute coronary syndromes (ACS), which requires prompt medical intervention. To this end, we propose a fine-grained Transformer with morphological enhancement for predicting hospital LOS in ACS. Considering the morphological features of blood vessels in computed tomography (CT) images, we design photometric and geometric transformations and combined self-supervised learning to extract enhanced morphological features. To fuse the CT image features with the physiological features contained in the electronic medical record, a multiscale attention is designed and capture the intra-modal and inter-modal fine-grained features with sparse strategy. We compare 16 state-of-the-art models, including classic machine learning models commonly used in ACS, the most advanced visual models, multimodal Transformers, and the latest LOS prediction models. The results show that our model achieved the best prediction results (mean absolute error is 1.33, Pearson correlation coefficient is 0.96). Further, we conduct the interpretability analysis, our model can well perceive the lesion regions, and the salient features have significant correlation with LOS (p = 0.0005). The ablation experiments also verified the effectiveness of each module. This work is expected to provide an effective tool for LOS prediction, thereby enabling the rational allocation of medical resources and offering timely intervention to patients in hospital.