Clinical time-series data are inherently complex, characterized by temporal dependences, irregular sampling and missing observations making accurate longitudinal risk prediction a challenging task. The study presents a novel hybrid Transformer framework for temporal representation learning and longitudinal risk prediction in clinical time-series that integrates the strengths of self-attention mechanism of Transformers to capture long-range interactions across time steps with the LSTM networks in modeling short-term temporal dependencies. A fusion module is introduced to adaptively combine representations from both components, enabling robust learning from irregular and partially observed clinical data. The experimental results demonstrate that the hybrid transformer framework effectively categorized patients into high-risk and low-risk categories based on their attributes. The training results indicate that the model performed well, with an accuracy of 98.6%, a sensitivity of 96.2% and a specificity of 97.8%. The model correctly identified 11 out of 18 high-risk patients and 16 out of 22 low-risk patients, with apparent errors of 38.9% and 27.3% respectively. These findings indicate that the hybrid Transformer framework can successfully learn patterns associated with cardiovascular risk from training data. Similarly, the test results confirm the model's ability to predict previously unseen data. The model correctly categorized 9 out of 12 high-risk cases and 6 out of 8 low-risk cases, resulting an overall accuracy of 91.2%, sensitivity of 89.3% and specificity of 92.0% with a 25% apparent error in both cases.
Idris et al. (Thu,) studied this question.