Medical time-series analysis faces challenges in handling heterogeneous data and limited clinical samples. This paper proposes an enhanced Long Short-Term Memory (LSTM) framework integrating multimodal fusion and transfer learning for disease prediction and postoperative risk assessment. We first design a dual-attention LSTM model optimized for chickenpox case forecasting, achieving an 18.79% reduction in MAE compared to baseline models. For clinical applications, a multimodal LSTM architecture is developed to fuse dynamic physiological signals (e.g., blood pressure, pulse) with static patient metadata (e.g., age, medical history), enabling binary classification of cerebral infarction thrombectomy-induced hemorrhage with 12–15% AUC-ROC improvement. To address data scarcity, a three-stage transfer learning strategy is introduced: domain adversarial training aligns feature distributions between source (public datasets) and target (stroke patient data), while progressive fine-tuning adapts pre-trained temporal patterns to postoperative monitoring. Evaluations on three datasets—Hungarian chickenpox cases, HRV time series, and 453 cerebral infarction patient records—demonstrate superior performance, with 93% HRV prediction accuracy and 90% precision in hemorrhage risk detection. The framework provides interpretable decision support for clinical scenarios, balancing computational efficiency and predictive robustness.