Abstract: This study proposes a hybrid prediction framework integrating a dynamically modified extended SEIR (Susceptible-Exposed-Infectious-Removed) model with an LSTM (Long Short-Term Memory) deep learning network to enhance the prediction accuracy of epidemic trends, based on COVID-19 pandemic data from Tokyo, Japan in 2022. Firstly, utilizing Tokyo's pandemic data from September 1 to October 12, 2022, the research first introduces dynamic time-varying vaccination rate parameters and spatial population density correction factors within the classical SEIR model framework to construct an extended SEIR model. Secondly, an LSTM network is incorporated to dynamically adjust the residuals between the predictions of the extended SEIR model and the actual data, leveraging its long-term and short-term memory mechanisms to learn nonlinear patterns from historical data and achieve real-time compensation for prediction errors. The results demonstrate that the optimized hybrid model exhibits a high degree of consistency with actual data in predicting the progression of the Tokyo pandemic, with its peak prediction performance significantly outperforming that of the standalone extended SEIR model.