This paper presents a novel parallel-stream long-term memory (LSTM) framework for predicting daily hospitalizations per week in the United States. To address the inaccuracy of existing prediction models during the COVID-19 pandemic, we incorporate social hospitalization proximity (SPH), a spatiotemporal feature based on the social connectivity index of Meta. SPH captures the transmission dynamics over space and time by reflecting weekly population movement. Our model captures both short-term and long-term temporal dependencies, and adjusts prediction consistency and error through a multi-horizontal ensemble strategy. The superiority of the proposed model is confirmed by comparative evaluation with COVID-19 prediction hub ensemble models during the delta and Omicron variant spread periods. In particular, during the Omicron variant spread period, the prediction performance is improved by an average of 27, 42, 54, and 69 hospitalizations per week for the 7-day, 14-day, 21-day, and 28-day forecast periods, respectively. We validate the predictive power of SPH through data removal experiments, highlighting the importance of spatiotemporal features in complex epidemiological modeling of infectious disease spread.