This study evaluates a deep-learning approach integrating serial 12-lead ECGs to accurately predict hospital admission for ED patients presenting with a range of acute cardiac symptoms, including chest pain, dyspnea, syncope, or presyncope. Using a large-scale, multicohort retrospective design, researchers leveraged the MIMIC-IV, MIMIC-IV-ED, and MIMIC-IV-ECG databases, encompassing over 30,000 ED visits with corresponding ECG and clinical data. The team developed a novel multimodal neural network that simultaneously integrates sequential ECG waveforms with dynamic vital signs, laboratory measurements, and static clinical variables to enhance predictive accuracy. The model was trained and externally validated on a cohort of 30,421 ED visits, with a subset of 11,273 patients having serial ECGs, to predict the likelihood of hospital admission within the ED encounter. Achieving an AUC of 0.924, the multimodal model significantly outperformed both ECG-only and tabular variable-only baselines, demonstrating superior discrimination for early identification of patients requiring hospital admission. Notably, the model provided accurate hospital admission predictions at the time of ECG acquisition and up to six hours before ED disposition, offering a potential tool for proactive resource allocation and early clinical decision-making.