Postoperative length of stay (LOS) is a key indicator of patient recovery and hospital resource utilization. We developed a multiclass deep learning model using preoperative structured EMR data from 81,084 surgical patients and externally validated it in an independent cohort of 56,575 patients. LOS was categorized into short (0–6 days), medium (7–13 days), and long (≥14 days). The model achieved AUROC of 0.876 (internal) and 0.858 (external), demonstrating robust generalizability. These results suggest the potential of EMR-based models for supporting clinical decision-making and hospital resource management.