Abstract
<title>Abstract</title> <p>Background Total hip arthroplasty (THA) is one of the most common elective surgeries in the U.S., but accurately predicting operative time and hospital stay before surgery remains difficult. Better estimates can improve OR scheduling, discharge planning, and resource use. Purpose This study developed and compared deep learning models with basic statistical approaches to predict THA resource use from preoperative patient data available at booking. Methods A total of 307,726 elective unilateral THA cases from the ACS NSQIP (American College of Surgeons National Surgical Quality Improvement Program) registry (2014–2023) were analyzed. Multiple models were developed to predict duration of surgery (DOS) and length of stay (LOS). Performance was compared to a simple mean model using mean absolute error (MAE), accuracy within clinical tolerance ranges, and binary classification accuracy for scheduling and discharge decisions. Results The deep learning model showed better technical performance, with an MAE of 24.2 minutes for DOS and 0.82 days for LOS. However, this advantage disappeared in clinically relevant categories. For predicting cases within a standard operative block (≤ 120 minutes), it outperformed the mean model by 0.1%, and for identifying short-stay patients (≤ 2 days), the improvement was only 3.5%. Conclusion Although deep learning achieved lower mean absolute error, gains in operationally relevant classification thresholds were small for DOS (0.1%) and modest for LOS (3.5% at the ≤ 2-day threshold). For highly standardized procedures such as THA, registries capturing patient-level variables alone may impose a practical ceiling on the operational utility of complex models. Whether incorporating institution- and surgeon-level operational variables would yield more clinically meaningful gains remains an open question.</p>