A retrospective study on 734 elderly hip fracture patients in Yichang Central People's Hospital (2016–2022) developed a machine learning model to predict postoperative length of stay (LOPS), with a mean value of 17.42 ± 3.77 days. Among the patients, 68.53% were women, the most frequent femoral neck fracture (59.26%) and hemiarthroplasty (41.96%). Predictors of LOPS were age, age-adjusted Charlson comorbidity index (ACCI) and surgical type, selected by Pearson correlation analysis and LASSO regression. The backpropagation neural network (BP-NN) model achieved an RMSE of 1.23 days, an MAE of 1.57 days, a MAPE of 7.69%, and an R² of 0.83, with most predictions within the 30% error threshold. SHAP analysis showed ACCI and age as main factors. Patients over 90 years of age had the longest LOPS (21.99 ± 0.66 days). The model aids clinical decision-making and resource planning, but requires external validation.