Hip fractures, highly prevalent among the elderly, are characterized by poor prognosis, frequent complications, and elevated mortality rates (1–3). Length of postoperative stay (LOPS) serves as a critical indicator of clinical outcomes and nursing quality, objectively reflecting healthcare efficacy for elderly patients with hip fracture (4). The unplanned prolongation of LOPS may lead to an increased risk of complications in patients and higher medical expenses (5–7). Previous studies have revealed that each additional day of hospitalization increases complications risk by 5%, and hospitalization costs increase by 5–8% accordingly (8, 9). In addition, the unplanned shortening of LOPS may adversely affect postoperative rehabilitation of patients, leading to an increased risk of complications or even death (10, 11). Therefore, accurately predicting the specific LOPS in elderly patients with hip fracture is crucial for healthcare institutions and patients.

Prior retrospective studies (12–16) have mostly used classification methods to explore LOPS distribution in elderly patients with hip fractures, mainly focusing on whether LOPS is prolonged. However, artificially classifying a continuous variable like LOPS essentially involves the subjective assumption that there are inherent differences between subjects below and above the threshold. This practice may introduce bias and information loss, leading to unreliable results (17, 18). Therefore, establishing a statistical model with LOPS as a continuous variable for regression analysis can fully preserve the information on individual differences, allowing for more precise quantification of the association between each influencing factor and LOPS. Currently, the revelated study predicting LOPS as a continuous variable has mostly focused on patients undergoing total hip arthroplasty (THA) (19). Thus, constructing a model capable of accurately predicting LOPS in all elderly patients with hip fractures will be more conducive to ward risk management and the allocation of precise medical services.

Machine learning (ML) algorithms, such as neural networks, can better handle nonlinear relationships between variables and complex interactions among multiple factors, which is particularly critical for LOPS prediction as it is influenced by multiple interconnected clinical factors (20). Compared with traditional statistical methods, ML is more capable of revealing the interactive effects and underlying associations among multiple factors, and has been widely applied in the field of medical outcome prediction with excellent performance (21, 22). Meanwhile, neural network is the most commonly used algorithm for time-related outcome prediction (23), further supporting the feasibility of this approach. Therefore, this study aimed to develop and validate a neural network for predicting the specific LOPS in elderly patients with hip fractures.

This retrospective cohort study developed and validated a ML model based on EHRs of elderly patients with hip fractures to predict the specific LOPS. We confirmed age and ACCI as dominant predictors, and impact of surgical type was mediated through interactions with age and ACCI. The goal of this model was to provide medical staff with a reference for the specific LOPS through accurate calculation of continuous variables, enabling clinicians to anticipate clinical risks and support precise resource allocation in hospital management. Most importantly, the model only adopted three commonly available clinical variables, which would not increase the workload of medical staff or medical costs, making it easier to implement in hospital settings. Additionally, all variables included in the model could be determined at the early stage of admission, enabling early planning for clinical surgeons and hospital managers.

In a retrospective study collecting clinical data of patients who underwent THA from January 2014 to December 2019, the authors reported that for each additional day of hospital stay, the risk of complications in patients increased by 13% (34). However, LOPS reduction was mainly achieved through the rational allocation of medical service providers and medical resources based on conditions of patients and treatment needs, with the key being the accurate identification of the actual LOPS required by patients. Previous studies had linked risk factors to the LOPS through linear regression analysis (12, 13, 16), but most of these adopted threshold-based classification, focusing on whether LOPS of patients was prolonged or not. Variable continuity was interrupted by threshold-based classification, which could lead to information loss (35). In contrast, continuous variable prediction could fully retain original data information and accurately capture the inter-patient heterogeneity. In particular, continuous variable prediction could output quantitative values. Once the model training was completed, it could help focus attention on groups who actually need a longer LOPS and act as a tool to assist in medical decision-making when integrated into the current medical system.

In addition, hospital management need to strategically plan bed capacity and allocation to better match medical service providers with medical resources. Bed congestion could lead to resource waste, increased medical costs, and higher in-hospital risks for patients (36–38). On the other hand, bed vacancies might result in resource underutilization and financial losses (39). Hence, accurately predicting specific LOPS values is extremely important for hospital management. For example, the workload of nursing wards depended heavily on patient volume and their length of stay. The ability to forecast workload would enable managers to deliver cost-effective care. A cross-sectional study described the impact of nursing staff allocation on the length of hospital stay and hospitalization costs for patients undergoing hip or knee surgery (40). Yankovic et al. also described nursing staff allocation depend on the prediction of patients’ hospital stay duration, further supporting the model’s value for hospital management (41). Therefore, it was necessary to develop a prediction model for forecasting specific LOPS in elderly patients with hip fractures to optimize hospital management.

The BP-NN model could also offer an additional benefit, enabling patients to plan their schedules and medical expenses more systematically. Cao et al found that the LOPS of patients was 12 days (16), while another study indicated that the average LOPS for such patients was 16.6 days (31). In our study, the average LOPS of elderly patients with hip fractures was 17.42 days. Therefore, if the specific LOPS could be predicted based on the patients’ personal characteristics, rather than relying on empirical estimates, patients would be better able to coordinate life and medical-related affairs in advance. In terms of medical expenses, patients could also make reasonable plans and preparations to avoid the impact of cost issues on treatment. Although there was a large difference in LOPS among elderly patients with hip fractures between domestic and international settings due to factors such as post-discharge residence and healthcare systems, the prediction method proposed in this study was worthy of recommendation. Other regions could retrain the model on local datasets to attain clinically acceptable predictive performance. If necessary, hyperparameter optimization methods tailored to local data could be employed for iterative optimization, thereby realizing accurate LOPS prediction.

This study employed a BP-NN model to predict the specific LOPS in elderly patients with hip fractures based on EHR dataset. The model demonstrated robust performance in validation dataset (RMSE = 1.23 days, MAE = 1.57 days, MAPE = 7.69%), outperforming prior similar studies on continuous outcome prediction (31, 39). The coefficient of determination (R² = 0.83) was moderately good, which was likely attributed to the inherent challenge with continuous outcome variable, including complexes conditions, interacting clinical factors and data noise. However, predicting LOPS as a categorize continuous might fail to reflect individual differences (42), since groups above or below the threshold were considered as inherently distinct. Critically, the relative error quantified by MAPE was of great significance for patient risk stratification and resource allocation. Nevertheless, RMSE, MAE, and MAPE used in this study had inherent limitations, with RMSE sensitive to extreme errors, MAE failing to distinguish error distributions, and MAPE tendance to exaggerate errors for short LOPS. Future studies could further incorporate indicators such as symmetric mean absolute percentage error and consistency correlation coefficient to measure precision and accuracy. Despite this fact, a specific LOPS prediction still provided a practical reference for individualized clinical decision-making.

SHAP analysis provided an interpretation of the LOPS prediction model, with age and ACCI being the main predictors, which were observed in previous retrospective analyses (16). With increasing age, elderly patients with hip fracture experienced varying degrees of decline in functional reserve, stress tolerance, and immunity, making them highly susceptible to various complications (43, 44), which may require an extended recovery period, thereby prolonging their LOPS. Sensitivity analysis with age treated as a categorical variable showed that the very old elderly group had the longest LOPS. This may be associated with factors such as relatively poor baseline health status and a higher burden of comorbidities in this population. Although age is an immutable clinical factor, this finding of the present study still holds important implications for the clinical risk management of specific patients and optimal allocation of medical resources. Cao et al. also reported that a high ACCI was associated with prolonged LOPS (16). ACCI cannot only serve as an indicator to evaluate the severity of preoperative diseases of patients (45) but also exhibit significant accuracy in predicting postoperative complications (46), and the severity of diseases and the incidence of complications largely affected patients’ LOPS (47). Although surgical type was a predictor of LOPS, there was no statistically significant difference in LOPS among the three surgical types after adjusting for confounding factors, suggesting that its independent effect on LOPS was modulated by age and ACCI. The specific mechanism underlying this result may be as follows: on the one hand, both age and ACCI were strongly correlated with LOPS in this study and served as the main dominant factors determining LOPS. On the other hand, the selection of surgical type in clinical practice was not an independent decision. Surgeons inherently made individualized decisions based on the patients’ age and health status, leading to a natural intertwining of the effects of surgical type with those of age and ACCI, and this ultimately results in the independent effect of surgical type being masked.

We interpret the potential limitations of this study in light of its findings. First, the retrospective design might introduce inherent selection and information biases. Although complete data of participants were obtained through standardized EHRs, there were still a few types of cases not included in this study. When patients with such transfer characteristics are encountered again, the model may have difficulty predicting their LOPS accurately. Second, this model was constructed solely based on data from elderly fracture patients in this region. However, LOPS may vary among patients in different countries, regions, and areas with varying economic levels. Therefore, it is necessary for other regions to develop models using local data to verify its generalizability. Third, this study collected only data from patients at admission based on EHRs to facilitate early intervention, while potential variables, such as frailty (48), preoperative medications (49), preoperative waiting time (admission-to-surgery interval), and American Society of Anesthesiologists’ (ASA) classification, were not included in model construction, which may limit the model’s performance. Fourth, this model was constructed based on the training dataset split in an 8:2 ratio. Although we supplemented 10-fold cross-validation to evaluate the model’s internal stability, no external validation has been conducted, and its external generalizability remains unclear. Future studies can perform multi-center external validation to clarify the model’s generalizability, and thereby enhance its clinical applicability. Fifth, this study constructed the LOPS prediction model only from the perspective of patient safety and did not incorporate objective socioeconomic factors such as income, medical expenses, and medical insurance status, so further research is needed to identify and overcome potential barriers in practical application.

This study developed and validated a BP -NN model using the EHRs of elderly patients with hip fractures to predict the specific LOPS in this population. The model identified age and ACCI as the strongest predictors, while also revealing that the impact of surgical type was not independent but was modulated through complex interactions with these patient-specific factors. These predictions indicate the potential to assist surgeons in making personalized decisions regarding the LOPS of elderly patients with hip fracture by considering the confluence of patient age, comorbidity, and planned surgical type, and in managing their in-hospital safety, as well as enable hospital administrators to more effectively plan medical resources strategically.