2025 年 10 巻 論文ID: 20250039
Objectives: Hip fractures in aging populations pose a major healthcare burden, and predicting discharge motor function may enable timely interventions. We aimed to develop a machine learning model with statistically validated reliability to predict motor Functional Independence Measure (mFIM) scores at discharge in patients who underwent bipolar hip arthroplasty (BHA).
Methods: This retrospective study was conducted at a regional hospital providing integrated care. A total of 201 hips treated with BHA for femoral neck fractures were analyzed. The primary outcome was the discharge mFIM score. Ten predictors were assessed: age; sex; presence of hypertension, diabetes, or heart failure; body mass index; Hasegawa Dementia Scale-Revised (HDS-R) score; time from admission to surgery; time from surgery to transfer to the convalescent ward; pre-fracture mobility status; pre-fracture independence level; and mFIM score at transfer. Six machine learning models were developed with hyperparameter tuning. Feature importance was evaluated using SHapley Additive exPlanations (SHAP), and results were compared with multiple regression for consistency.
Results: Light gradient boosting machine achieved the highest predictive performance (R2 = 0.84). SHAP analysis revealed that the most influential predictors were mFIM score at transfer to the rehabilitation ward, HDS-R score, and pre-fracture independence level. These factors were identified as significant variables in multiple linear regression.
Conclusions: We developed a reliable and interpretable machine learning model to predict discharge mFIM scores in patients who underwent BHA. Key predictors of the model were supported by SHAP and multiple regression analysis. This model can assist clinicians in setting appropriate rehabilitation goals and discharge plans early during recovery.
With the global progression of population aging, the prevalence of femoral fractures is predicted to increase, reaching approximately 6.26 million cases annually by 2050.1) Proximal femoral fractures in older adults are associated with increased costs of medical care and long-term care, as well as higher mortality rates.2) The growing number of such patients may place a significant strain on healthcare resources. To address this situation, it is essential to establish individualized rehabilitation goals and facilitate early discharge planning, which requires the development of reliable predictive indicators that can support such clinical decision-making.3)
Following surgical treatment during the acute phase, patients with hip fractures are typically transferred to a convalescent rehabilitation ward (CRW) for further functional recovery and eventual return to home.4) However, accurately predicting functional outcomes at discharge remains challenging owing to substantial inter-individual variability in pre-fracture physical function, cognitive status, and comorbidities among older patients. Previous studies have identified several factors influencing discharge function, including age, cognitive function, comorbidities, and pre-fracture activities of daily living (ADLs).5,6,7) To comprehensively evaluate these variables, machine learning-based predictive models have been developed8); however, these models still face limitations in clinical applicability, and further improvements in predictive accuracy are needed. Accurate prediction of functional outcomes at the time of transfer to the CRW may facilitate early determination of appropriate rehabilitation goals, length of hospitalization, and discharge destination, thereby enabling more efficient allocation of limited healthcare resources.
This study focused on patients with femoral neck fractures who underwent bipolar hip arthroplasty (BHA). This focus was based on the rationale that BHA is applied to a relatively uniform fracture type and surgical technique, thereby reducing potential bias in the construction of the predictive model. This study aimed to develop a machine learning model for predicting motor function at discharge, evaluated using the motor Functional Independence Measure (mFIM), utilizing patient background information and physical function assessments at the time of CRW admission. We also aimed to statistically evaluate the accuracy and reliability of the model.
This retrospective study was conducted at a regional hospital with 250 beds, including 50 beds dedicated to CRW. A notable advantage of this setting is that the entire care pathway—from admission to surgery, postoperative rehabilitation in the CRW, and discharge—can be completed within the same institution. Between January 2018 and December 2024, among the 347 hips with femoral neck fractures treated at the study hospital, 57 hips with Garden stage I or II fractures underwent internal fixation, whereas the remaining 290 hips with Garden stage III or IV fractures were treated by performing BHA. Among this latter group, patients with a total of 222 hips were transferred postoperatively to the CRW, where the rehabilitation program was successfully completed. Patients with conditions potentially interfering with the rehabilitation course, such as prior neurological or psychiatric disorders, perioperative complications, or missing functional data, were excluded from the analysis. Notably, these exclusion criteria were intended to apply to conditions such as severe sequelae of strokes or schizophrenia that would markedly hinder participation in postoperative rehabilitation. Conversely, dementia was not considered an exclusion criterion, because mild to moderate cognitive impairment is commonly observed in older adults. Finally, 201 hips from 195 patients were included in the final analysis (Fig. 1). A total of 17 cases included patients with a prior history of proximal femoral fracture on the contralateral side, with a mean interval of 39.2 months (range: 3–144 months) from the previous discharge. Among these individuals, six patients underwent bilateral BHA during the study period, and in all these cases, the second fracture occurred more than 4 months after discharge from the initial admission (mean, 20.8 months; range, 4–36 months). This study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Shunan Memorial Hospital (Approval No. R07-03). The requirement for informed consent was waived by the committee owing to the retrospective nature of the study and the use of anonymized data.
Flowchart of patient recruitment.
All surgeries for displaced femoral neck fractures were performed by the same orthopedic surgical team. The Taperloc Complete Hip Ste (Zimmer-Biomet; Warsaw, Indiana, USA) was used for all procedures. A posterolateral approach was used in all cases. Among the short external rotators, only the inferior gemellus and obturator externus were detached, and the remaining muscles were preserved. Postoperatively, in the acute care ward, physical therapy (PT) was administered six times per week, with each session lasting from 20 to 40 min. The primary focus here was on early mobilization, training in the use of portable toilets, and gait rehabilitation. Having initially confirmed satisfactory wound healing and that weight-bearing and gait training on the affected side could be initiated, patients were transferred to the CRW. At this point, rehabilitation transitioned from the acute phase to the convalescent phase.
In the CRW, where this study was conducted, a multidisciplinary team provided comprehensive rehabilitative care. Medical management, nursing, PT, occupational therapy (OT), and social work services were coordinated. PT was provided eight times per week, with each session lasting 40–60 min, focusing on transfer training, gait training, stair climbing, and range of motion exercises. OT was conducted six times per week in sessions of 40–60 min and was focused on improving basic ADLs, safe bathing and bed mobility to prevent dislocation, cognitive assessment and stimulation, and training in the use of assistive devices with safety education. A multidisciplinary team conference was held weekly to evaluate progress and adjust treatment plans, during which individualized rehabilitation plans were reviewed and revised as necessary.
Outcome Measures and Patient CharacteristicsData were retrospectively collected from electronic medical records. Functional status was assessed using the Functional Independence Measure (FIM),9) which consists of 18 items covering three domains: ADL, mobility, and cognitive function. Each item is rated on a 7-point scale, with higher scores indicating greater independence. The FIM includes 8 items related to ADL, 5 items related to mobility, and 5 items related to cognition. The motor FIM (mFIM) comprises 13 items related to ADL and mobility, with a maximum possible score of 91. FIM assessments were conducted at two time points: upon admission to the CRW and at discharge. Evaluations were performed during multidisciplinary team conferences.
Pre-fracture independence level and mobility status were assessed by nursing staff at the time of hospital admission based on interviews with patients and/or their families. Independence level was evaluated using the “Independence Level in Daily Living for the Elderly,” a standardized classification system developed by the Japanese Ministry of Health, Labour, and Welfare. This system categorizes patients into four levels: J (independent, able to go out alone), A (independent indoors but requires assistance to go out), B (mostly bedridden but able to maintain a sitting position), and C (completely bedridden and fully dependent for ADLs). Patients classified into Level C were excluded from the study. Mobility status was evaluated on a four-level scale: independent walking, cane-assisted walking, walker-assisted walking, and wheelchair use. All variables were treated as ordinal variables based on their respective rankings in statistical analysis.
Cognitive function was assessed using the Hasegawa Dementia Scale-Revised (HDS-R), which was administered by nursing staff after admission. The HDS-R is a widely used dementia screening tool that is used to evaluate multiple cognitive domains, including orientation and memory. In addition, three common chronic conditions frequently observed in older adults—hypertension, diabetes mellitus, and heart failure—were selected as comorbidities because of their potential influence on prognosis and rehabilitation course. Other background variables included age at the time of surgery, body mass index (BMI), time from admission to surgery, and time from surgery to CRW admission. These variables are listed in Table 1. All continuous variables were summarized as mean ± standard deviation, and categorical variables were expressed as counts and percentages.
| Characteristic | (n = 201 hips) |
| Age at surgery, years | 82.6 ± 9.3 |
| Female | 158 (78.6%) |
| Comorbidities | |
| Hypertension (hips) | 57 (28.4%) |
| Diabetes mellitus (hips) | 38 (18.9%) |
| Heart failure (hips) | 6 (3.0%) |
| HDS-R (/30) | 17.9 ± 9.4 |
| BMI, kg/m2 | 20.6 ± 3.6 |
| Time waiting for surgery, days | 5.2 ± 2.2 |
| Time from surgery to transfer to CRW, days | 13.3 ± 7.5 |
| Length of hospital stay, days | 61.1 ± 25 |
| Pre-fracture mobility status | |
| Independent walking | 92 (45.8%) |
| Walking with a cane | 44 (21.9%) |
| Using a walker | 54 (26.9%) |
| Wheelchair bound | 11 (5.5%) |
| Pre-fracture independence level | |
| J: Independent in daily life; goes out alone despite disabilities | 93 (46.3%) |
| A: Independent indoors; does not go out without assistance | 86 (42.8%) |
| B: Assisted indoors; mostly bedridden but able to sit up | 22 (10.9%) |
| mFIM score at transfer to CRW | 46.3 ± 21.7 |
| mFIM score at discharge | 65 ± 23.9 |
Data presented as mean ± standard deviation or number (percentage).
The dataset was randomly split into training and holdout (test) datasets at a ratio of 80:20. Five-fold cross-validation was used for internal validation, and the test dataset was reserved for final model evaluation. The target variable was mFIM score at discharge, which was predicted using ten input features: age at surgery, sex, comorbidities (hypertension, diabetes mellitus, and heart failure), HDS-R score, BMI, time from admission to surgery, time from surgery to CRW admission, pre-fracture mobility status, pre-fracture ADL independence level, and mFIM score at the time of transfer to CRW. The computational environment was implemented in Google Colaboratory (Google LLC, Mountain View, CA, USA) using Python, version 3.11.13.
Six regression models were compared: linear regression, Lasso regression (Least Absolute Shrinkage and Selection Operator), support vector regression, random forest (RF), light gradient boosting machine (LightGBM), and extreme gradient boosting. These models represent a range of algorithmic approaches, including linear models, regularized regression, nonlinear kernel-based methods, bagging-based ensemble methods, and boosting-based ensemble methods. This diversity was intended to balance interpretability and generalization performance, thereby optimizing predictive accuracy. All models were trained using five-fold cross-validation, and hyperparameters were optimized using grid search. The hyperparameter ranges explored and the final selected values for each machine learning model are summarized in Table 2.
| Model | Hyperparameter | Search range | Selected value |
| Linear regression | fit_intercept | [True, False] | True |
| positive | [False, True] | False | |
| Least Absolute Shrinkage and Selection Operator | alpha | [0.001–10.0] | 0.464 |
| Support vector regression | kernel | [‘linear’, ‘rbf’] | linear |
| C | [0.1, 1, 10] | 1 | |
| gamma | [‘scale’, ‘auto’] | scale | |
| Random forest | n_estimators | [100, 200] | 200 |
| max_depth | [None, 10, 20] | None a | |
| min_samples_split | [2, 5] | 5 | |
| max_features | [‘auto’, ‘sqrt’] | sqrt | |
| Light gradient boosting machine | n_estimators | [100, 200] | 200 |
| learning_rate | [0.05, 0.1] | 0.05 | |
| max_depth | [3, 5, 7] | 3 | |
| Extreme gradient boosting | n_estimators | [100, 200] | 200 |
| learning_rate | [0.05, 0.1] | 0.1 | |
| max_depth | [3, 5, 7] | 3 |
All hyperparameters were tuned using grid search with fivefold cross-validation.
a Indicates that nodes are expanded until all leaves are pure.
The predictive performance of the regression models was evaluated using the coefficient of determination (R2), mean absolute error (MAE), and mean squared error (MSE). R2 reflects the goodness of fit of the model, whereas MAE and MSE quantify the magnitude of prediction errors. MAE is less sensitive to outliers, whereas MSE (and its square root, RMSE) penalizes larger errors more heavily. These metrics were calculated for the training and validation datasets. Model performance was assessed using validation and test datasets.
For models that demonstrated particularly high accuracy, SHapley Additive exPlanations (SHAP) summary plots (dot plots) were used to visualize the contribution of each feature to the predictions and to assess interpretability. In the SHAP summary plot, features are listed on the vertical axis in descending order of importance, whereas the horizontal axis displays SHAP values, representing the magnitude and direction of the contribution of each feature to the predicted outcome. Each dot represents an individual sample, with color indicating the feature value (red = high, blue = low).
Multiple Linear Regression Analysis of Factors Affecting Discharge mFIMTo identify factors associated with mFIM score at discharge, multiple linear regression analysis was performed using the same ten input features used in the machine learning models. In addition, to evaluate functional improvement during the convalescent phase, a paired t-test was conducted to compare the mFIM scores at transfer to the convalescent ward and at discharge. A significance level of P < 0.05 was applied, and all statistical analyses were conducted using SPSS version 25 (IBM, Armonk, NY, USA).
Power AnalysisAn a priori power analysis was conducted using G*Power version 3.1.9.7 (Düsseldorf, Germany) to determine the required sample size for multiple linear regression. With the assumptions of a medium effect size (f2 = 0.15), an alpha level of 0.05, a statistical power of 0.80, and ten predictors, the minimum required sample size was calculated to be 76. However, the actual sample size used in this study was 201. This larger sample size was chosen to enhance the stability and generalizability of the regression model. Additionally, a post hoc analysis was performed based on the results of the regression analysis.
For the validation dataset, although no substantial differences in predictive performance were observed across models, RF (R2 = 0.77 ± 0.08, MAE=8.75 ± 1.39, MSE=127.44 ± 39.85) and LightGBM (R2 = 0.76 ± 0.09, MAE=8.45 ± 1.33, MSE=135.56 ± 49.94) showed the best performance in descending order. For the test dataset, both models demonstrated high predictive accuracy, with LightGBM (R2 = 0.84, MAE=7.47, MSE=95.96) outperforming RF (R2 = 0.82, MAE=8.08, MSE=111.64) (Table 3).
| Model | Validation (fivefold cross-validation on training dataset) | Test (hold-out test dataset) | |||||
| R2 | MAE | MSE | R2 | MAE | MSE | ||
| Linear regression | 0.74 ± 0.12 | 9.28 ± 2.39 | 148.20 ± 63.84 | 0.77 | 9.58 | 139.82 | |
| Least Absolute Shrinkage and Selection Operator | 0.74 ± 0.11 | 9.15 ± 2.32 | 144.95 ± 61.44 | 0.77 | 9.36 | 139.33 | |
| Support vector regression | 0.74 ± 0.11 | 9.17 ± 2.29 | 147.17 ± 60.35 | 0.77 | 9.5 | 142.77 | |
| Random forest | 0.77 ± 0.08 | 8.75 ± 1.39 | 127.44 ± 39.85 | 0.82 | 8.08 | 111.64 | |
| Light gradient boosting machine | 0.76 ± 0.09 | 8.45 ± 1.33 | 135.56 ± 49.94 | 0.84 | 7.47 | 95.96 | |
| Extreme gradient boosting | 0.75 ± 0.06 | 8.68 ± 1.11 | 137.02 ± 31.18 | 0.76 | 9.04 | 145.48 | |
Data presented as mean ± standard deviation.
SHAP summary plot analysis revealed that the top three features contributing to outcome prediction were consistent across RF (Fig. 2A) and LightGBM (Fig. 2B) models. These features were, in order of importance, mFIM score at transfer to the CRW, HDS-R score, and pre-fracture independence level. In all cases, higher values of these variables positively contributed to the predicted mFIM score at discharge.
SHAP summary plots for the RF model (A) and the LightGBM model (B). Each dot represents the SHAP value of one case. The x-axis indicates the contribution to the prediction (positive: right; negative: left), and the color represents the value of the feature (red: high; blue: low). The top three important features were consistent across both models and higher values of these features contributed positively to the predicted outcomes. Gray bars indicate the mean of the absolute SHAP values, representing the overall effect of each feature on model prediction.
Using discharge mFIM score as the dependent variable, the following predictors were found to be statistically significant, in descending order of standardized coefficients (β): mFIM score at transfer to the CRW [β=0.44, B=0.49, P < 0.001, 95% confidence interval (CI) 0.35–0.62], HDS-R score (β=0.33, B=0.83, P < 0.001, 95% CI 0.56–1.11), and pre-fracture independence level (β=0.21, B=3.41, P < 0.001, 95% CI 1.55–5.28) (Table 4). This ranking was consistent with the order of feature importance observed in the SHAP summary plots for the RF and LightGBM models. Furthermore, paired t-test analysis revealed a statistically significant improvement in mFIM scores during the convalescent phase (mean ± standard deviation: 18.8 ± 10.6, P < 0.001), thereby indicating a substantial functional gain between transfer to the CRW and discharge.
| Parameter | Standardized β | B (Unstandardized) | t value | P value | 95% CI |
| Age at surgery | 0.07 | 0.19 | 1.81 | 0.072 | −0.02 to 0.39 |
| Female | 0.03 | 1.76 | 0.86 | 0.394 | −2.31 to 5.83 |
| Comorbidities | 0.04 | 1.44 | 0.99 | 0.322 | −1.42 to 4.30 |
| HDS-R | 0.33 | 0.83 | 5.91 | < 0.001 | 0.56 to 1.11 |
| BMI | −0.06 | −0.36 | −1.56 | 0.12 | −0.82 to 0.10 |
| Time waiting for surgery | −0.02 | −0.24 | −0.64 | 0.522 | −0.99 to 0.50 |
| Time from surgery to transfer to CRW | −0.07 | −0.21 | −1.87 | 0.063 | −0.43 to 0.01 |
| Pre-fracture mobility status a | −0.07 | −1.60 | −1.39 | 0.167 | −3.88 to 0.67 |
| Pre-fracture independence level b | 0.21 | 3.41 | 3.61 | < 0.001 | 1.55 to 5.28 |
| mFIM score at transfer to CRW | 0.44 | 0.49 | 6.94 | < 0.001 | 0.35 to 0.62 |
R=0.88, R2 = 0.78, Adjusted R2 = 0.77, F (10, 190) = 67.29, P < 0.001.
a Lower score is better; b Higher score is better.
A post hoc power analysis was performed based on the R2 value (0.78) of the regression model. Given the large effect size (f2 = 0.354) and a sample size of 201 participants, the statistical power (1−β) was calculated to be greater than 0.999, indicating sufficient power for detecting the observed effect.
In this study, multiple machine learning models were developed and compared to predict motor function at discharge, which was evaluated using the mFIM. Among the models developed, RF and LightGBM had the highest predictive performance. For the test dataset, LightGBM achieved the highest R2 (0.84) and the lowest MAE (7.47), followed closely by RF (R2 = 0.82), both showing excellent accuracy. Feature importance analysis based on SHAP summary plots revealed the same top three contributing variables in both models: mFIM score at transfer to the CRW, HDS-R score, and pre-fracture independence level. Consistent results were observed in the multiple linear regression analysis. The agreement between the statistical and machine learning approaches supports the reliability and explainability of the predictive models developed in this study.
This study was conducted to develop a machine learning model to predict motor function at discharge, measured using the mFIM score, based on patient background information and physical function assessments at the time of transfer to the CRW in patients who underwent BHA. The developed models showed high predictive accuracy, with LightGBM yielding the best performance. In addition, the results of paired t-test analysis confirmed a statistically significant improvement in mFIM scores between transfer and discharge, therefore indicating that the predicted outcome represents a clinically meaningful measure of functional recovery. Shtar et al.8) previously developed a predictive model for discharge mFIM scores in a larger cohort of 1625 patients with proximal femoral fractures, reporting R2 values of 0.73 and 0.81 for the validation and test datasets, respectively. In contrast, our model, developed using a smaller, more homogeneous dataset of 201 BHA cases, achieved comparable or better accuracy, with R2 values of 0.76 and 0.84 for validation and test data, respectively. These findings support our initial hypothesis that BHA cases are less susceptible to modeling bias owing to the relatively consistent surgical procedures and fracture patterns, thereby enabling more stable and accurate prognostic predictions.
The machine learning model developed in this study was statistically reliable and explainable, suggesting its practicality and interpretability for clinical use. To enhance model interpretability for discharge mFIM score, we evaluated feature importance using SHAP alongside conventional multiple linear regression. The three variables identified as top predictors in RF and LightGBM models (mFIM score at CRW admission, HDS-R score, and pre-fracture independence level) represent clinically relevant indicators of cognitive and motor function, both of which are known to influence postoperative outcomes in patients with femoral neck fractures,10) thus reinforcing their importance from a medical perspective.
Although SHAP and multiple regression are fundamentally different analytical methods, the identification of common key predictors across both approaches further supports the interpretability and reliability of the prediction model. The standardized regression coefficients in multiple linear regression analysis indicate the relative influence of each predictor on the outcome variable across the entire dataset. In contrast, SHAP provides a quantitative visualization of the contribution of each feature to outcome prediction for individual cases, even in nonlinear models. Despite the methodological differences between these two approaches, their consistent identification of the same key predictors suggests that these features are robust predictors of discharge mFIM score. On the basis of multiple regression analysis, although the influence of features ranked below the top three failed to reach the level of statistical significance, the findings of SHAP analysis indicated that these variables, such as BMI, surgical waiting time, and comorbidities, may still make meaningful contributions to individual predictions. This accordingly highlights the potential utility of SHAP for capturing nuanced effects that may not be evident when applying traditional statistical methods. This makes it a valuable tool for bridging the gap between predictive modeling and clinical application.
In the field of proximal femoral fractures, machine learning is increasingly being applied to predict various outcomes such as primary fractures,11,12) secondary fractures,13) mortality,14,15) and functional outcomes.8) However, only a few studies have been conducted to statistically validate the reliability of these models using traditional analytical approaches. The present study is notable in that we developed a highly accurate machine learning model to predict discharge mFIM scores in patients who underwent BHA and further verified its validity through statistical methods. Clinicians can use this model to estimate a patient’s approximate physical function level at discharge as early as the time of transfer to a CRW. This may facilitate the early establishment of rehabilitation goals and discharge planning tailored to individual patients. Ultimately, this approach can contribute to optimizing the length of hospital stay and improving the efficiency of healthcare resource allocation.
Our findings offer valuable insights, but some limitations of this study should be acknowledged. First, this was a retrospective study conducted at a single institution, which may limit the external validity of the results. However, the single-center design also ensures consistency in the rehabilitation protocols, evaluation criteria, and documentation methods, thereby enhancing the quality and integrity of the dataset. To mitigate the risk of overfitting, the model was developed using a hold-out validation method, with a strict separation between the training and testing datasets. This approach allowed for an initial assessment of the generalizability of the model and was intended to simulate the application of the model to external clinical data. Therefore, future validations using multicenter datasets and prospective studies are necessary. Furthermore, considering the retrospective nature of the study, detailed information on patients’ previous diagnoses, such as osteoarthritis or vertebral compression fractures, was inconsistently recorded and, consequently, was not included in the analysis, even though these conditions may have influenced postoperative ADL recovery. Second, this study focused on patients with femoral neck fractures who underwent BHA. Previous studies have shown that fracture type and surgical procedure can influence the duration and outcomes of functional recovery.16,17,18) Therefore, the current model should only be applied to patients undergoing BHA, and further studies are warranted to extend its applicability to other fracture types and surgical methods. Third, inclusion of the mFIM score at transfer to the convalescent ward as a predictor may have partially masked the effects of some preoperative variables, because this variable may act as a mediator between early postoperative recovery and discharge outcomes. Although this variable enhanced the predictive performance and clinical utility of the model, caution is warranted when interpreting the relative importance of other predictors. Fourth, the exclusion of patients with perioperative complications or intensive care needs may have led to the selection of a cohort with relatively favorable recovery potential, thereby introducing a level of selection bias. Consequently, the generalizability of our findings with respect to more medically complex or frail populations may be limited. Furthermore, although some patients had a history of contralateral proximal femoral fracture, they were included as independent cases and were handled in the same manner as unilateral cases in the data analysis, which may also have introduced a potential source of bias. Finally, because this study was conducted solely among Asian patients, caution should be exercised when generalizing the findings to other racial and ethnic populations.
The machine learning model developed in this study achieved high predictive performance for discharge mFIM scores in patients with femoral neck fractures who underwent BHA. Furthermore, the consistency between the feature importance of the model and the results from traditional statistical analyses supports its clinical interpretability and reliability. This model may support individualized goal setting for rehabilitation and assist in making timely decisions regarding the length of hospital stay and discharge destination, thereby contributing to the efficient use of limited healthcare resources.
The authors declare no conflict of interest.
