Objective: This study aimed to develop a predictive model for early ambulation using clinical indicators obtained immediately after surgery in patients with proximal femoral fractures.

Methods: Patients who sustained a proximal femoral fracture and underwent surgery between April 2022 and April 2024, and whose medical records confirmed independent ambulation of at least 10 m before injury, were included. Those who died or had postoperative weight-bearing restrictions were excluded. The outcome variable was the ability to walk 10 meters without assistance at two weeks postoperatively. Predictive features included body mass index (BMI), abbreviated mental test score (AMTS), American Society of Anesthesiologists physical status, use of walking aids pre injury, intraoperative blood loss, and surgical method. A gradient boosting decision tree was used to develop the model.

Results: A total of 122 patients were included. Key predictors of ambulation at two weeks were AMTS, BMI, and the use of an intramedullary nail. The model achieved a recall of 72.7%, a precision of 66.6%, and an ROC AUC of 0.80 in an independent test dataset.

Conclusion: This study demonstrated the feasibility of a machine learning model to predict early ambulation using immediate postoperative indicators. As walking ability at two weeks is associated with long-term gait recovery and discharge outcomes, this model may aid in optimizing rehabilitation planning and discharge strategies.

Objective: This study aimed to identify the characteristics of cases in which walking level improves even in the later stages of convalescent rehabilitation, and to predict such cases.

Methods: The Walking LEVEL Scale (WaLS) was measured over time in 192 patients (brain-related and orthopedic diseases) admitted to a convalescent rehabilitation ward. The “delayed recovery group (DR)” was defined as the group showing improvement of WaLS in the later period, while the “non-delayed recovery group (NDR)” was defined as the group for whom improvement of WaLS did not extend into the same period. Then, changes in WaLS over time and each factor were compared to identify differences between DR and NDR. Using receiver operating characteristic analysis, cutoff values, sensitivity and specificity, likelihood ratios for the relevant factors were calculated and examined for the predictability of DR utilizing Bayesian estimation.

Results: WaLS temporal changes showed a sigmoid curve only in the DR of brain-related disease. Other groups showed logarithmic curves. The WaLS score at admission was selected as a significant variable. The positive and negative likelihood ratios for the score at cutoff values of 4 or less were 2.00 and 0.19 for brain-related disease, and 1.49 and 0.63 for orthopedic disease, respectively. Adopting the DR frequency of the participants in this study as the prior probability, the DR posterior probability of brain-related disease was calculated to be 5% using the negative likelihood ratio.

Conclusion: The negative likelihood ratio of the WaLS score at admission is useful for predicting DR for brain-related disease.