Related Experiment Video
Updated: Aug 29, 2026
Deep-Learning Based Multi-Joint Synchronous Tracking for Objective Quantification of Hindlimb Locomotor Kinematics in Rats
Published on: April 3, 2026
Deep neural network-based prediction of walking independence after hip fracture surgery and analysis of contributing
Keisuke Nakamura1, Yuya Nagasawa2, Masayuki Shimizu3
1Department of Physical Therapy, School of Health Sciences, Shinshu University, Matsumoto, 390-8621, Japan; Department of Rehabilitation, Matsumoto City Hospital, Matsumoto, 390-1401, Japan.
Background:
Early prediction of walking independence after hip fracture surgery can support perioperative decision-making and rehabilitation planning. As functional recovery is dynamic, the predictive value of clinical variables may differ across postoperative phases. This study aimed to develop and internally validate predictive models for postoperative walking independence.
Methods:
We conducted a single-center retrospective cohort study of patients aged ≥65 years who underwent surgery for femoral neck or trochanteric fractures at an acute care hospital. Walking independence was defined as a Functional Independence Measure walking item score ≥5 and was assessed at postoperative week 1 (POW1) and postoperative week 4 (POW4). We developed and internally validated prediction models using logistic regression (LR), a nonlinear support vector machine (NLSVM), and a deep neural network (DNN). The data were split into training (80%) and test (20%) sets using stratified sampling. Class imbalance was addressed using simple oversampling in the training set. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) and standard classification metrics. For the DNN, variable contributions were examined using a partial derivative-based method.
Results:
A total of 368 patients were analyzed. At POW1, all models showed considerable discrimination (AUC, 0.892-0.926), with NLSVM achieving the highest AUC (0.926) and DNN showing the highest sensitivity (0.882). At POW4, LR and NLSVM demonstrated moderate discrimination (AUC, 0.800 and 0.756, respectively), whereas DNN showed comparable discrimination (AUC, 0.778) with the most balanced classification performance. Contribution analysis consistently highlighted positive contributions from early ambulation (within 2 postoperative days) and predominantly negative contributions from older age and longer waiting times to surgery and rehabilitation initiation.
Conclusions:
Predictive performance varied according to postoperative phase. The DNN provided the most balanced prediction at POW4 and, together with contribution analysis, may help characterize time-dependent patterns of walking recovery.
