Oral Presentation 51st International Society for the Study of the Lumbar Spine Annual Meeting 2025

Development and Validation of a Machine Learning Model for Classifying Postoperative Physical Activity Patterns Using Wearable Accelerometer Data in Patients with Lumbar Degenerative Disease (115775)

Tatsuya Endo 1 , Naohiko Kanemura 2 , Keita Sato 1 , Masumi Iwabuchi 1 , Toshikazu Ito 1 3 , Osamu Shirado 1
  1. Department Of Rehabilitation/ Orthopaedics & Spinal surgery AIZU Medical Center, Fukushima Medical University, Aizuwakamatsu City, FUKUSHIMA, Japan
  2. Saitama Prefectural University, Koshigaya
  3. Hokkaido Chitose College of Rehabilitation, Chitose City, Hokkaido, Japan

INTRODUCTION:
In recent years, reports on physical activity using wearable accelerometers have increased in the field of spinal cord and spine diseases. However, the vast amount of information can make data interpretation challenging, creating a barrier to clinical application. The purpose of this study was to develop a machine learning model to classify physical activity patterns using raw accelerometer data and to verify its accuracy.

METHODS:
The subjects were 56 patients (26 males, 30 females; mean age 70.7 years) who underwent posterior lumbar interbody fusion (PLIF) for lumbar degenerative disease between April 2022 and May 2024. We used a triaxial wearable accelerometer (OMRON HJA-750C Active Style Pro) to record activity intensity, walking time, and step count between 4 to 14 days postoperatively. Patients were categorized based on activity levels using hierarchical clustering analysis. The three resulting cluster classifications were used as training data to create a machine learning model using raw accelerometer data from postoperative days 4 to 8, and model accuracy was evaluated.

RESULTS:
Cluster analysis classified patients into three groups. Cluster 1 was the least active, older than Cluster 3, and exhibited delayed postoperative improvement in activities of daily living (ADL) with a sedentary lifestyle. Cluster 2 demonstrated moderate activity levels, showing lower moderate-intensity activity time, step count, and walking time compared to Cluster 3, with a higher rate of cane use at discharge. Cluster 3 was the most active, showing a favorable postoperative course. Among the models evaluated, k-nearest neighbors (kNN) and logistic regression demonstrated the highest classification performance (accuracy: 77.78%, precision: 80.56%, recall: 83.33%, F-measure: 77.46, ROC-AUC: 0.84).

DISCUSSION:
This study demonstrated that early postoperative physical activity patterns during hospitalization can be accurately predicted based on raw accelerometer data, enabling classification into three distinct groups according to physical activity levels. Using this predictive model, patients’ physical activity patterns can be anticipated within the first postoperative week, allowing for tailored rehabilitation program planning according to individual activity patterns. Additionally, it allows for setting specific step goals for the following week. Future studies may expand clinical applications by exploring long-term recovery patterns and their associations with final clinical outcomes.

 

 

67382364b98cb-Endo+T,+ISSLS2025+%E7%94%BB%E5%83%8F.png