Deep Learning Predicts Postoperative Mobility, Activities of Daily Living, and Discharge Destination in Older Adults from Sensor Data

Abstract

What are the main findings?

Deep learning models applied to lumbar IMU data predict postoperative mobility (CHARMI) and activities of daily living (Barthel Index) in older surgical patients, with R² values of 0.65 and 0.70, respectively.

Recommended discharge destinations were predicted with 82% accuracy using lumbar IMU data and deep learning.

What is the implication of the main finding?

IMU-based assessment and deep learning offer the potential for automated, objective, and continuous monitoring of functional recovery.

The growing proportion of older adults in the population necessitates improved methods for assessing functional recovery. Objective, continuous monitoring using wearable sensors offers a promising alternative to traditional, often subjective assessments. This study aimed to investigate the utility of inertial measurement unit (IMU)-based data, combined with deep learning, to predict postoperative mobility, activities of daily living, and discharge destination in older adults following surgery. Data from the SURGE-Ahead project was analyzed, involving 39 patients (mean age 79.05 years) wearing lumbar IMU sensors for up to five postoperative days. Deep learning models (TabPFN) were applied and validated using leave-one-out cross-validation to predict the Charité Mobility Index (CHARMI), the Barthel Index, and discharge destination. The TabPFN model achieved R² values of 0.65 and 0.70 for predicting CHARMI and Barthel Index scores, respectively, with moderate to strong agreement with human assessments (weighted kappa ≥ 0.80). Discharge destination was predicted with an accuracy of 82%. The z-channel IMU data and parameters related to walking bouts were most predictive of outcomes. IMU-based data, combined with deep learning, demonstrates potential for automated functional assessment and discharge decision support in older adults following surgery.