# Automatic multi-IMU-based deep learning evaluation of intensity during static standing balance training exercises

## TL;DR

This study uses wearable sensors and deep learning to automatically assess balance exercise intensity, which could help monitor home-based rehabilitation.

## Contribution

A novel CNN-based model using IMU data to estimate physical therapists' intensity ratings for balance exercises.

## Key findings

- A CNN model using data from 13 IMUs achieved RMSE of 0.66, comparable to PT inter-rater variability.
- Model performance stabilized with four strategically placed IMUs on thighs and back.

## Abstract

Effective balance rehabilitation requires training at an appropriate level of exercise intensity given an individual’s needs and abilities. Typically balance intensity is assessed through in-clinic visual observation by physical therapists (PTs), which limits the ability to monitor and progress intensity during home-based components of training programs. The goal of this study was to train and evaluate machine learning models for estimating physical therapists’ perceived balance exercise intensity using data from full-body wearable sensors to support the development of home-based training exercise dosage monitoring.

Balance exercise participants (n = 47) participated in a single-day balance training session where they were filmed performing static standing exercises at various levels of intensity. Kinematic data from 13 full-body wearable inertial measurement units (IMUs) and self-ratings of balance intensity were also collected. An additional cohort of PT participants (n = 42) was recruited to watch the videos of the balance exercise participants and provide ratings of balance intensity. The mean PT rating for each video was used as a ground truth (GT) label of balance intensity. We trained and evaluated Convolutional Neural Networks (CNN)-based models to predict balance intensity based on performance as captured through the IMUs. Model performance was evaluated by calculating the root-mean-square error (RMSE) of predications. A sensitivity analysis was also performed to assess the effect of the number of IMUs used on model performance.

Models trained on orientation derived from all 13 IMUs achieved good predictive performance as indicated by a RMSE of 0.66 [0.62, 0.69], which was within the threshold defined by typical inter-rater variabilities between PTs (RMSE of 0.74 [0.72, 0.76]). Sensitivity analysis indicated that model performance stabilized at four sensors with the best performance corresponding to sensors placed on both thighs and the lower and upper back.

Findings from this study indicated that balance intensity assessment can be achieved through wearable sensors and a CNN model, which could support the supervision and effectiveness of home-based balance rehabilitation.