Towards an AI-Based In-Bed Posture Detection System for Pressure Injury Prevention

Lindsay Stern, Atena Roshan Fekr

Keywords: Posture Detection, Classification, Pressure Injury, Deep Learning, Convolutional Neural Networks, Vision Transformer

TL;DR: Investigating the use of a pressure sensitive smart mat and neural networks to classify 10 sub-postures using pressure distribution image for pressure injury prevention.

Abstract: Pressure injuries (PIs) are common wounds among patients with decreased mobility who are unable to periodically redistribute their body weight. The most common technique to prevent PI development is through frequent repositioning, often requiring support from caregivers, which can be a costly and laborious task. Therefore, this paper investigates the use of a pressure sensitive sheet to automatically capture in-bed body postures to prevent PI development. Five Neural Networks were evaluated to classify 10 sub-postures using pressure distribution images. Two techniques were explored: directly classifying all 10 postures, and a hierarchical architecture. Although the hierarchical architecture with the ShuffleNet algorithm achieved the highest F1-Scores of 99.75% ± 1.43% for holdout (20% test set) and 93.53% ± 7.37% for Leave-One-Subject-Out (LOSO) cross-validation, direct classification provides more stable results. These results suggest that this approach has promising potential to detect common sub-postures and could be used to remind caregivers to facilitate timely repositioning, thereby preventing PI development.

Track: 4. AI-based clinical decision support systems

Registration Id: PHN5FZ4XT4H

Submission Number: 214