Assessing Fatigue with Multimodal Wearable Sensors and Machine Learning

TL;DR

This study develops a multimodal wearable sensor system combined with machine learning to objectively assess cognitive and physical fatigue, outperforming existing methods in accuracy.

Contribution

The paper introduces a novel experimental setup with a sensor suit and ML models for fatigue detection, providing a new approach to fatigue assessment.

Findings

Random Forest detects physical fatigue with 80.5% accuracy

LSTM detects cognitive fatigue with 84.1% accuracy

Sensor-based system outperforms previous methods

Abstract

Fatigue is a loss in cognitive or physical performance due to physiological factors such as insufficient sleep, long work hours, stress, and physical exertion. It adversely affects the human body and can slow reaction times, reduce attention, and limit short-term memory. Hence, there is a need to monitor a person's state to avoid extreme fatigue conditions that can result in physiological complications. However, tools to understand and assess fatigue are minimal. This paper primarily focuses on building an experimental setup that induces cognitive fatigue (CF) and physical fatigue (PF) through multiple cognitive and physical tasks while simultaneously recording physiological data. First, we built a prototype sensor suit embedded with numerous physiological sensors for easy use during data collection. Second, participants' self-reported visual analog scores (VAS) are reported after each task to confirm fatigue induction. Finally, an evaluation system is built that utilizes machine learning (ML) models to detect states of CF and PF from sensor data, thus providing an objective measure. Our methods beat state-of-the-art approaches, where Random Forest performs the best in detecting PF with an accuracy of 80.5% while correctly predicting the true PF condition 88% of the time. On the other hand, the long short-term memory (LSTM) recurrent neural network produces the best results in detecting CF in the subjects (with 84.1% accuracy, 0.9 recall).