Body Temperature Monitoring for Regular COVID-19 Prevention Based on Human Daily Activity Recognition

Lei Zhang, Yanjin Zhu, Mingliang Jiang, Yuchen Wu, Kailian Deng, Qin Ni2021 · DOI

TL;DR

This paper introduces a wearable system that combines human activity recognition (HAR) with body surface temperature monitoring. By integrating temperature data into HAR algorithms, they achieve improved activity recognition accuracy, which is crucial for effective COVID-19 prevention strategies, especially considering how body temperature can fluctuate with activity.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical need for effective public health measures, including temperature monitoring for fever detection. However, existing methods for body temperature measurement can be inaccurate, especially in dynamic environments where factors like physical activity and ambient temperature can influence readings. Furthermore, while Human Activity Recognition (HAR) using wearable sensors is well-established for applications like health and fitness, it has rarely incorporated body temperature as a key data point. This paper addresses the challenge of developing a more robust and accurate system for continuous body temperature monitoring that can adapt to a person's daily activities, thereby improving the reliability of fever detection for COVID-19 prevention.

Methods

The researchers developed a wearable device equipped with an STM32 microcontroller, a 6-axis inertial sensor (accelerometer and gyroscope), and a temperature sensor (LMT70). This device collects raw data on human activity and body surface temperature. The collected data undergoes pre-processing, including signal standardization and data stacking (using a sliding window approach with varying overlap percentages).

For Human Activity Recognition (HAR), they implemented and compared several machine learning (ML) and deep learning (DL) algorithms, including Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Stochastic Gradient Descent (SGD), Logistic Regression (LR), Naive Bayes (NB), Stacked Denoising Autoencoder (SDAE), Random Forest (RF), Extra Trees (ET), and Deep Forest (DF). The performance of these algorithms was evaluated using metrics like accuracy, precision, recall, and F1-score, considering both the inertial sensor data alone and the combined inertial and temperature sensor data.

A key aspect of their work is the investigation into the impact of including body temperature data on HAR accuracy. They also analyzed the temperature deviations observed during different activities and proposed a statistical approach using normal distribution curves to help calibrate body temperature readings based on activity levels.

Datasets

The study utilized a newly collected dataset from 10 healthy participants performing four distinct daily activities: sitting, walking, walking upstairs, and walking downstairs. A total of 31,713 data entries were collected, with specific counts for each activity (8365 for sitting, 9416 for walking, 6581 for walking upstairs, and 7351 for walking downstairs). This dataset, which includes inertial sensor data and body surface temperature, serves as the benchmark for their HAR model evaluation.

Key Insights and Ideas

  • Temperature Data Enhances HAR: Incorporating body surface temperature data significantly improved the accuracy of human activity recognition across various ML and DL algorithms, demonstrating its value beyond just fever detection.
  • Activity Influences Body Temperature: A surprising finding was that participants' body surface temperature was consistently lower during dynamic activities (walking, upstairs, downstairs) compared to sitting. This highlights the need to account for activity levels when interpreting temperature readings for fever screening.
  • Random Forest and Extra Trees Excel: Among the tested algorithms, Random Forest (RF) and Extra Trees (ET) consistently showed superior performance in terms of accuracy and efficiency, especially when combined with data stacking.
  • Data Stacking Improves Robustness: Data stacking, particularly with a 50% overlap, proved effective in improving the accuracy and robustness of the HAR models, especially for differentiating between similar dynamic activities like walking upstairs and downstairs.
  • Statistical Calibration for Temperature: The study proposes using normal distribution curves of temperature deviations based on activity to provide a more accurate and context-aware interpretation of body temperature readings, crucial for reliable COVID-19 prevention.

Key Results

  • Highest Accuracy: The Random Forest (RF) algorithm achieved the highest recognition accuracy of 92% with 50% data stacking and temperature data included. Extra Trees (ET) followed closely with 90% accuracy.
  • Baseline Accuracy: Without temperature data, the best-performing algorithms (RF and ET) achieved only around 78% recognition accuracy.
  • Temperature Impact: Adding temperature data improved accuracy by 2-11% for most classifiers.
  • Dynamic Activity Temperature Drop: Body surface temperature during walking activities was, on average, 1.15°C to 1.17°C lower than when sitting.
  • Algorithm Efficiency: RF demonstrated a good balance between accuracy and running time, with a 50% stacked model taking 2.38 seconds to process.

Limitations and Future Work

  • Limited Participant Pool: The study involved only 10 participants, which may limit the generalizability of the findings to a broader population. Future work could involve larger and more diverse participant groups.
  • Activity Specificity: While four activities were tested, future research could explore a wider range of daily activities and more complex scenarios.
  • Environmental Factors: The paper acknowledges that extreme weather conditions can influence temperature readings. Further investigation into mitigating these environmental effects would be beneficial.
  • Real-World Deployment: The authors suggest that the proposed statistical analysis of temperature deviations based on activity can provide a valuable reference for epidemic prevention personnel. Future work could focus on the practical implementation and validation of this system in real-world settings.

Summary updated

BibTeX
@misc{zhang2021body,
  author = {Lei Zhang and Yanjin Zhu and Mingliang Jiang and Yuchen Wu and Kailian Deng and Qin Ni},
  title = {{Body Temperature Monitoring for Regular COVID-19 Prevention Based on Human Daily Activity Recognition}},
  year = {2021},
  doi = {10.3390/s21227540},
}
Figure