Extending Anxiety Detection from Multimodal Wearables in Controlled Conditions to Real-World Environments

Abstract

This study quantitatively evaluated whether and how machine learning (ML) models built by data from controlled conditions can fit real-world conditions. This study focused on feature-based models using wearable technology from real-world data collected from young adults, so as to provide insights into the models' robustness and the specific challenges posed by diverse environmental noise. Feature-based models, particularly XGBoost and Decision Trees, demonstrated considerable resilience, maintaining higher accuracy and reliability across different noise levels. This investigation included an in-depth analysis of transfer learning, highlighting its potential and limitations in adapting models developed from standard datasets, like WESAD, to complex real-world scenarios. Moreover, this study analyzed the distributed feature importance across various physiological signals, such as electrodermal activity (EDA) and electrocardiography (ECG), considering their vulnerability to environmental factors. It was found that integrating multiple physiological data types could significantly enhance model robustness. The results underscored the need for a nuanced understanding of signal contributions to model efficacy, suggesting that feature-based models showed much promise in practical applications.

Keywords: anxiety; machine learning; multimodal; transfer learning; wearable technology.

Figure 1

(Left) Experimental setup for in-lab and real-world data in RADWear and WEAR studies. (Middle) Experimental setup for data collection using wearable sensors, and (Right) pipeline for processing and analysis.

Figure 2

STAI score for one of the participants at the different test conditions experienced. The threshold considered to be indicative of the observation of anxiety is 11 for the STAI Y6. It can be observed that the meditation session reduced the anxiety level, while the cold pressor test (CPT) increased it. For this participant, the Trier Social Stress Test (TSST) seemed to be the highest cause of anxiety.

Figure 3

Flowchart outlining pipeline for processing and analyzing wearable sensor data from raw data collection to machine learning.

Figure 4

The F1 score of transfer learning models are shown with the left column showing F₁ score for models trained at different noise levels at different signal-to-noise ratios (SNRs), shown in the x-axis. The right column shows F₁ score for each of the tested models for all SNRs. F₁ score, as tested on calibration data for WEAR and RADWear in the top row, WEAR in-lab sessions in the middle row, and on RADWear in-the-wild in the bottom row.