JMIR Publications

ABSTRACT

Background:

Monitoring and understanding affective states are important aspects of healthy functioning and treatment of mood-based disorders. Recent advancements of ubiquitous wearable technologies have increased the reliability of such tools in detecting and accurately estimating mental states (e.g., mood, stress, etc.), offering comprehensive and continuous monitoring of individuals over time.

Objective:

Previous attempts to model an individual’s mental state relied on subjective measurements or the inclusion of only a few modalities (i.e., phone, watch). Thus, the goal of our study was to investigate the capacity to more accurately predict affect through a fully objective approach using multiple commercial devices.

Methods:

Longitudinal physiological data (such as sleep and heart rate), as well as daily assessments of emotions were collected from a sample of college students at the University of California, Irvine using smart wearables and phones (Oura ring and Samsung smart watches) for over a year. The data was z-normalized using mean and variance of the training set. Distributions of selected participants for both positive and negative affect (PA and NA) were observed. PA had a wider distribution, indicating a higher discriminative power. 7 out of the 20 participants provided us with 200 or more days worth of data, and we used their data for training our model. Data Points with middle 20% affect values were removed such that the classification task only predicts if the value is above or below the participant’s own median score. The data was trained on models such as random forests (RF), support vector machine (SVM), multilayer perceptron (MLP) and K-nearest neighbor (KNN). Moreover, because this study was conducted during the COVID-19 pandemic, it investigated the impact of the very first pandemic-induced lockdown in March 2020 by comparing t-values for emotional states in months before and after March.

Results:

Results showed that RF, the best performing model, was able to predict emotion with 16% and 4% higher accuracies respectively, when compared to state-of-the-art methods. The Area Under Curve (AUC) for affect prediction at 0.82, was 8.1% higher than state-of-the-art. Correlation analyses were also conducted between features and PA (or NA). Oura ring features generally had higher correlations to both PA and NA. Finally, March, 2020 gave the highest t-values, indicating a significant impact of the pandemic induced lockdown on students’ emotional states.

Conclusions:

This study was designed to investigate the viability of predicting an individual's emotion only by leveraging objective measurements without a need to intervene users for collecting feedback. According to results, generic machine learning models, with an accuracy of ~81%, can outperform alternative methods, without requiring historical information about an individual or training personalized models. Additionally, sleep and heart rate are some of the best objective predictions of next-day emotion.