Social anxiety is a fear of social situations in which the individual is exposed to possible scrutiny by others [1], and high levels of social anxiety are associated with a low quality of life in various domains [2,3]. Even when not clinically diagnosable (ie, subclinical or threshold social anxiety), it can greatly affect young people’s lives. Fehm et al [4] showed that young adults with social anxiety who do not receive treatment are at risk of developing social anxiety disorder (SAD) and comorbid mental health problems such as depression, both of which cause further adverse life impairments [3,5]. SAD is one of the most common anxiety disorders [6]. One UK study in 2000 [7] revealed that the annual health care cost per person with SAD was £609 (US $834.59), with annual productivity losses and social security benefits adding to £1920 (US $2631.22) per person with SAD, whereas those with SAD and a comorbidity incurred even higher costs. Nevertheless, many individuals do not receive treatment owing to limited availability or lack of awareness of social anxiety among health care professionals [3,4,8]. Some may not even seek treatment owing to a fear of being negatively evaluated by health care professionals [8]. Thus, it is imperative to empower both individuals and health care professionals in early detection of social anxiety before it potentially escalates into SAD and other related problems.

Common methods for assessing social anxiety involve using subjective measures, usually in a clinical setting. Owing to the rising prevalence of social anxiety, however, it is becoming evident that traditional approaches are inadequate and unsustainable for health care services [4,5,9]. In recent years, increasing focus has been given to technological advances that might help in the early detection and subsequent intervention for anxiety-related problems. In terms of social anxiety, objective methods used to assess symptoms include monitoring physiological changes typically caused by anxiety such as an elevated heart rate (HR), increased electrodermal activity (EDA), variation in skin temperature (ST), and trembling [1,10-12].

Nevertheless, despite extensive and promising research into stress and emotion detection based on physiological indices applicable to social anxiety collected from wearable sensors (Table 1), there has been little effort to predict social anxiety particularly using this approach. This might be ascribed to the recent shift in attention toward social anxiety reported by Heimberg and Butler [13], owing to widening of the diagnostic criteria and leading to a rise in those who fit the criteria for social anxiety.

Although not without its problems, detection via wearable sensors has the potential to underpin solutions addressing the growing needs of individuals with social anxiety and complement traditional therapeutic approaches. If subclinical social anxiety could reliably and validly be detected using wearable sensors, initial treatment could subsequently transition to digital self-help solutions to aid social anxiety at earlier stages when treatment is less extensive and costly [7]. Furthermore, self-help solutions may be a more appropriate method of treatment as individuals with social anxiety often feel nervous to seek treatment in clinical settings [8]. Detecting social anxiety using evidence-based objective methods could also complement current therapeutic approaches.

This study aimed to investigate whether social anxiety in young people with subclinical social anxiety can be detected using physiological data (based on HR, ST, and EDA) recorded from an existing multi-sensor wearable. The study aims to explore if models can be trained to differentiate (1) between baseline and socially anxious states, (2) among baseline, anticipation anxiety, and reactive anxiety states, and (3) between social anxiety among individuals with social anxiety of differing severity. This study also aims to explore the predictive capability of the singular modalities.

Young adult participants were recruited using posters around Imperial College London. The initial sample comprised 13 individuals who self-identified as shy or socially fearful. An exclusion criterion was created to ensure that only young adults with subclinical social anxiety were recruited, as described in Figure 1. To assess participants’ social anxiety levels, the self-reported version of the Liebowitz Social Anxiety Scale (LSAS-SR) was initially used (mean 64.33, SD 13.12, range 38-80). In total, 13 individuals attended the experiment. One participant who showed up for the experiment was known to the experimenter and had their data subsequently excluded owing to likely bias. The final study sample thus comprised 12 participants (58% female; mean age, 19.75 years, SD 1.76 years; 67% Asian, 25% White, and 8% Mixed race).

The University Ethics Committee approved all the procedures and measures used in the study. Throughout the procedure, participants were reminded that their participation was voluntary and that they could withdraw their data at any time until used for statistical analysis. The collected data were anonymized and stored in a password-protected folder.

The data were collected using the E4 Empatica research-grade multi-sensor wristband wearable. The device was selected as it simultaneously monitors various types of physiological data at predetermined sampling rates [34]. However, only HR, EDA, and ST data were explored in this study as they could be considered social anxiety markers [26-29]. E4 Empatica has not yet been used in many studies of this nature, although other multi-sensor wrist-worn wearables have demonstrated effectiveness [16,17].

Using the default sampling rates of the E4 [35], EDA was measured in microSiemens (μS) at 4 Hz using stainless steel electrodes positioned on the inner side of the wrist. HR was measured in beats per minute (BPM) at 1 Hz using data derived from a photoplethysmography sensor. ST was measured in °C at 4 Hz using an infrared thermophile [35]. The data were collected throughout the duration of the experiment, an example of which is shown in Figure 2. The full data set and code needed to recreate the classification models and reproduce the results, as well as functions that enable further experimentation, is available in a designated GitHub repository (Multimedia Appendix 1).

The experiments had an approximate duration of 30 minutes. Similar to previous studies [1,11,23], the experiment was split into 3 stages involving relaxation (baseline), task preparation (anticipation anxiety), and performance (reactive anxiety), as responses might differ across these stages [25]. Figure 3 illustrates the experimental stages. The timestamps for the stages were also recorded for labeling purposes. The experimental protocol is listed below.

First, the wearable was attached to the participant’s wrist. The procedure commenced with a 10-minute baseline period. During this time, participants were offered magazines and ocean sounds were played to create a calming effect.

Second, the nature of the task was then announced, and the participant was given 5 minutes to prepare a 3-minute speech on a selected subject from a choice of topics chosen on the basis of their anxiety-inducing potential. These included “Is Brexit good or bad, and why?,” “Intelligence is not enough,” and “The history of Western Europe until the 2000s.”

Third, a “judging” panel comprising experimenter confederates entered the room, and the participant performed the speech while being timed.

Finally, the participant was debriefed, and the wearable was removed.

The HR data were first upsampled to 4 Hz, similar to ST and EDA. A Moving Average Filter (Equation 1) was then applied to the data to remove noise [17] and reduce the risk of model overfitting [36,37].

The investigations were framed as supervised learning tasks owing to their classification nature. Four classification algorithms were explored: SVM, Random Forest, Decision Tree, and KNN. Furthermore, for classification investigation (2), a “One Vs. Rest” strategy was utilized as the investigation involved a multi-class data set.

The trained models were evaluated using 10-fold cross-validation. The method involves dividing the data set into k-folds with 1 fold for testing and the others for training. Confusion matrices were also utilized to calculate the average classification accuracy for each class.

All study descriptives are shown in Table 2. In this sample, women had higher mean levels for all study variables than men (though the differences were nonsignificant, which is likely owing to the small sample). This is uncharacteristic, as women typically have a higher risk to develop anxiety and higher mean levels of social anxiety than men [41]. However, the self-reported LSAS scores were highly correlated with SPSQ scores (r=0.63; P=.05).

For classification investigation (1), the yielded accuracies were between 97.54% and 99.48%, as shown in Table 3. For investigation (2), the accuracies were between 95.18% and 98.10%, as shown in Table 4. Additionally, for investigation (3) the yielded accuracies were between 98.86% and 99.52%, as shown in Table 5. In each classification investigation, Radial SVM outperformed other classifiers (Tables 3-5).

There were common class misclassification patterns among all classifiers. For investigation (1), the models were less able to classify anxious states (Table 3). For investigation (2), reactive anxiety was misclassified the most and often mistaken for anticipation anxiety (Figure 8). Additionally, in investigation (2), the baseline class was most accurately classified, as shown in Table 4. For investigation (3), the models were not as effective at classifying anxiety category 2 (Table 5).

The singular modality results are shown in Table 6 and Figure 9. In classification investigation (1) EDA yielded 80.46% and was shown to have the highest predictive capability. EDA was also shown to have the highest classification accuracy of 70.02% for investigation (2), whereas ST was the most effective modality for investigation (3) with an accuracy of 89.47%. For each classification investigation, HR was observed to be the least effective modality. Furthermore, KNN generally outperformed other classifiers (Table 6).

Despite these promising results, these findings are preliminary. The sample size was small, with the COVID-19 pandemic preventing further data collection. Prior to the COVID-19 pandemic, we intended to collect test data in “real-world” settings to evaluate the models’ ability to detect social anxiety in practice. Instead, the models were evaluated using a subset of data from the experiment. Although this approach is often used in ML studies [16,17], it does not offer a realistic indication of model generalizability. Therefore, given the small sample size, our results need to be interpreted cautiously.

Additionally, classifiers may have been biased toward certain classes owing to the moderately differing class sizes (Multimedia Appendix 2). This may have accounted for the high accuracies but reduced model generalizability [44]. Like other studies of a similar nature (such as affect recognition studies using physiological data [45]), it was difficult to establish the ground truth of the data with respect to the presence and nature of social anxiety. Therefore, labeling was assumed to be aligned with the experimental protocol.

Furthermore, the physiological responses from the individuals could have been influenced by external factors such as caffeine and alcohol consumption [46,47], though this was not mitigated in the current study design. It is also important to note that EDA measurements can be affected by environmental conditions such as humidity and room temperature [27]. Although the experiments took place in the same room, these variables were not monitored and controlled. In sum, all of these limitations remain challenges for future research.

Despite its limitations, this study has extended previous work and applications focusing on supervised machine learning in the field of physiological anxiety detection. This experiment was informed by existing study protocols, such as using an impromptu speech task, which is a cornerstone of experimental work invoking social anxiety [11,22,23]. Additionally, the study utilized the LSAS-SR measure, which is a widely used measure demonstrating good psychometric properties in previous research [24], and the social anxiety self-reports were cross-validated using another well-known indicator of subclinical social anxiety (SPSQ [32]). Overall, our findings also align with those of prior studies indicating that EDA is a “directed and undiluted” representation of the SNS [27]. Although prior work has focused on EDA as an indicator, physiological measurement from the anatomical site of the wrist had not been explored in a social anxiety context.

This study examined whether social anxiety could be detected in young adults using physiological data (HR, ST, and EDA) from wrist-worn sensors. The findings indicate that it is possible to detect social anxiety and its severity using this approach. Future work in this area has the potential to identify novel methods of detecting and monitoring subclinical social anxiety in young adults, which could help counteract development into SAD. As mental health provision is transitioning toward digital interventions, it is crucial that they are evidence-based and can target individuals with subclinical levels of social anxiety. The ability for future interventions to detect social anxiety before it escalates further could have great social and economic benefits for health care, society and those who experience its consequences.