Wearable devices for anxiety assessment: a systematic review

Figures

• 

  Fig. 1. PRISMA diagram showing study selection process. The diagram illustrates the study identification, screening, eligibility assessment, and inclusion process. A total of 517 records were retrieved from Scopus, IEEE, and PubMed, of which 14 duplicates were removed. After title and abstract screening, 402 records were excluded. Of 101 full-text reports assessed, 75 were excluded for reasons including stress detection (not anxiety), reviews, absence of automatic anxiety detection, non-anxiety outcomes, non-wearable devices, or studies in children. Finally, 26 studies were included in the review. ECG electrocardiogram, EDA electrodermal activity, EMG electromyography, PPG photoplethysmography, RSP respiratory signal, SKT skin temperature.

• 

  Fig. 2. Study design and data analysis workflow for wearable-based anxiety detection review. The study design and data analysis workflow consisted of four steps: search and selection of studies, data extraction, feature analysis, and performance analysis. Step 1: Search and selection of studies. a shows the wearable signal modalities analyzed: ECG, RSP, EDA, and PPG. b presents the number of reviewed studies using each signal modality, including single- and multi-modality approaches. Some studies contributed to multiple categories; for example, Jain & Kumar44 reported results using ECG alone, EDA alone, and ECG + EDA combined, and were counted in all relevant categories. Step 2: Data extraction. c highlights the elements extracted from each study, including dataset, features, model, validation, and evaluation metrics. Step 3: Feature analysis. d–g summarize the most commonly used features for ECG, RSP, EDA, and PPG, respectively, and their frequencies of use. Step 4: Performance analysis. h displays boxplots illustrating the distribution of accuracy results across different signal modalities, comparing single-modality and multi-modality approaches, while i presents pooled accuracies for…

• 

  Fig. 3. Overview of anxiety detection methodologies across reviewed studies. a Anxiety induction methods: Proportions of methods used to induce anxiety in the studies. b Anxiety measurement methods: Proportions of methods used to measure anxiety across studies. c Models used for anxiety detection: Proportions of machine learning models employed for anxiety detection. d Validation methods: Proportions of strategies employed to validate models. Note: In all panels, n refers to the number of independent studies reporting the use of the respective method. LOTO CV leave-one-trial-out cross-validation, LOSO CV leave-one-subject-out cross-validation, N/R not reported, VR virtual reality, SVM support vector machines, KNN K-Nearest Neighbors, CV cross-validation.

• 

  Fig. 4. Placement of ECG and RSP WDs in this study. a Wearable ECG device. RA is right arm lead, LA is left arm lead and LL is left leg lead. b T-shirt ECG device. c Inductive RSP device. d Piezoelectric RSP device. e Wrist-mounted EDA and PPG device. f Finger-mounted EDA and PPG device. ECG electrocardiogram, EDA electrodermal activity, PPG photoplethysmography, RSP respiratory signal.

• 

  Fig. 5. Overview of features extracted from physiological signals across reviewed studies for anxiety detection. a Unique features by signal: Number of unique features extracted from each signal modality (ECG, RSP, EDA, and PPG). b Most used ECG features: ECG features reported more than once across studies, highlighting HRV features (e.g., Mean HR, pNN50, and SDNN). c Most used RSP features: RSP features reported more than once across studies, focusing on respiratory timing and interval measures (e.g., Ti/Te mean, Mean IBI). d Most used EDA features: EDA features reported more than once across studies, highlighting amplitude and peak-related metrics (e.g., Mean EDA, Mean Peak Amp SCR). e Most used PPG features: PPG features reported more than once across studies, emphasizing heart rate and variability metrics (e.g., Mean HR, SD PPG). Note: In (a), n indicates the number of unique features identified across studies for each signal modality. In (b–e), n indicates the number of independent studies reporting the use of each specific feature. ECG features: Mean HR mean heart rate, pNN50 percentage of successive RR intervals differing by more than 50 ms, SDNN standard deviation of RR…

• 

  Fig. 6. Accuracy results for anxiety detection across studies by signal(s) used. a Accuracy results by signal(s) used: Boxplots illustrate the distribution of reported accuracies across studies, where the centre line indicates the median, the box represents the interquartile range (IQR), and whiskers show the minimum and maximum values. Categories with fewer than three studies are shown as individual points without boxplots. Single-modality results are highlighted in green, while multi-modality results are shown in purple. b Pooled accuracies for anxiety detection: Bar chart displaying pooled accuracy values for each signal modality or combination, calculated as weighted averages based on the number of studies reporting each result. Note: In (a), n indicates the number of independent study results contributing to each boxplot. ECG electrocardiogram, EDA electrodermal activity, PPG photoplethysmography, RSP respiration, SKT skin temperature, EMG electromyography.

• 

  Fig. 7. Limitation levels across included studies and their distribution by category. a Limitation levels across studies: Heatmap summarizing the limitation levels for population representation, sample size, study completion rate, selective reporting, outcome measurement validity, algorithmic transparency, validation robustness, and reproducibility (L1–L8) for each study. Levels are categorized as Low (green), Moderate (orange), Serious (red), or Could Not Be Determined (gray). b Distribution of limitation levels by category: Bar chart illustrating the proportion of studies rated as Low, Moderate, Serious, or Could Not Be Determined for each limitation category (L1–L8), highlighting the prevalence of Serious limitations in population representation (L1) and sample size (L2). Note: L Limitation.