Table 2.

Characteristics of samples for studies using machine learning approaches. All studies used supervised learning, with sample sizes ranging from 40 to 727 (median 326, IQR 66-410).

Study	Population type	Participants, N	Age (years), mean (SD) Sex (female), %	Ground truth assessment, anxiety measure (type)	Wearable device (consumer or research-grade)	Machine learning task	Predictors	Algorithms	Validation method
Coutts et al, 2020 (United Kingdom) Trial 1 [70]	University students	68	Age: 21 (NRa) Female: 64%	DASS-Ab (state) and STAIc (trait and state)	Biobeam band (research)	Binary classification (supervised)	Heart rate variability	Deep neural networks (LSTMd)	Train-test split: training (80%), test (10%), validation (10%)
Coutts et al, 2020 (United Kingdom) Trial 2 [70]	University students	584	Age: NR Female: NR	DASS-A (state) and STAI (trait and state)	Biobeam band (research)	Binary classification (supervised)	Heart rate variability	Deep neural networks (LSTM)	Train-test split: training (80%), test (10%), validation (10%)
Fukuda et al, 2020 (Japan) [71]	Office workers	60	Age: NR Female: NR	DAMSe (state)	Fitbit Charge 3 (Consumer)	Binary classification (supervised)	Sleep actigraphy data (13 features, eg, total sleep time and number of wake)	RFf	Leave-one-person-out cross-validation
He et al (2025) [72]	University students	40	Age: 21.7 (3.11) Female: 52.5%	STAI Y6 (state)	E4 Empatica Hexoskin smart shirt	Binary classification (unsupervised)	Wrist-worn wearable: skin temprature and EDAg Smart shirt: ECGh	SVMi, RF, KNNj, naïve Bayes	Leave-one-participant-out
Lee et al, 2022 (Korea) [73]	Older adults with mild cognitive impairment	352	Age: 72.48 (5.9) Female: 73%	Clinical diagnosis (Trait)	Fitbit Alta HR2 (Consumer)	Binary classification (supervised)	Model 1: 24-hour activity rhythms and sleep patternk Model 2: Model 1+ minimal K-GAIl (5 items)	Logistic regression, SVM, RF, GBMm	10-fold stratified cross-validation
Lee et al, 2024 (Korea) [74]	Older adults with mild cognitive impairment	352	Age: 72.5 (5.9) Female: 73%	Clinical diagnosis (Trait)	Fitbit Alta HR2 (Consumer)	Binary classification (supervised)	Model 1: activity, sleep Model 2: activity + minimal K-GAI	Convolutional neural network, LSTM, residual network	Cross-validation
Saylam et al, 2023 (United States) [75]	Office workers	727	Age: NR Female: NR	Single itemn (state)	Garmin smartwatch (consumer)	Regression	Activity, stress, sleep, heart rate	RF, XGBoosto, LSTM	80% training, 20% testing
Saylam et al, 2024 (United States) [76]	College students	Baseline: ~700 Follow-up: 300	Age: NR Female: NR	BAIp	Fitbit (consumer)	Regression	Activity and sleep	RF, XGBoost, LSTM	Training: 5 semesters Validation: 2 semesters Testing: 2 semesters

^aNR: not reported.

^bDASS-A: Depression, Anxiety, and Stress Scale-Anxiety Subscale.

^cSTAI: State Trait Anxiety Inventory

^dLSTM: Long Short-Term Memory networks.

^eDAMS: Depression and Anxiety Mood Scale.

^fRF: random forest.

^gEDA: electrodermal activity.

^hECG: electrocardiogram.

ⁱSVM: Support Vector Machine.

^jKNN: k-nearest neighbors.

^kModel 1: 24-hour activity rhythms (interdaily stability [the stability of an activity rhythm of a daily pattern], intra-daily variability [the variability of the activity rhythms throughout the day], dominant rest phase onset [the start time of the 5-hour period with the least activity within 24 hours], and sleep patterns (total sleep time, sleep onset latency, wake after sleep onset, and sleep quality).

^lK-GAI: Korean version of Geriatric Anxiety Inventory.

^mGBM: gradient boosting machine.

ⁿSingle-item question “Please select the response that shows how anxious you feel at the moment. Response scale: 1 (not at all anxious), 2 (a little anxious), 3 (moderately anxious), 4 (very anxious), and 5 (extremely anxious).”

^oXGBoost: extreme gradient boosting.

^pBAI: Beck Anxiety Inventory.