Table 2.
Components of wearable technology mechanism (n = 18).
| Author (year) | Data Extraction |
Data Preparation |
Model Comparisons | Best Model (Type) | External Validation | Distinguish COVID-19 from another ILI | ||||
|---|---|---|---|---|---|---|---|---|---|---|
| WT | Data Input | Feature Selection or Extraction | Data Augmentation | Internal Validation | Data Labelling/Data Segmentation | |||||
| Alavi et al. (2021) | Fitbit, Apple Watch | Overnight RHR | NR | NR | NR | NA | RHRAD, CuSum, NightSignal, Isolation Forest | NightSignal (Deterministic Finite State Machine) | NR | No |
| Bogu and Snyder (2021) | Fitbit | RHR | NR | Seven time-series data augmentation techniques | Time series cross-validation | 7 days before and 21 days after symptom onset were considered as infectious | NA | Long Short-Term Memory Networks-based autoencoder (LAAD) (Unsupervised deep learning) | NR | No |
| Cleary et al (2021) | Fitbit, Apple watch | RHR, sleep and steps | NR | NR | NR | 0–7 days after symptom onset as test periods | RHRmetric, SLEEPmetric, STEPmetric, SENSORmetric | SENSORmetric (Statistical Analysis) | NR | No |
| D’Haese et al (2021) | Oura | Overnight HR, HRV, RR, activity, sleep and symptom report | NR | NR | K-fold cross-validation | Label symptoms suspicious of viral-like symptoms | NA | Markov network and Association Rule Mining Algorithm (Supervised deep learning and unsupervised machine learning) | NR | No |
| Gadaleta et al. (2021) | Fitbit, Apple watch | RHR, sleep, activity and symptom report | Feature Extraction | NR | k-fold cross validation | NR | NA | CatBoost (Supervised machine learning) | NR | No |
| Hassantabar et al. (2021) | Empatica E4 | Galvanic skin response, Temperature, inter-beat interval, oxygen saturation and symptom report | NR | Synthetic data generation with the TUTOR framework | Unspecified | NR | Naïve Bayes, Random Forest, Ada Boost, Decision Tree, SVM, k-NN, deep neural network model with grow-and-prune synthesis | Deep neural network model with grow-and-prune synthesis (Supervised deep learning) | NR | No |
| Hirten et al. (2021) | Apple watch | HRV | NR | NR | Bootstrapping | Defined being symptomatic as the first day of a reported symptom | NA | Mixed-effect Cosinor model (Statistical Analysis) | NR | No |
| Liu et al (2021) | Fitbit | HR | NR | NR | leave one subject out cross-validation | NR | CNN, MLPs, LSTM | Contrastive CAE (Unsupervised deep learning) | NR | No |
| CAE, contrastive CAE | ||||||||||
| Lonini et al (2020) | NR | RR intervals, steps, RR and frequency spectrum of cough signals | Feature Selection | NR | leave-one-subject-out nested cross validation | Labelled snapshots as COVID-19 positive and negative | NA | Logistic Regression (Supervised machine learning) | NR | No |
| Miller et al (2020) | WHOOP | Overnight RHR, HRV and RR | NR | NR | Unspecified | Meeting or exceeding threshold was equivalent to classifying healthy or infected days as COVID-19 positive. | NA | Gradient boosted classifier (Supervised machine learning) | NR | No |
| Mishra et al. (2020) | Fitbit | HR and steps | Feature Extraction | NR | NR | Dates of symptom onset and diagnosis to define sick periods | RHR-Diff, HROS-AD, CuSum | CuSum (Statistical Analysis) | NR | No |
| Natarajan et al. (2020) | Fitbit | RR, HR, and HRV | NR | NR | k-fold cross validation | Data from 2nd to 6th day of symptom onset labelled as sick | NA | CNN (Supervised deep learning) | NR | No |
| Nestor et al (2021) | Fitbit | Night-time RR, RHR, HRV and symptom report | Feature Extraction | NR | Time series cross-validation | Days between self-reported symptom onset and self-reported recovery labelled as positive | XGBoost, XGBoost and GRU-D | XGBoost and GRU-D (Supervised machine and deep learning) | Prospective evaluation | Yes |
| Quer et al (2021) | Fitbit, Apple Watch | RHR, sleep, activity and symptom report | NR | NR | Bootstrapping | First date of symptoms to seven days after symptoms considered infectious | RHRMetric, SleepMetric, ActivityMetric, SymptomMetric, SensorMetric, OverallMetric | OverallMetric (Statistical analysis) | NR | No |
| Sarwar and Agu (2021) | Fitbit | RHR and sleep | Feature Selection and Extraction | Synthetic Minority Over-sampling Technique (SMOTE) | k-fold cross-validation | 14 days after the symptom onset was considered as the infectious period | Naïve Bayes, Random Forest, Ada Boost, Logistics Regression, SVM, Gradient Boosting Classifier, LSTM Autoencoder | Gradient Boosting Classifier (Supervised machine learning) | NR | No |
| Skibinska et al (2021) | Fitbit | HR and steps | Feature Extraction | Synthetic data generation with the TUTOR framework | k- fold Stratified Cross-Validation | NR | XGBoost, k-NN, SVM, Logistic Regression Decision Tree Random Forest | k-NN (Supervised machine learning) | NR | Yes |
| Smarr et al. (2020) | Oura | Dermal temperature | NR | NR | NR | “Symptom window” as each individual’s window of reported symptoms | NA | Minimum and maximum temperature threshold (Statistical analysis) | NR | No |
| Zhu et al (2020) | Huami/Amazfit | RHR and sleep | NR | NR | NR | NR | NA | CDNet – CatNN and DenNN (Supervised deep learning) | NR | No |
Note. CAE = conventional convolutional auto-encoder, CNN = conventional neural network, HR = heart rate, HROS-AD = heart rate over steps anomaly detection, HRV = heart rate variability, ILI = influenza-like illness, k-NN = k-nearest neighbour, LSTM =Long Short-Term Memory Networks, MLPs = Multilayer Perceptrons, NA = not applicable, NR = not reported, RHR = resting heart rate, RHRAD = resting heart rate anomaly detection, RHR-Diff = resting heart rate difference, RR = respiratory rate, SVM = support vector machines, WT = wearable technology