Table 1.
Performance comparison of classical machine learning algorithms for wearable device datasets.
| Author(s)/Database | Types of Diseases/Data |
CML-Algorithms | Application | Evaluation |
|---|---|---|---|---|
| Shao et al. (2018) [20]/ 2017 PhysioNet CinC Challenge (CinC: Computing in Cardiology) |
AF/ECG | DT, AdaBoosted DT ensemble |
Classification (4 classes) |
F1-score: 0.82 |
| Fallet et al. (2019) [21]/ 17 patients (catheter ablation of cardiac arrhythmia) |
AF and Ventricular arrhythmia/PPG, ECG, ACC-signals (ACC signals: three-axis accelerometer signals) | DT | Classification (2 classes) |
ACC: 95.0% SPE: 92.8% SEN: 96.2% |
| Ghiasi et al. (2020) [22]/ Z-Alizadeh Sani CAD dataset: 303 patients |
CAD/Databank: 55 independent parameters |
DT-based CART (classification and regression tree) |
Classification (2 classes) |
ACC: 92.41%, TNR: 77.01%, TPR: 98.61% |
| Tozlu et al. (2021) [23]/ 33 MI patients, 22 CAD patients, 26 normal. |
MI and CAD/ Electronic noses (19 gas sensors) |
SVM | Classification (2 classes) |
ACC: MI: 97.19%, CAD: 81.48% |
| Qureshi et al. (2020) [24]/ ~250 patients, Extracted CVD dataset |
CVD/Physiological signals and clinical data | SVM and DT | Classification (2 categories) |
ACC: 86.72%, SEN: 67.0%, SPE: 89.0% |
| Mei et al. (2018) [25]/ CinC 2017, (MIT-BIH AF) database |
AF/ECG | SVM and Bagging trees |
Classification (2 classes, 3 classes) |
ACC: 92.0%-96.6% (Varies noise levels), 82.0% (3 classes) |
| Iftikhar et al. (2018) [26]/ 23 healthy people, 40 AF, 21 CAD, 21 MI patients |
AF/SCG and GCG (seismo- and gyro- cardiogram-signals) |
RF and SVM | Multiclass model (SR, AF, CAD, STEMI) | ACC: 75.24% F1: 74% (RF) |
| Sengupta et al. (2018) [27]/ 188 subjects |
Abnormal Myocardial Relaxation (AMR)/spECG (spECG: Signal Processed Surface ECG) | RF/Monte Carlo cross-validation | Prediction | AUC: 91%, SEN: 80%, SPE: 84% |
| Sopic et al. (2018) [28]/ Physionet (PTB Diagnostic ECG database) |
MI/ECG | RF | Classification and prediction |
ACC: 83.26%, SEN: 87.95%, SPE: 78.82% |
| Meng et al. (2019) [29]/ Activity tracker data |
SIHD/Tracker data |
HMM | Output health status over time | AUC: 0.79 |
| Akbulut & Akan (2018) [30]/ 30 participants |
CVD/ECG | Decision Forest (DF), Logistic Regression (LR), NNs | Risk assessment | ACC: 96.0% |
| Dunn et al. (2021) [31]/ 54 integrative personal omics profiling (iPOP) participants |
CVD/PPG, wVS HR, Electrodermal activity (EDA), physical activities |
RF and Lasso models, canonical correlation analysis (CCA) | Prediction | wVS (wearable vital sigh) models outperform cVS (clinical vital sigh) models |
| Han et al. (2019) [32]/ 9530 controls, 306 cases |
AF/AF burden signatures |
Convolutional NN (CNN), RF and L1 regularized LR (LASSO) | Prediction of short-term stroke in 30-day window |
AUC: RF: 0.662, Ensemble: 0.634 |
| Hill et al. (2019) [33]/ CPRD (CPRD: UK Clinical Practice Research Datalink) 2,994,837 individuals (3.2% AF) |
AF/ECG | Statistical/Models (NNs, LASSO, RF, SVM and Cox Regression) | Prediction | AUROC: 0.827 SEN: 75% |
| Jabeen et al. (2019) [34]/ UCI repository, 100 cardiac patients |
CVD/Medical records |
SVM, Naïve Bayes (NB), RF, Multilayer Perceptron (MLP) | Classification (8 classes) |
ACC: 98% for Community-based heuristic approach |
| Kantoch E. (2018) [35]/ 5 participants, SPPB (SPPB: Short Physical Performance Battery task) test task |
Sedentary Behavior (CVD risk)/Ambulatory and Daily activities | Linear Discriminant Analysis (LDA), DT, KNN, SVM, NB, Artificial NNs (ANNs) |
Classification (6 activities) |
ACC: 95.00% ± 2.11% |
| Kwan et al. (2021) [36]/ 50 participants |
AF/PPG | XGBoost, RF, SVM and Gradient Boosting DT | Prediction | AF predicted 4 h in advance |
| Li, B. et al. (2019) [37]/ Hypertension patients, 3 datasets (stroke, HF, renal failure) |
CVD/Medical records |
Spark MLlib library (LR, SVM, NB) |
A risk early warning model | LR(HF): AUC: 0.9269, ACC:0.8529, F1: 0.8456 |
| Yang et al. (2018) [38]/ MIT-BIH arrhythmia Database |
Arrhythmia/ECG | PCANet andand L-SVM, Back Propagation (BP)-NN, KNN | Identification (5 types) |
ACC: 97.77% (skewed) 97.08% (noised) |
| Yang et al. (2020) [39]/ 20 AS patients, 20 health persons |
AS/SCG and GCG | DT, RF and ANNs | Classification (2-classes, multi-classes) |
ACC: (2/multi-classes): RF 97.43%/92.99% |
| Yang and Wei, (2020) [40]/ MIT-BIH AF database |
Cardiac Arrhythmias/ECG |
KNN, SVM and NNs | Classification (6 main types) |
Best ACC: 97.70% (KNN) |
| Bumgarner et al. (2018) [41]/ 100 patients |
AF/ECG | Kardia Band (KB) algorithm supported by Physician |
Classification (2 classes) |
SEN: 99%, SPE: 83%, K coefficient: 0.83 |
| Dörr et al. (2019) [42]/ 672 participants |
AF/PPG, iECG | Heartbeats PPG algorithm |
Classification (2 classes) |
ACC: 96.1%, SEN: 93.7%, SPE: 98.2% |
| Fan et al. (2019) [43]/ 112 participants |
AF/Waveform recording from PPG |
PRO AF PPG algorithm |
Classification (2 classes) |
Smart bands: ACC: 97.72%, SEN: 95.36%, SPE: 99.70% |
| Green et al. (2019) [44]/ 19 patients and 64 healthy volunteers |
oHCM (with left ventricular outflow tract obstruction)/PPG |
Multiple-instance ML model | Classification (2 classes) |
SEN: 95%, SPE: 98%, C-statistic: 0.99 |
| Guo et al. (2019) [45]/ 187,912 used smart devices |
AF/PPG | Discrimination rule PPG algorithm | Prediction | Positive predictive value: 91.6% (95% CI: 91.5% to91.8%) |
| Karwath et al. (2021) [46]/ 18,637 patients (LVEF < 50) |
HFrEF/ECG | Hierarchical clustering |
Statistical analysis |
Mean Jaccard score: 0·571 (SD 0·073; p < 0·0001) |
| Khan and Algarni, (2020) [47]/ UCI dataset https://www.kaggle.com/datasets, accessed on 15 April 2020. |
Heart disease/LoMT (LoMT: Internet of Medical Things) Sensor data and medical records | MSSO-ANFIS | Prediction | ACC: 99.45%, PRE: 96.54% |
| Zeng et al. (2020) [48]/ PTB database:290 subjects, in which 148 patients with MI and 52 controls |
MI/ECG | TQWT-VMD- Radial Basis Function (RBF) |
Classification (2 classes) |
ACC:97.98% |
| Perez et al. (2019) [49]/ 419,297 participants |
AF/PPG, ECG patch |
Irregular pulse notification algorithm |
Identification | Positive predictive value: 84% (95% CI, 76 to 92) |
| Shao et al. (2020) [50]/ AFDB-2017, MIT-BIH AF (MITBIH-AFDB) |
AF/ECG patch | CatBoost-based method |
Classification (4 classes) |
F1: 0.92 |
| Spaccarotella et al. (2020) [51]/ 100 participants, 54 STEMI, 27 non-STEMI, 19 normal |
Acute coronary syndromes/ECG |
Cohen κ coefficient and Bland–Altman analysis | Earlier diagnosis | For STEMI: SEN: 93%, SPE: 95% |
| Stehlik et al. (2020) [52]/ 100 subjects |
HF/PPG | Similarity-based | Prediction | SEN: 88%, SPE: 85% |
| Steinhubl et al. (2018) [53]/ 2659 participants |
AF/ECG | Statistical analysis | Assessment | 3.0% difference (immediate vs. delayed monitoring) |
| Samuel et al. (2020) [54]/ UCI repository Cleveland HF disease dataset: 303 patients |
HF/Medical records |
HNCL (HNCL: Hierarchical Neighborhood Component-based-Learning)/adaptive multi-layer networks (AMLN) | Prediction | ACC: 97.8%, SEN: 95.45%, SPE: 100% |