Table 4.

Artificial intelligence–related features of wearable devices for forecasting.

Author	Machine learning category	Algorithm used for forecasting or predictions	Input	Data validation method	Data set decomposition	Algorithm best performance	Reported best diagnostic performance of model	Ground truth
Hina et al [22]	Classical	Linear regression, FGSVRa, SVRb, and ensemble-boosted trees	PPGc signal	Train or test split	Training: 60% of 200 subjects Validation: 40%	FGSVR	Coefficient of determination (R2): 0.937 mARDd: 7.62% RMSEe: 11.20 mg/dL Clarke error grid: 95%	Med-device
Alfian et al [23]	Modern	LSTMf	Insulin dose, BGg level, meal ingestion, and exercise activity	Train or test split	Two BG data sets: Data set 1: 148 records CGMh data set: single diabetes patient taken: 26,167 records Both data sets used 80% for training and rest for testing	LSTM	Correlation coefficient (r) and RMSE (Data set 1) RMSE: 25.621 mg/dL, r:0.647 (Data set 2) RMSE: 2.285 mg/dL, r: 0.999	Expert and med-device
Alarcon-Paredes et al [24]	Modern	NDi—ANNj	Fingertip images	K-fold cross validation, train or test split	Training: 514 hostograms Train or test (for model selection): 70% of whole data set Validation subset (model validation): 30%	ANN	MAEMk: 10.37 Clarke grid error: 90.32%	Medical
Islam et al [10]	Modern	CNNl	PPG signal, GSRm, and BG	Train or test split	210 data points Training: 204 data points Testing: 6 data points (4 nondiabetic and 2 diabetic)	CNN	Clarke grid error: 80%	Med-device
Kularathne et al [25]	Classical	Linear regression	Age, BMI, current blood glucose level, genetic factors, smoking, HbA1c,Carbs	NAn	NA	Linear regression	MSEo: 0.0150 R2 score: 0.7834 Variance score: 0.7346	Medical and med-device
Joshi et al [26]	Classical and modern	Deep neural network, MPRp	NIRq signals	Train or test split	Training samples: 187 Testing samples: 46	MPR3	mARD: 4.86% AvgE: 4.88% Mean absolute deviation: 9.42% RMSE: 13.57 mg/dL	Medical and med-device
Zhou et al [27]	Classical	RFr	NIR signals, heart rate variability, pulse transfer time, BG level	Train or test split	Training samples: ND Test samples: 168	RF	Clarke grid error: 80.35% Avg RMSE: 1.44 mg/dL	Med-device
Mahmud et al [28]	Modern	CNN	Infrared channels, GSR and temperature signal	Train or test split	Training samples: 15 instance data of 15 subjects Testing: another 25 data Chosen data length: 1024	CNN	Clarke grid error (values not mentioned)	Med-device
Bent et al [29]	Classical	Multiple regression model, RF	Interstitial glucose summary and glucose variability metrics	Train or test split, leave on out	Training: 16 participants Testing: 10 participants	RF	RMSE: 35.7 mg/dL Mean absolute percentage error: 5.1%	Med-device
Lee et al [9]	Modern	CNN	PPG signal	Train or test split	349 of PPG data samples: Training: 279 sets Testing: 70 sets	CNN	Clarke error grid: 84.29%	Med-device
Shrestha et al [30]	Classical	SVMs	NA	NA	NA	SVM	Accuracy: 97%	Med-device
Li et al [31]	Classical	RF	NIR signal	NA	NA	RF	MAE: 17.27% Clarke error grid: 56.52%	Med-device

^aFGSVR: fine Gaussian support vector regression.

^bSVR: support vector regression.

^cPPG: photoplethysmography.

^dmARD: mean absolute relative difference.

^eRMSE: root mean square error.

^fLSTM: long short term memory.

^gBG: blood glucose.

^hCGM: continuous glucose monitoring.

ⁱND: not defined.

^jANN: artificial neural network.

^kMAE: mean absolute error.

^lCNN: convolutional neural network.

^mGSR: galvanic skin response.

ⁿNA: not applicable.

^oMSE: mean square error.

^pMPR: multiple polynomial regression.

^qNIR: near-infrared.

^rRF: random forest.

^sSVM: support vector machine.