Table 1.

Machine learning research work for healthcare wearables for fall detection, activity recognition, eating monitoring, fitness tracking, and stress detection.

Task	Research work	ML technique(s)	Sensors/signals used	Dataset(s)
Fall detection	[48]	J48 (96.7%), logistic regression (94.9%), MLP (98.2%)	3D accelerometer and gyroscope in smartphone	MobiAct (https://bmi.hmu.gr/the-mobifall-and-mobiact-datasets-2/)
	[49]	KNN (84.1), naive Bayes (61.5%), SVM (68.25%), and ANN (72%)	Accelerometer, gyroscope, and magnetometer	UMAFall dataset (https://figshare.com/articles/dataset/UMA_ADL_FALL_Dataset_zip/4214283)
	[30]	Temporal signal angle measurements	Inertial measurement unit (IMU)	12 features for 7 subjects performing 5 fall types
		(93.3%@200 Hz to 91.8%@10 Hz)		(9 times each with 3 different speeds)
	[50]	KNN and RF	Accelerometer and gyroscope	SisFall dataset [51]
		(99.80% KNN and 96.82% for falling activity recognition)		(For falling and non-falling activities)
	[52]	SVM (97% F1 score and 99.7% recall)	Accelerometer and gyroscope	Public fall detection dataset [27]
Activity recognition	[25]	CNN	Accelerometer and gyroscope	UCI-HAR dataset and study set
		(UCI-HAR dataset: 95.99%, study set: 93.77%)		21 participants and 6 ADLs
	[53]	Locally linear embedding transfer learning	Accelerometer, magnetometer, gyroscope	UCI-HAR dataset
	[26]	Sequence-to-sequence matching network	Tri-axis accelerometer, tri-axis gyroscopes, magnetometer (depending on the dataset)	Postures dataset, mini MobiAct, and UCI-HAR dataset
	[54]	SVM: 90%	sEMG signals of the upper limb by Delsys, accelerometer	6 males and 6 females for 3 motion states of virtual vehicle: left turn, stop, and right turn
	[39]	ATRCNN: 97%	Tri-axis accelerometer, tri-axis gyroscope	6550 pieces of data for 4 activities: walking, sitting down, running, and climbing stairs
Eating monitoring	[34]	Proximity-based active learning	3D accelerometer	A public dataset for performing different activities including eating [34]
	[55]	Random forest (89.6% in the laboratory and 72.2% outside the laboratory)	One IMU and a proximity sensor on ear and one IMU on the upper back and a microphone	Two datasets: 12.5 hrs for 16 participants in semi-controlled setting with 6 labels and 3 hrs for each of 15 participants outside the laboratory with chewing and non-chewing labels
	[37]	DBSCAN clustering	3D accelerometer	A public dataset for performing different activities including eating [34]
	[56]	Random forest and DBSCAN clustering algorithm (average precision of 92.3%)	Inertial sensor on the downside of the lower jaw	A study dataset of 25 participants, 10 in a laboratory setting and 15 in the wild doing different activities including eating a meal of different food types
	[33]	Gradient boosted decision tree (80.27% accuracy)	Gyroscope and accelerometer in Apple Watch	79 features for 16 subjects taking pills
Fitness tracking	[38]	Logistic regression (0.9356), random forest (0.9203), extremely randomized trees (ERT) (0.9177), and SVM (0.9328)—best accuracy reported in different scenarios	2 accelerometers (hip and ankle)	Study set of 39 participants with a total of 55 days in which sport and jogging activities were logged
	[57]	L2-SVM	3-Axis accelerometer and 3-axis gyroscope	114 participants over 146 sessions
Stress detection	[2]	BN, SVM, KNN, J48,	Zephyr BioHarness for ECG	2 participants with 324 instances
		RF and AB learning methods	Shimmer3 GSR for EDA	At rest and exercise sessions
	[24]	Neural network model (92% accuracy for metabolic syndrome patients and 89% for the rest)	ECG, GSR, body temperature, SpO2, glucose level, and blood pressure	312 biosignal records from 30 participants
	[58]	LR (87% accuracy) and SVM (93%)	ECG sensor in a chest strap	HR and RR data for 44 children (26 with ASD and 18 without ASD) while at rest (7 min) and while engaged in stressful tasks (9 min)