Table 2.

Machine learning-based systems for fall detection using wearable systems.

Reference	Year	Dataset used	Sensors/dataset used	Sensor placement (if wearable system)	Methodology	Observed performance
[30]	2011	UCI dataset	3-Axes accelerometer, 2-axis gyroscope	Chest, thigh	Comparison of ML algorithms for fall detection using single node and two nodes	Accuracy of classification = 99.8%, with 2 nodes—one on the waist and one on the knee Naïve-Bayes classifier gave best results
[34]	2012	Generated from experiments	Accelerometer	Mobile phone	Comparison of SVM, SMLR, Naive Bayes, decision trees, kNN, and regularized logistic regression for fall detection	Support vector machines and regularized logistic regression were able to identify a fall with 98% accuracy and classify the type of fall (trips, left lateral, slips, right lateral) with 99% accuracy. Naïve-Bayes reported least accuracy
[29]	2014	Generated from experiments	Accelerometer, gyroscope, magnetometer	6 different positions on the body	Comparison of k-NN classifier, LSM, SVM, BDM, DTW, and ANNs algorithms	k-NN classifier and LSM gave above 99% for sensitivity, specificity, and accuracy
[22]	2014	Generated from experiments	Accelerometer	Mobile phone	Accelerometer data from wearable sensors to generate alarms for falls, combined with context recognition using sensors in an apartment, for inferring regular ADLs, using Bayesian networks	Provides statistical information regarding the fall risk probability for a subject
[48]	2015	Publicly available activity recognition dataset	Accelerometer, gyroscope	Smartphone	Comparison of Naive Bayes classifier, decision trees, random forests, classifiers based on ensemble learning (random committee), and lazy learning (IBk) algorithms for activity detection	Naive Bayes classifier performs reasonably well for a large dataset, with 79% accuracy, and it is fastest in terms of building the model taking only.5.76 seconds Random forests are better in terms of both accuracy and model building time, with 96.3% accuracy and 14.65 seconds model building time. k-Means clustering performs poorly with 60% classification accuracy and 582 seconds model building time
[47]	2016	Generated from experiments	3-Axes accelerometer	Not specified	Comparison of decision tree, decision tree ensemble, kNN, neural networks, MLP algorithms for soft fall detection	Decision tree ensemble was able to detect soft falls at more than 0.9 AUC
[31]	2016	MobiFall dataset	Accelerometer, gyroscope	User's trouser pocket	Comparison of Naïve-Bayes, LSM, ANN, SVM, kNN algorithms for fall detection	k-NN, ANN, SVM had the best accuracy—results for kNN: Accuracy = 87.5 Sensitivity = 90.70 Specificity = 83.78
[26]	2016	Generated from experiments	3-Axis accelerometer	Smartwatch	Threshold-based analysis of acceleration	Accuracy = 96.01%
[40]	2017	Generated from experiments	Accelerometer, gyroscope	Vest	Kalman filter for noise reduction, sliding window, and Bayes network classifier for fall detection	With Kalman filter Accuracy = 95.67%, Sensitivity = 99.0% Specificity = 95.0%
[38]	2017	Generated from experiments	3D accelerometer	Smartphone	Combination of threshold-based and ML-based algorithms—K-Star, Naive Bayes, J48	Energy saving = 62% compared with ML-only techniques Sensitivity = 77% (thresholding only), 82% (ML only), 86% (hybrid) Specificity = 99.8% (thresholding only), 98% (ML only), 99.5% (hybrid) Accuracy = 88.4% (thresholding only), 90% (ML only), 92.75% (hybrid)
[46]	2017	Generated from experiments	3-Axes accelerometer	Waist	Combination of threshold-based and knowledge-based approach based on SVM to detect a fall event	Using a knowledge-based algorithm: Sensitivity = 99.79% Specificity = 98.74% Precision = 99.05% Accuracy = 99.33%
[49]	2017	Generated from experiments	3-Axes accelerometer	Smartwatch	Spectrum analysis, combined with GA-SVM, SVM, and C4.5 classifiers	GA-SVM gave best results with Accuracy = 94.1% Sensitivity = 94.6% Specificity = 93.6%
[50]	2017	MobiFall dataset	3-Axes accelerometer	Not specified	Comparison of multilevel fuzzy min-max neural network, MLP, KNN, SVM, PCA for fall detection	Multilevel fuzzy min-max neural network gave best results with Sensitivity = 97.29% Specificity = 98.70%
[37]	2017	FARSEEING dataset	3-Axes accelerometer	5 locations on the upper body - neck, chest, waist, right side, and left side	Sensor orientation calibration algorithm to resolve issues arising out of misplaced sensor locations and misaligned sensor orientations, HMM classifiers	Sensitivity = 99.2% (experimental dataset), 100% (real-world fall dataset)
[11]	2017	Generated from experiments	3-Axes accelerometer	Chest	LWT-based frequency domain analysis and SVM-based time domain analysis of RMS of acceleration	Accuracy = 100% Sensitivity = 100% Specificity = 100%
[32]	2017	Generated from experiments	3-Axis accelerometer, 3-axis gyroscope	Waist	Backpropagation neural network (BPNN) for fall detection	Accuracy = 98.182% Precision = 98.33% Sensitivity = 95.161% Specificity = 99.367%
[39]	2010	Generated from experiments	Accelerometer	Chest, thigh	Naïve-Bayes, SVM, OneR, C4.5 (J48), neural networks	Naïve-Bayes gave best results Accuracy = 100% Sensitivity = 87.5%
[43]	2016	Generated from experiments	Accelerometer	Different parts of the body	Bayesian framework for feature selection, Naïve-Bayes, C4.5	Better accuracy with improved classification than Naïve-Bayes and C4.5
[33]	2016	Generated from experiments	3D accelerometer	Chest	SVM, kNN, complex tree algorithms applied on data generated by accelerometers	Accuracy and precision of SVM were the highest Recall was highest for complex tree
[44]	2017	Generated from experiments	Accelerometer (MobiAct dataset)	Not applicable	ENN + kNN (where ENN was applied to remove outliers), ANN, SVM, and J48	For ENN + kNN: Sensitivity = 95.52% Specificity = 97.07% Precision = 91.83%
[41]	2018	Generated from experiments	Triaxial gyroscope	Waist	Decision tree	Accuracy = 99.52% Precision = 99.3% Recall = 99.5%
[45]	2018	Cogent dataset, SisFall dataset	3D accelerometer, 3D gyroscope-Cogent dataset Accelerometer, gyroscope-SisFall dataset	Chest, waist	Event-ML, classification and regression tree (CART), kNN, logistic regression, SVM	Better precision and F-scores with Event-ML than FOSW and FNSW-based approaches
[42]	2019	Public datasets	Accelerometer, gyroscope	Chest, thigh	ANN, kNN, QSVM, ensemble bagged tree (EBT)	Extraction of new features from acceleration and angular velocity improved the accuracy of all 4 classifiers. Accuracy of EBT was highest (97.7%)
[51]	2019	SisFall dataset	Accelerometer, gyroscope	Waist	kNN, SVM, random forest	Accuracy for fall detection was the highest for kNN (99.8%). Accuracy for recognizing fall activities was the highest for random forest (96.82%)
[52]	2018	SisFall dataset, generated from experiments	Accelerometer	Chest/thigh, waist	SVM, kNN, Naïve-Bayes, decision tree	Accuracy and sensitivity of SVM were the highest (97.6% and 98.3%, respectively) for both datasets
[63]	2018	UMA dataset	Accelerometer, gyroscope, magnetometer	Wrist, waist, chest, ankle	kNN, Naïve-Bayes, SVM, ANN, decision tree	Without risk categorization: 81% for decision tree With risk categorization: 85% for decision tree
[56]	2019	Public datasets	Accelerometer	Not specified	CNN-based models for feature extraction	Highest accuracy reported = 99.86%
[57]	2018	SisFall dataset-original and manually labelled	Accelerometer	Not specified	RNN	Highest accuracy reported for fall detection: 83.68% (before manual labelling), 98.33% (after manual labelling)
[36]	2018	Generated from experiments	Accelerometer, gyroscope, magnetometer	Near the waist	kNN	Accuracy = 99.4%
[16]	2018	Generated from experiments	Accelerometer	Waist	Decision tree	Accuracy = 91.67% Precision = 93.75%
[54]	2018	SisFall dataset	Accelerometer	Waist	RNN with LSTM	Highest accuracy (after hyperparameter optimization) = 97.16%
[53]	2017	Generated from experiments	Accelerometer, gyroscope, proximity sensor, compass	Right, left, and front pockets	SVM, decision tree, kNN, discriminant analysis	Highest accuracy = 99% for SVM
[59]	2018	Generated from experiments	Depth camera, accelerometer	Waist	CNN	Accuracy of fall detection = 100%
[55]	2017	Public datasets	Accelerometer	Not specified	CNN-based analysis on time series accelerometer data converted to images	Accuracy = 92.3%
[58]	2017	Generated from experiments	Accelerometer, radar, depth camera	Wrist	Ensemble subspace discriminant, linear discriminant, kNN, SVM	Overall accuracy of ensemble classifier was the highest, after fusion of radar, accelerometer, and camera = 91.3%. This is an improvement of 11.2% compared to radar-only and 16.9% compared to accelerometer-only results
[62]	2018	Generated from experiments	Accelerometer, gyroscope, magnetometer	Hip	SVM, random forest	Without sensor fusion: Accelerometer precision = 86.23% Accelerometer recall = 87.46% With sensor fusion: precision = 94.78%, recall = 94.37%, with random forest