Table 2.

Task-specific performance comparison.

Task type and methods/ model			Key features		Performance metrics (number of studies)		References
Classification (n=28)
	RFa (n=13)	Handcrafted features: time/frequency features (eg, mean, SD, percentiles, lag-1 autocorrelation), ensemble of decision trees		Lab (n=7): F1-score 91.9%, accuracy 94.0% Free-living (n=4): F1-score 81.0%, accuracy 87.4% Lab and free-living (n=2): F1-score 88.1%, accuracy 93.8%		[28,36-38,50-53,64,67,72,74,80]
	ANNb (n=7)	Handcrafted features: time/frequency features (eg, spectral entropy, signal power), multilayer perceptron		Lab (n=7): F1-score 88.0%, accuracy 93.1% Free-living (n=1): F1-score 75.4%, accuracy 82.1%		[19,29,58,60,62,65,73]
	SVMc (n=4)	Kernal-based classification on RBFd, advanced cross-correlation metrics (xy, xz, yz)		Lab (n=1): F1-score 75.4%, accuracy 88.4% Free-living (n=3): accuracy 86.5%		[28,50,55]
	DTe (n=4)	Tree-based splits, integrate with ANN outcomes		Lab (n=3): F1-score 86.6%, accuracy 87.8% Free-living (n=1): F1-score 75.4%, accuracy 82.1%		[58,62,64,74]
	Gradient boosting (n=3)	Gradient boosting framework, handling missing data		Lab (n=2): F1-score 91.6%		[37,54,64]
	HMMf (n=3)	Temporal sequence modeling, Viterbi smoothing		Lab (n=1): F1-score 99.8% Free-living (n=2): F1-score 73.5%, accuracy 94.0%		[71,75,76]
	QDAg (n=1)	Quadratic decision boundaries, probabilistic classification		Lab (n=1): F1-score 100%, accuracy 99.9%		[71]
	LASSOh (n=1)	L1 regularization, sparse solutions		Lab (n=1): F1-score 83.6%		[64]
	CNNi (n=1)	Automated feature extraction via convolutional filters on raw signals		Free-living (n=1): F1-score 73.4%, accuracy 96.8%		[68]
Estimation (n=10)
	RF (n=6)	Regression trees, bootstrapped subsets of ActiGraph data		Lab (n=5): F1-score 83.5%, accuracy 86.1% Free-living (n=1): F1-score 80.0%, accuracy 91.4%		[56,57,59,70,77]
	ANN (n=2)	Nonlinear activation functions, raw signal processing		Lab (n=2): F1-score 91.1%, accuracy 85.7%		[61,66]
	SVM (n=1)	Kernal-based regression.		Lab (n=1): F1-score 90.7%, accuracy 88.7%		[70]
	k-NNj (n=2)	Instance-based learning, Euclidean distance metrics		Lab (n=2): F1-score 96.4%, accuracy 95.8%		[37,70]
	XGBoostk (n=1)	Gradient boosting framework, handling missing data		Lab (n=1): F1-score 100%, accuracy 100%		[37]
	Gradient boosting (n=1)	Iterative error correction, additive regression trees		Lab (n=1): F1-score 93.2%, accuracy 92.1%		[70]
Deep learning (n=5)
	Bi-LSTMl (n=3)	Bidirectional temporal modeling, raw signal processing		Free-living (n=2): F1-score 73.6%, accuracy 93.6% Lab and free-living: F1-score 53.3%, accuracy 53.7%		[31,33,79]
	CNN (n=2)	Automated feature extraction via convolutional filters on raw signals		Free-living (n=2): F1-score 71.9%, accuracy 94.4%		[33,68]
	ViTm (n=1)	Self-attention mechanisms for long-range dependencies		Free-living (n=1): F1-score 79.8%, accuracy 95.0%		[33]
	CNN-LSTMn or CNN-BiLSTMo (n=2)	Hybrid architecture, integrate spatial and temporal learning		Lab (n=1): F1-score 82.1% Free-living (n=1): F1-score 91.4%, accuracy 97.7%		[33,34]
	ViT-BiLSTMp (n=1)	Vision Transformer + BiLSTM, gravity-based acceleration analysis.		Free-living (n=1): F1-score 98.4%, accuracy 99.0%		[33]

^aRF: random forest.

^bANN: artificial neural network.

^cSVM: support vector machine.

^dRBF: radial basis function.

^eDT: decision tree.

^fHMM: hidden Markov model.

^gQDA: quadratic discriminant analysis.

^hLASSO: least absolute shrinkage and selection operator.

ⁱCNN: convolutional neural network.

^jk-NN: k-nearest neighbor.

^kXGBoost: extreme gradient boosting.

^lBiLSTM: bidirectional long short-term memory.

^mViT: vision transformer.

ⁿCNN-LSTM: convolutional neural network and bidirectional long short-term memory.

^oCNN-BiLSTM: convolutional neural network and bidirectional long short-term memory.

^pViT-BiLSTM: vision transformer bidirectional long short-term memory.