Table 3.

Taxonomy of machine learning and deep learning technologies for moderate-to-vigorous physical activity (MVPA) detection, categorized by learning paradigm.

Learning paradigm	Key features	Algorithms	Strengths	Limitations	References
Supervised	Require labeled data (activity intensity labels)	RFa, ANNb, SVMc, DTd, XGBooste, HMMf, QDAg, LASSOh, k-NNi, and gradient boosting	High accuracy with sufficient labeled data Interpretable feature importance, Robust to noise and nonlinear patterns	Dependency on large, labeled datasets Overfitting risk Poor generalization to free-living environments	[28,29,33,36-38,50-60, 62,64-67,69-72,74-78,80]
Unsupervised	Work with unlabeled data, focus on clustering or feature learning	k-means, SOMj, and autoencoders	No need for labeled data Identifies hidden patterns in raw signals Reduces dimensionality	Limited direct applicability to MVPA classification Lower accuracy for intensity-specific tasks Interpretability challenges	[19,29,63]
Hybrid	Combine supervised and unsupervised components, integrates multiple architectures	CNN-BiLSTMk, ViT-BiLSTMl, DLENm, and multitask learning frameworks	Capture spatial and temporal dependencies Improve generalizability State-of-the-art performance in free-living settings	High computational complexity Require large datasets Synchronization challenges for multisensor data	[29,31,33,34,68,79]

^aRF: random forest.

^bANN: artificial neural network.

^cSVM: support vector machine.

^dDT: decision tree.

^eXGBoost: extreme gradient boosting.

^fHMM: hidden Markov model.

^gQDA: quadratic discriminant analysis.

^hLASSO: least absolute shrinkage and selection operator.

ⁱk-NN: k-nearest neighbor.

^jSOM: self-organizing maps.

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

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

^mDLEN: deep learning ensemble network.