Table 1. Literature review for existing PHM models.

Human dynamics prediction via motion sensors
State-of-the-art models	Sensors details	Main contributions	Limitations
Jalal, Quaid & Kim (2019)	Acc	An accelerometer-based motion detection methodology is proposed using multi-features and random forest for classification. The system produced features including variance, positive-negative peaks, and signal magnitude features.	Although the model achieved good accuracy, it considered limited static activities such as drink glass, and pour water.
Chen et al. (2020)	Acc Gyro Mag ECG	A pattern-balanced semi-supervised deep model is proposed for imbalanced activity recognition from multimodal sensors. The study focused on multimodal sensors, limited labeled data and class-imbalance issues. Further, it has exploited the independence of multiple sensors based data and to identify salient regions that recognize human activities.	Imbalanced data distribution is a challenge, which authors tried to void. However, the system performance was low when compared to other methods.
Batool, Jalal & Kim (2019)	Acc Gyro	Method to recognize physical activity detection is proposed via features extraction like Mel-frequency cepstral coefficients (MFCCs). Further, particle swarm optimization and support vector machines (SVM) is used for classification.	Limited motion activities are recognized using Motion-Sense dataset, which will not fit over dynamic activities.
Javeed, Jalal & Kim (2021)	IMU MMG EMG	An effective model for healthcare monitoring has been proposed using multiple features, feature reduction, and recognizer engine. A novel multi-layer sequential forward selection technique has been proposed along with bagged random forest for classification.	The system recognized limited exercise-based activities but was unable to attain good accuracy rates.
Jalal et al. (2020)	Acc Gyro Mag	A detailed study on the physical activities detection systems has been presented in this research. Further, a quality of life improving method has been proposed for indoor-outdoor environments. Both statistical and non-statistical features extraction methods have been fused together to recognize multiple physical activity patterns.	Although the model achieved good accuracy, it recognized only static activities including downstairs, upstairs, and walking.
Xia et al. (2021)	Acc Gyro	The research presents twofold contributions towards sensor-based human activity recognition. First, it proposed a skinned multi-person linear model to build a large dataset based on forward kinematics. Second, it presented a novel deep learning model named multiple level domain adaptive learning model to learn the disentangled representation for the multi-sensors-based data.	The system was able to achieve acceptable rates but due to all the activities mixed together, the performance attained was not up-to-the-mark.
Azmat & Jalal (2021)	Acc Gyro GeoMag	The paper proposed a combination of template matching and codebook generation to eliminate the orientation errors and lessen the computational complexity. The overall methodology involves pre-processing, windowing, segmentation, features extraction, and classification techniques.	Method proposed template matching for static and dynamic activities, however, accuracy achieved for dynamic activities was low.
Ayman, Attalah & Shaban (2019)	Acc Gyro Mag	The paper proposed a novel framework for human activity recognition using machine learning based sensors fusion technique. It also utilized random forest, bagged decision tree, and SVM classifiers for the features selection. The proposed framework consists of data collection, segmentation, features extraction, and classification along with features selection methods.	Limited gestures have been predicted using Handy and PAMAP2 datasets, which will not be able to perform acceptable over dynamic activities.
Jalal et al. (2020)	Acc Gyro Mag	A combination of multiple sensors like accelerometer, gyroscope, and magnetometer have been used to recognize physical activities. Multiple types of features including statistical, MFCCs, and Gaussian mixture model have been extracted followed by the classification of multiple activities via decision tree.	Imbalanced data distribution is eluded. However, the system performance was very low when compared to other state-of-the-art methodologies.
Tao et al. (2021)	Acc Gyro Mag	They proposed a novel attention-based approach for human activity recognition. First, they extracted sensor-wise features using convolutional neural networks (CNN). Then, they used attention-based fusion method for learning body locations and generating features representations. Lastly, inter-sensor features extraction has been applied to learn inter-sensor correlations and predict activities.	The model was able to achieve acceptable rates but due to all the activities mixed together, the performance accomplished was not decent enough.
Javeed et al. (2020)	Acc Gyro ECG EMG	Hybrid-features based sustainable physical healthcare patterns recognition (HF-SPHR) has been proposed in this research. The system includes pre-processing, features extraction, features fusion and reduction, codebook generation, and classification using deep belief networks.	Limited motion activities have been detected via selected datasets that is not sufficient to accomplish well over dynamic activities.