A Hybrid Posture Detection Framework: Integrating Machine Learning and Deep Neural Networks

Correct posture is key to good muscle and bone health. Posture is how one holds the body when moving, standing, sitting, or even sleeping. A good posture supports healthy living, while faulty posture may lead to shoulder, neck, and back pain. Technology has contributed to a sedentary lifestyle and prolonged working hours in front of the computer, which may decrease muscle strength and affect body posture, making it vital to consciously or with the help of technology check on posture.

Posture detection and monitoring are crucial for remote health monitoring, especially among the elderly and vulnerable, enabling them to live independently. Apart from geriatrics healthcare, posture detection has been widely applied in human-computer interaction, surveillance, environmental awareness, and physical training.

The posture recognition is done using multisensory and LoRa (Long range) technology. LoRa WAN is low-cost technology with the advantage of long-distance transmission. The network of sensors, Artificial Intelligence (AI), and pressure-sensing technologies are used in comfortable wearable clothes, chairs, wheelchairs, smart cushions, and mattresses to study postures.

The researchers proposed a novel hybrid approach that integrates machine learning (ML) and deep learning (DL) to achieve better posture identification and prediction than DL and ML separately. The three significant contributions of the proposed work are:

• Implement a novel convolutional neural network (CNN) and long short-term memory (LSTM) for automatic posture identification.
• A novel hybrid approach based on deep learning (1D-CNN, 2D-CNN, LSTM, BiLSMT) and machine learning (random forest, K-nearest neighbours (KNN), Naive Bayes, decision tree, Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA) and support vector machines (SVM) methods have been developed to identify the posture prediction.
• A benchmarking of the proposed approach with state-of-the-art approaches shows the accuracy and dominance of the proposed method.

Machine learning, deep learning, and hybrid classifier selection are the main focus for improved performance in remote posture detection with maximum accuracy. Six features are used for posture prediction: skew, percentile, square root (SR), standard deviation (SD), mean, and kurtosis.

After feature extraction, the most critical task is to determine the best feature combination for posture prediction in terms of accuracy. The study evaluated six features and a combination of these features using different machine learning models, including SVM, logistic regression, KNN, decision tree, Naïve Bayes, random forest, LDA, and QDA.

Deep learning methods were used because combinations of features for posture prediction are time-consuming. These do not require feature extraction, and they usually perform better than machine learning models. Deep learning methods, including 1D-CNN, 2D-CNN, LSTM, and BiLSTM have been applied to the raw dataset.

The hybrid approach of remote posture detection contains different predictions of machine learning and deep learning to train meta-learning. The method is used to enhance the system's performance. The experimental results show that posture detection using the hybrid approach performs better than deep learning and machine learning algorithms. The comparative results show that the hybrid approach achieved superior performance in evaluation metrics, including accuracy, precision, recall, and f-measure, compared to machine learning and deep learning.