An integrated multi-layer hybrid framework for accurate, real-time posture assessment in healthcare and rehabilitation contexts is proposed, although all solutions trade off accuracy, computational cost, and practical generalizability.
Abstract
Purpose: This study aims to critically review hybrid yoga posture recognition systems in order to propose an integrated multi-layer hybrid framework for accurate, real-time posture assessment in healthcare and rehabilitation contexts.
Design / Methodology / Approach: Yoga posture recognition methods were taxonomically analysed using a four-dimensional framework, including input modality, feature representation, learning paradigms (SVM, CNN, LSTM), and system-level integration, with multi-metric performance evaluation.
Research Limitation: The proposed hybrid framework is conceptually validated but has not yet been empirically tested in a real-world clinical cohort, representing a direction for future experimental work.
Findings: Hybrid deep learning systems have the best accuracy (94%- 97%), although all solutions trade off accuracy, computational cost, and practical generalizability.
Practical Implication: The hybrid CNN-LSTM system developed will support real-time posture monitoring and provide corrective feedback for remote rehabilitation, fitness coaching, and edge-deployable healthcare.
Social Implication: These applications help reduce healthcare inequalities, lower treatment costs, and improve quality of life for diverse populations worldwide.
Originality/Value: This integrated architecture is evaluated using a comprehensive multi-metric assessment protocol for recognising yoga postures.
This study develops a skeleton-imposed images-based CNN model, contrasting traditional CNN approaches that typically rely solely on body shape or key point representations, that enhances the accuracy of yoga pose classification, providing a more nuanced understanding of complex poses.
Miral Desai, H. Mewada· SN Computer Science· 0 citations
In response to the limitations of traditional posture assessment (subjective, low quantification) and single sensing modalities (vision, pressure, IMU), this paper proposes a multimodal fusion framework integrating plantar pressure and kinematic sensing. Using the MovePort public dataset and a clinical dataset, sliding window augmentation generated 1,216 labeled samples. Five statistical features were extracted from COP and IMU signals. A random forest classifier was adopted with GroupKFold (public dataset) and LOOCV (clinical dataset). Unlike previous work, this study has innovated in the following areas. Firstly, the modal level ablation experiment quantified the fusion gain: Full fusion achieved an accuracy of 85.32% on the MovePort public dataset, outperforming only COP (70.29%) and only IMU (76.83%), with a quantization gain that was+8.49 percentage points higher than the optimal single modality. Secondly, feature importance analysis revealed for the first time the relative contribution of each modality: IMU accounted for 72.6%, COP accounted for 27.4%, providing quantitative basis for sensor selection. Thirdly, strictly adopt topic independent validation (GroupKFold) to avoid overly optimistic generalization estimates commonly seen in random partitioning. On a clinical small sample dataset (N=37, three types of tasks), the framework achieved an accuracy of 56.76%, establishing a baseline for future clinical studies.
Churan Tao, Shanjian Liu, Lin Wang et al.· 2026 IEEE International Conf...· 0 citations
An auxiliary action-recognition evaluation framework incorporating a Big Generative Adversarial Network (BigGAN)-based data augmentation mechanism that offers a reproducible foundation for data augmentation, action classification, and intelligent feedback in sports motion monitoring applications is developed.
Fangge Zhang, Tianli Hao, Longyu He· Journal of Mechanics in Medi...· 0 citations
Human Activity Recognition (HAR) is a fast-growing research area that focuses on identifying human actions using data collected from sensors and vision-based devices. It plays an important role in applications like health monitoring, smart homes, surveillance, sports analysis, and human-computer interaction. In recent years, several methods have been developed to improve the performance of HAR systems using machine learning, deep learning, and hybrid models. This paper presents a detailed review of different methods used in HAR. The study is divided into three main categories: vision-based methods, sensor-based methods, and hybrid approaches that combine both types. Each method is discussed with examples from recent research, along with their advantages and limitations. A comparison is also provided in the form of a table to highlight the performance and challenges of each approach. Although HAR systems have achieved good results in controlled environments, several challenges still remain. These include poor generalization to new users or unknown environments, difficulty in recognizing complex or overlapping activities, dependence on large datasets, and lack of real-time performance. This paper also discusses these research gaps based on recent findings. The future of HAR depends on building more accurate, reliable, and real-time systems that can adapt to different situations. The paper concludes by suggesting possible directions for future work, such as the development of lightweight models, use of standard datasets, better handling of real-time data, and making models more interpretable.
The recognition of emotions is a crucial research field in the domain of intelligent healthcare, as emotional states are related to diagnosis, adherence to therapy, patient safety, mental health, pain perception, and quality of care. In the clinical setting, patients present affect not just by way of words, but also through their facial expression, vocal quality, body movements and reactions. In fact, traditional emotion recognition algorithms—especially unimodal ones—do not capture the complexity of these emotions, as they rely on the information provided from a single source and are therefore sensitive to noise, occlusion, missing information and individual differences. To tackle this drawback, multimodal emotion recognition combines the complementary information from multiple modalities, which increases the robustness and predictive reliability. Meanwhile, deep-learning breakthroughs have ushered in two hugely popular architectural families for this use case: Convolutional Networks, which are adept at local feature extraction, and Transformers, which are very good at the long-range dependency and contextual relationships. This article is the first part of a comparative research paper of a hybrid CNN-Transformer approach to multimodal emotion recognition in healthcare. This study investigates the integration of CNN and Transformer components for enhancing emotional understanding in complex medical contexts like telemedicine, mental health surveillance, pain analysis, rehabilitation, and elderly care. A comparative approach is taken, where hybrid architectures are compared to traditional machine learning approaches, as well as to standalone CNN-based and standalone Transformer-based systems. The main idea is that a combination of local analysis and global reasoning is required for emotion recognition within the healthcare domain, and that such a combination is especially appropriate for the application of emotion recognition in healthcare. The article first presents why the problem of emotion-aware Artificial Intelligence (AI) is relevant in healthcare, and then it summarizes the concepts behind multimodal affective computing. It then elaborates the pros and cons of each modeling paradigm, and suggests a hybrid paradigm that leverages CNNs to extract features from each modality and Transformer layers to fuse features across modalities. The expected advantages of the hybrid model are accuracy, robustness, temporal sensitivity and adaptability to heterogeneous healthcare data sources as compared to the other models, that are detailed in a comparative discussion. Some of the main challenges, such as the availability of annotated clinical data, computational requirements, interpretability, privacy, fairness, and deployment feasibility in real-world healthcare systems, are also discussed. It is concluded that hybrid CNN-Transformer approaches are a promising and practically relevant approach for emotion recognition in healthcare in multimodal settings. Beyond their technical performance, their significance is in being able to effectively support emotionally aware, context-sensitive and patient-oriented intelligent systems. But, to make it successful, it is imperative to have evaluation based on clinical evidence, explainable outputs, and ethical design practices. This paper is a comparative study which serves as a baseline for further studies and development on the reliable, transparent and deployable emotion recognition systems for modern healthcare environments. Multimodal emotion recognition is an essential task in health care, playing a vital role in patient wellbeing. Emotion recognition is an important task in healthcare and has a critical impact on patient well-being, especially in multimodal recognition.
Ancy T A, S.P Swornaibiga· International journal of com...· 0 citations
Pain is a complex and pervasive phenomenon affecting a large percentage of the population, and accurate assessment is essential for effective clinical management and intervention. Computational pain recognition systems enable continuous monitoring, support clinical decision-making, and help mitigate pain-related distress and functional decline. This study introduces a unified tokenization framework for heterogeneous 3D modalities in pain recognition that provides a single processing pipeline across behavioral and brain-activity 3D data, without requiring separate architectures for each modality or handcrafted inductive biases. The framework preserves spatial, temporal, and time--frequency structure while mapping diverse inputs into a shared token space. Extensive experiments show that the proposed approach effectively processes facial videos and fNIRS data in both raw-signal and spectrogram-based representations. On the AI4Pain benchmark dataset, the proposed framework achieves state-of-the-art performance while maintaining high computational efficiency and enabling real-time assessment on both GPU and CPU hardware.
Stefanos Gkikas, Christian Arzate Cruz, V. Becchetti et al.· arXiv.org· 6 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
