• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Friday, November 20, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Health

Hybrid Vision Transformer BiLSTM Model Achieves Near-Perfect Physical Activity Classification

by
November 9, 2026
in Health, Technology
Reading Time: 6 mins read
0
Hybrid Vision Transformer BiLSTM Model Achieves Near-Perfect Physical Activity Classification

Hybrid Vision Transformer BiLSTM Model Achieves Near-Perfect Physical Activity Classification

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A deep learning architecture combining Vision Transformers with bidirectional long short-term memory networks has set a new benchmark in classifying physical activity intensity in adults, reaching 99.63 percent accuracy from wrist-worn accelerometer signals. Published in BMC Biomedical Engineering by investigators Lin Wang, Zizhang Luo, and Tianle Zhang, the framework translates raw gravitational acceleration into two-dimensional Gramian Angular Field images before extracting paired spatial and temporal features. Across multi-class evaluations, the ViT-BiLSTM model classified behaviors into sedentary time, light physical activity, and moderate-to-vigorous physical activity. By calculating acceleration with the Euclidean Norm Minus One algorithm rather than relying on traditional metabolic equivalent cut-points, the system showed that computer vision principles can overcome longstanding signal-processing constraints. The model demonstrated remarkable stability across diverse temporal windows and movement categories, offering an accurate methodology for continuous behavioral assessment.

Accurate quantification of daily movement is critical for identifying how activity levels influence cardiovascular and metabolic health, yet traditional analytical approaches often suffer from systematic measurement errors. Accelerometer-based tracking has largely replaced self-reported questionnaires in clinical and epidemiological research because continuous objective monitoring avoids recall bias. However, conventional analytical methods have historically depended on rigid counts or metabolic equivalent thresholds, classifying activities below three metabolic equivalents as light, between three and 5.99 as moderate, and six or above as vigorous. Cumulative count thresholds within fixed epochs produce inconsistent cut-points across studies, with moderate-to-vigorous physical activity cut-offs ranging between 1952 and 2020 counts per minute for identical adult cohorts. These variations complicate comparative analyses across populations, while metabolic-equivalent estimation remains vulnerable to individual physiological differences, ambient environmental factors, and device placement, frequently misclassifying transitional movements.

To overcome the limitations of fixed empirical thresholds, researchers have increasingly evaluated machine learning methods, progressing from classical classifiers to modern deep neural networks. Early implementations using support vector machines, k-nearest neighbors, random forests, and hidden semi-Markov models achieved moderate recognition accuracy but relied on hand-crafted features that required labor-intensive selection and often omitted subtle signal dynamics. Feature selection wrappers proved computationally demanding and prone to overfitting, while statistical filters struggled to model complex feature interactions. Convolutional neural networks subsequently demonstrated the ability to learn hierarchical spatial representations directly from raw data without manual engineering. However, previous convolutional models that attained high accuracy required participants to wear up to five accelerometers simultaneously, severely limiting practical deployment in free-living conditions. Furthermore, unidimensional time-series models often miss global topological features, making it difficult to capture comprehensive spatial and temporal dependencies.

To resolve these structural challenges within a single-device paradigm, the researchers leveraged the extensive Capture-24 dataset collected from Axivity AX3 tri-axial wrist trackers in Oxfordshire between 2014 and 2016. The benchmark repository contains continuous 100-hertz recordings from 151 adults aged 18 to 91 years over approximately 24 hours, generating nearly 4,000 total recording hours with over 2,500 hours annotated against validated ground truth. Within the demographic subset of 131 participants, the sample encompassed 74 young adults aged 18 to 39 years including 32 men and 42 women, 42 middle-aged adults aged 40 to 59 years including seven men and 27 women, and 15 older adults aged 60 and above comprising 14 men and nine women. The analysis focused specifically on data captured between 8 AM and 10 PM, reflecting typical waking hours when sedentary, light, and moderate-to-vigorous behaviors predominantly occur.

The analytical workflow began by processing tri-axial acceleration signals into gravitational values using the Euclidean Norm Minus One algorithm. Expressed as the square root of the summed squared acceleration components minus one standard gravitational unit of 9.81 meters per second squared, this calculation isolates dynamic motion without complex frequency filtering. Ground-truth activity intensity labels were defined using validated adult thresholds established by Hildebrand and colleagues: values from zero to 10 gravitational units per second represented sedentary behavior, 10 to 42 gravitational units denoted light physical activity, and values exceeding 42 units signified moderate-to-vigorous activity. Because raw data were recorded at 100 hertz, these defined intensities corresponded to cumulative acceleration over precise numbers of data points within specific time frames, scaling from 100 points in a one-second window to 500, 1000, and 1500 points across five, ten, and fifteen-second windows.

To enable computer vision architectures to process linear acceleration signals, the normalized sequences were transformed into two-dimensional Gramian Angular Field images. After using the Euclidean Norm Minus One method to remove static gravity and eliminate negative values, the resulting signal sequence was normalized into a bounded interval between negative one and one. These values were projected into polar coordinates, where each normalized amplitude was converted to an angle using an inverse cosine function, and the temporal index was mapped to a normalized radius. The Gramian Angular Field matrix was then calculated by computing the cosine of the sum of angles across all paired time points, encoding local variations and global temporal correlations. After rescaling matrix values between zero and one, the authors generated synthetic images with a resolution of 224 by 224 pixels, matching the standard dimensions required by Vision Transformer architectures.

The hybrid ViT-BiLSTM architecture processes these visual representations by combining spatial feature extraction with bidirectional temporal sequence modeling. Input images are partitioned into fixed-size square patches that are flattened and projected into lower-dimensional embeddings via a learnable linear transformation matrix. Trainable positional encodings are added to preserve the spatial ordering of patches across the image grid. The resulting sequence passes through transformer encoder layers containing multi-head self-attention mechanisms and feed-forward networks with layer normalization, capturing global spatial configurations and subtle motion shifts. These extracted spatial feature vectors are subsequently fed into a bidirectional long short-term memory network, where forward and reverse recurrent cells process past and future context across the sequence. The concatenated hidden states from both directions are then passed to a fully connected classification layer equipped with a softmax function, generating final predictive probability distributions across activity categories.

Training and validation routines were conducted on a dedicated workstation equipped with NVIDIA RTX 2080 Ti graphical processing units, using a data split of 60 percent training, 20 percent validation, and 20 percent testing. The network was optimized using the Adam algorithm with a Cross-Entropy loss function, a sequence length of four, a batch size of 16, an initial learning rate of 1e-5, and a weight decay of 0.001 to prevent overfitting. Learning rates were regulated through a step scheduler with a decay factor of 0.8, alongside moment decay parameters of 0.9 and 0.999, supported by mixed-precision training with an automatic gradient scaler. The training span was capped at ten epochs because preliminary evaluations indicated that validation accuracy improved by less than 0.1 percent per epoch after epoch eight, while extending training up to twenty epochs yielded minimal gains and elevated overfitting risk.

Experimental evaluations demonstrated that the gravity-based ViT-BiLSTM framework significantly outperformed conventional machine learning baselines and alternative deep neural architectures. Operating across a 30-second temporal window, the model attained an overall accuracy of 99.63 percent, achieving class-specific accuracies of 99.5 percent for light physical activity, 98.9 percent for moderate-to-vigorous activity, and 99.5 percent for sedentary behavior. When trained using traditional metabolic equivalent labeling, the same model achieved 96.9 percent overall accuracy, confirming the advantage of gravity-based acceleration. By comparison, a CNN-BiLSTM model achieved 92.01 percent accuracy, a standalone Vision Transformer reached 80.36 percent, an independent BiLSTM attained 80.11 percent, and a standard convolutional neural network showed 74.57 percent. The proposed framework also exceeded previous Capture-24 benchmarks, including traditional machine learning algorithms that scored between 80 and 88 percent and the AccNet24 architecture, which achieved 98.6 percent for sedentary, 95.6 percent for light, and 94.7 percent for vigorous activity.

Statistical testing demonstrated exceptional robustness across different temporal windows and activity categories, with an overall accuracy of 98.5 percent plus or minus 1.48 percent across one, five, ten, fifteen, and thirty-second windows. Accuracies reached 98.7 percent for one second, 98.1 percent for five seconds, 98.2 percent for ten seconds, 99.0 percent for fifteen seconds, and 98.65 percent at thirty seconds. A two-way analysis of variance confirmed no statistically significant accuracy differences across physical activity intensities with an F-value of 2.18 and p-value of 0.13, or across temporal windows with an F-value of 0.52 and p-value of 0.72. Receiver operating characteristic curves showed area-under-the-curve values approaching 1.0 for all intensity levels. Sedentary behavior consistently exhibited the highest accuracy at 98.9 percent, while light and moderate-to-vigorous activities each averaged 98.2 percent. Training and validation loss curves declined steadily toward zero, indicating robust convergence across epochs.

While the proposed architecture established exceptional benchmark performance, the authors outlined distinct limitations and essential engineering directions for translating these algorithms into commercial wearable hardware. The investigation focused exclusively on adult participants between 18 and 91 years of age and implemented a single activity threshold standard, highlighting the necessity of future testing across diverse populations, clinical cohorts, and alternative acceleration boundaries. Additionally, the authors noted an operational trade-off across window durations: while one-second and fifteen-second windows achieved superior accuracies, they exhibited higher loss variance during training due to rapid movement fluctuations. Because transformer architectures require substantial memory and computational bandwidth, deploying the model onto low-power edge microcontrollers will require systematic optimization techniques. Future work will investigate model pruning to eliminate redundant network weights, post-training quantization to reduce floating-point precision, and compact lightweight topologies, paving the way for real-time physical activity monitoring directly within consumer wristbands.

Subject of Research: Deep learning classification of adult physical activity intensity using wrist accelerometer signals

Article Title: A novel ViT-BILSTM model for physical activity intensity classification in adults using gravity-based acceleration

Article References: Wang, L., Luo, Z., & Zhang, T. (2025). A novel ViT-BILSTM model for physical activity intensity classification in adults using gravity-based acceleration. BMC Biomedical Engineering, 7(1), Article 2. https://doi.org/10.1186/s42490-025-00088-2

Image Credits: AI Generated

DOI: 10.1186/s42490-025-00088-2

Keywords: Physical Activity, Deep Learning, Vision Transformer, BiLSTM, Accelerometer, Biomedical Engineering, Pattern Recognition, Wearable Technology, novel, ViT-BILSTM, model, physical

News Source: Denise Maddox. (November 9, 2026). Hybrid Vision Transformer BiLSTM Model Achieves Near-Perfect Physical Activity Classification. Scienmag.

Tags: accelerometerBiLSTMBiomedical Engineeringdeep learningmodelnovelpattern recognitionphysicalphysical activityvision transformerViT-BILSTMwearable technology
Share12Tweet7Share2ShareShareShare1

Related Posts

Clinicians Propose Framework to Protect Medicaid Care Amid Projected Federal Cuts

Clinicians Propose Framework to Protect Medicaid Care Amid Projected Federal Cuts

November 20, 2026
Plant-Disjoint Audit Shows High Accuracy in Automated Coconut Disease Recognition Model

Plant-Disjoint Audit Shows High Accuracy in Automated Coconut Disease Recognition Model

November 20, 2026

Caspofungin and Clindamycin-Primaquine Show Similar Mortality in Refractory Pneumocystis Pneumonia

November 20, 2026

BADF-Net Outperforms Eleven Baseline Models in Segmenting Abdominal Ultrasound Structures

November 20, 2026

POPULAR NEWS

  • CMOS Potts Chip Harnesses Multiphase Oscillations for Low-Power Memory and Graph Optimization

    CMOS Potts Chip Harnesses Multiphase Oscillations for Low-Power Memory and Graph Optimization

    29 shares
    Share 12 Tweet 7
  • Optimized Radiolabeling Improves iPSMA Homogeneity and Supports Single Vial Cold Kit Formulation

    29 shares
    Share 12 Tweet 7
  • Cross-Sectional Bahrain Survey Links Weather, Demands, and Gender to Activity Rates

    29 shares
    Share 12 Tweet 7
  • Biologists Propose Teleonome Framework to Assess Animal Welfare and Agency

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Electrosprayed Starch Microparticles Stabilize Blackberry Extract and Protect Bioactive Antioxidants

Clinicians Propose Framework to Protect Medicaid Care Amid Projected Federal Cuts

Plant-Disjoint Audit Shows High Accuracy in Automated Coconut Disease Recognition Model

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.