CoChat

Deep CNN-LSTM With Self-Attention Model for Human Activity Recognition Using Wearable Sensor

Open access

The DOI matches its Crossref record; 55 of the 61 references it cites were found in a registry or an index; 5 of the 51 sources its claims cite were checked for support.

Abstract

Human Activity Recognition (HAR) systems are devised for continuously observing human behavior - primarily in the fields of environmental compatibility, sports injury detection, senior care, rehabilitation, entertainment, and the surveillance in intelligent home settings. Inertial sensors, e.g., accelerometers, linear acceleration, and gyroscopes are frequently employed for this purpose, which are now compacted into smart devices, e.g., smartphones. Since the use of smartphones is so widespread now-a-days, activity data acquisition for the HAR systems is a pressing need. In this article, we have conducted the smartphone sensor-based raw data collection, namely H-Activity, using an Android-OS-based application for accelerometer, gyroscope, and linear acceleration. Furthermore, a hybrid deep learning model is proposed, coupling convolutional neural network and long-short term memory network (CNN-LSTM), empowered by the self-attention algorithm to enhance the predictive capabilities of the system. In addition to our collected dataset (H-Activity), the model has been evaluated with some benchmark datasets, e.g., MHEALTH, and UCI-HAR to demonstrate the comparative performance of our model. When compared to other models, the proposed model has an accuracy of 99.93% using our collected H-Activity data, and 98.76% and 93.11% using data from MHEALTH and UCI-HAR databases respectively, indicating its efficacy in recognizing human activity recognition.We hope that our developed model could be applicable in the clinical settings and collected data could be useful for further research.

Overview

CoChat Summary

TL;DR

The paper proposes a deep CNN-LSTM model with self-attention for smartphone/wearable-sensor human activity recognition, trained on a new H-Activity dataset and evaluated on MHEALTH and UCI-HAR. It reports 99.93% accuracy on H-Activity, 98.76% on MHEALTH, and 93.11% on UCI-HAR.

Key findings

  • The proposed M4 model achieved 99.93% accuracy on H-Activity (4,888 correctly classified instances), 98.76% on MHEALTH, and 93.11% on UCI-HAR.
  • On H-Activity, recall was 99.92% for sitting/standing, 100% for walking, 99.75% for jogging, and 100% for running; precision was 100%, 99.83%, 100%, and 99.83%, respectively, with weak differentiation between jogging and running.
  • On UCI-HAR, sitting and standing were confused (recall 74.95% and 82.70%, respectively); on MHEALTH, the confusion matrix shows low recall for some classes (e.g., Std 19.90%, Sit 42.79%) despite 98.76% overall accuracy.
  • Compared with prior models, the proposed model outperformed LSTM-CNN (95.56%) and CNN,LSTM (93.80%) on MHEALTH, and on UCI-HAR outperformed CNN-LSTM (92%), CNN (91.98%), Stacked LSTM (93%), and Res-LSTM (91.6%).

Methods

  • Design: hybrid deep CNN-LSTM with self-attention; proposed M4 has 19 layers (CNN, LSTM, attention, dense, dropout, batch normalization, concatenate) with 633,188 trainable parameters, input size 10×9, trained for 150 epochs with batch size 1000, Adam optimizer, categorical cross-entropy, sigmoid/softmax activations.
  • Data: H-Activity was collected via an Android app from 10 volunteers aged 20–30, smartphone in the right trouser pocket, for 4 activities (sitting/standing, walking, jogging, running), at 10 Hz, yielding 48,920 samples with 9 features (accelerometer, gyroscope, linear acceleration on x/y/z); benchmark datasets were MHEALTH (10 volunteers, 12 activities, 50 Hz) and UCI-HAR (30 participants, 6 activities, 50 Hz, 10,299 samples).
  • Preprocessing and analysis: under/oversampling, linear interpolation for missing values, sliding-window segmentation of size 10, one-hot encoding; evaluation used accuracy, precision, recall, F1, and AUC, with 25% of the whole dataset used for testing and four architectures (M1–M4) compared.

Limitations

  • Stated: future work will add more participants and adjust the network structure; the paper also notes that comparison using final statistical measures with other existing approaches was not conducted because datasets and experimental settings differ.
  • Stated/evident: H-Activity is small (10 volunteers, ages 20–30), uses a single sensor placement (right trouser pocket), a low 10 Hz sampling rate, and only four activities; the data were not subjected to noise cancellation, and the Android platform cannot read sensor data at a particular moment, only when a shift is detected.
  • Evident: class imbalance required resampling; confusion remains between jogging/running and sitting/standing, and the MHEALTH confusion matrix shows low recall for some classes despite high overall accuracy.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/jtehm.2022.3177710

Cites

55 of 55 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

FIGURE 1. The schematic diagram of our proposed workflow. Raw data are firstly acquired from sensors. After preprocessing, segments of data are extracted (known as Segmentation) and a classifier is designed. Fine tuning is used to adjust the hyper-parameters. The classifier is then trained and evaluated using those features (known as Classification).
FIGURE 1. The schematic diagram of our proposed workflow. Raw data are firstly acquired from sensors. After preprocessing, segments of data are extracted (known as Segmentation) and a classifier is designed. Fine tuning is used to adjust the hyper-parameters. The classifier is then trained and evaluated using those features (known as Classification).CC BY · PubMed Central
FIGURE 2. Inertial sensor placement in various datasets (a) H-Activity (b) MHEALTH (c) UCI-HAR.
FIGURE 2. Inertial sensor placement in various datasets (a) H-Activity (b) MHEALTH (c) UCI-HAR.CC BY · PubMed Central
FIGURE 3. Representation of (a) user interface of the Android-based data collection application, ‘Sensors Data Collector’ before and after the data collection (b) inertial sensors in the smartphones and the direction of the accelerometer, gyroscope, and linear acceleration.
FIGURE 3. Representation of (a) user interface of the Android-based data collection application, ‘Sensors Data Collector’ before and after the data collection (b) inertial sensors in the smartphones and the direction of the accelerometer, gyroscope, and linear acceleration.CC BY · PubMed Central
FIGURE 4. The proposed Deep CNN-LSTM with self-attention model architecture for human activity recognition.
FIGURE 4. The proposed Deep CNN-LSTM with self-attention model architecture for human activity recognition.CC BY · PubMed Central
FIGURE 5. Graphical representation of training (a) Accuracy, (b) losses, (c) F1-Scores, and (d) AUCs of M1, M2, M3, and M4, respectively.
FIGURE 5. Graphical representation of training (a) Accuracy, (b) losses, (c) F1-Scores, and (d) AUCs of M1, M2, M3, and M4, respectively.CC BY · PubMed Central
FIGURE 6. Graphical representation of validation (a) Accuracy, (b) losses, (c) F1-Scores, and (d) AUCs of M1, M2, M3, and M4, respectively.
FIGURE 6. Graphical representation of validation (a) Accuracy, (b) losses, (c) F1-Scores, and (d) AUCs of M1, M2, M3, and M4, respectively.CC BY · PubMed Central
FIGURE 7. Influence of the optimizer on model performance during training.
FIGURE 7. Influence of the optimizer on model performance during training.CC BY · PubMed Central
FIGURE 8. Performance comparison of different (LSTM, CNN-LSTM with Self-Attention, LSTM-CNN, and Parallel CNN-LSTM) models.
FIGURE 8. Performance comparison of different (LSTM, CNN-LSTM with Self-Attention, LSTM-CNN, and Parallel CNN-LSTM) models.CC BY · PubMed Central

Verification

Claims the paper makes

2 Valid · 13 Not valid · 34 Not checked · 58 Own results and interpretations

29 more claims

Create a free account to see all 107 claims and to verify the 46 references still open.

See all claims — free account
  • It may state two separate claims as one. 37
  • Probably not a substantial claim. 52

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

5 of 8 · the rest aren’t loaded

3 more findings

Create a free account to see every finding of this check.

See all findings — free account

Citations

Cites

Show all 55 references

Research works printed but not linked: 6

  1. Yang, Deep convolutional neural networks on multichannel time series for human activity recognition, 3995
  2. Weiss, The impact of personalization on smartphone-based activity recognition
  3. Anguita, A public domain dataset for human activity recognition using smartphones, 437
  4. 2020, Sensors Data Collector—Apps on Google Play
  5. Krizhevsky, Imagenet classification with deep convolutional neural networks, 25
  6. 10.3390/s16010115 · DOI · Paper page

Cited by

See all 288

Metadata

Authors
Khatun, M. A., Yousuf, M. A., Ahmed, S., Uddin, M. Z., Alyami, S. A., Al-Ashhab, S., Akhdar, H. F., Khan, A., Azad, A., & Moni, M. A.
Affiliations
  • Jahangirnagar University (BD)
  • SINTEF Digital (NO)
  • Imam Mohammad ibn Saud Islamic University (SA)
  • The University of Queensland (AU)
  • Swinburne University of Technology (AU)
Corresponding author
Moni, Mohammad Ali
Venue
IEEE Journal of Translational Engineering in Health and Medicine
Published
2022
Volume
10
Pages
1-16
ISSN
2168-2372 (ISSN-L)
OpenAlex
W4285079306
Cited by
288
References
61
Citations by year
2026: 68 · 2025: 84 · 2024: 75 · 2023: 54 · 2022: 7
FWCI
17.97
Type
journal article
Language
en
Open access
Locations
  • IEEE Journal of Translational Engineering in Health and Medicine (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Charles Sturt University Research Output (CRO) (submittedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • NORA - Norwegian Open Research Archives (submittedVersion, cc-by)
  • Greater South Information System
Funders
Deanship of Scientific Research at Imam Mohammad Ibn Saud Islamic University through Research Group (RG-21-09-20)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Physical Sciences › Engineering › Biomedical Engineering › Gait Recognition and Analysis
  • Health Sciences › Medicine › Public Health, Environmental and Occupational Health › Physical Activity and Health
Keywords
human activity recognition, wearable sensors, smartphone sensors, inertial sensors, accelerometer, gyroscope, linear acceleration, CNN-LSTM, self-attention, mobile health, UCI-HAR, Android application
MeSH terms
Wearable Electronic Devices, Algorithms, Attention, Humans, Human Activities, Neural Networks, Computer

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-07 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Deep CNN-LSTM With Self-Attention Model for Human Activity Recognition Using Wearable Sensor. https://app.cochat.ai/papers/10.1109/jtehm.2022.3177710
The paper
Khatun, M. A., Yousuf, M. A., Ahmed, S., Uddin, M. Z., Alyami, S. A., Al-Ashhab, S., Akhdar, H. F., Khan, A., Azad, A., & Moni, M. A. (2022). Deep CNN-LSTM With Self-Attention Model for Human Activity Recognition Using Wearable Sensor. IEEE Journal of Translational Engineering in Health and Medicine. https://doi.org/10.1109/jtehm.2022.3177710

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.