Using Wearable Sensors for Sex Classification and Age Estimation from Walking Patterns
Abstract
:1. Introduction
2. Methods and Data
2.1. Methodology
2.2. Data Collection
2.3. Equipment
2.4. Data Preprocessing
2.5. Feature Extraction
- : difference between the ranks of the corresponding values in the two variables.
- n: number of observations.
2.6. Feature Selection
2.7. Features Description
2.7.1. Sex Classification
- (a)
- Avg Step Interval: This feature calculates the average interval period (in milliseconds) between positive peaks of accl-net. An example of this is shown in Figure 7.
- (b)
- Variability in Interval Between Peaks of Vertical Acceleration: This feature computes the standard deviation of the interval period between peaks of accelerometer-y-axis values within a sample. An example of this is shown in Figure 8.
- (c)
- Avg Interval Between Peaks of Lateral Twist Motion: This feature computes the average interval period between positive peaks of gyroscope-y-axis values within a sample. An example of this is shown in Figure 9.
- (d)
- Power Concentration of Forward Acceleration in Beta Band (9–12 Hz): This feature computes the percentage of total power within the beta band (9–12 Hz) of the power spectral density (PSD) domain for accelerometer-z-axis values. An example of beta band in PSD domain of accl-z is shown in Figure 10.
- (e)
- Lower Bound of Bandwidth of Lateral Twisting Motion: This feature computes the frequency value where the cumulative power exceeds 10% of the total power in the PSD domain for the gyroscope-y-axis. You can find an example of this in Figure 11.
2.7.2. Age Estimation
- (a)
- Avg Step Interval: This feature calculates the average interval period (in milliseconds) between positive peaks of accl-net as described previously.
- (b)
- Range of rate of change of Vertical Swing Motion: This feature computes the inter-quartile range of the derivative of gyroscope-z signal with regard to time. An example of jerk of gyro-z is shown in Figure 12.
- (c)
- Deviation of Forward Acceleration: This feature computes the mean absolute deviation (MAD) of the accelerometer-z signal. An example of this is shown in Figure 13.
- (d)
- Range of Forward Tilt Motion: This feature computes the inter-quartile range of the gyroscope-x signal. An example of this is shown in Figure 14.
- (e)
- Power Concentration of Net Acceleration in Delta Band (0–3 Hz): This feature computes the percentage of total power within the delta band (0–3 Hz) of the power spectral density (PSD) domain for accl-net signal. An example of delta band in PSD domain of accl-net is shown in Figure 15.
3. Results
3.1. Features Analysis
3.1.1. Sex Classification
- (a)
- Avg Step Interval: We can see from Figure 16 that the average value of this feature for males is and females is . This is because males generally have longer strides, leading to a greater interval between positive peaks. On the other hand, females usually take smaller steps.
- (b)
- Variability in Interval Between Peaks of Vertical Acceleration: Based on Figure 17, the average value of this feature for males is and that for females is . This feature assesses the variability in the timing of peaks in acceleration along the y-axis. A higher standard deviation for males suggests a more dynamic movement in the y-axis, whereas a lower standard deviation for females implies a more stable movement in the y-axis.
- (c)
- Avg Interval Between Peaks of Lateral Twist Motion: As depicted in Figure 18, the average value of this feature for males is and that females is . A positive peak in the gyroscope Y-axis may suggest a twisting motion of the torso to the right. Males typically have longer and wider strides, resulting in longer intervals between gyroscope Y-axis positive peaks. On the other hand, smaller strides in females may be the cause for shorter intervals between positive peaks of gyro-y.
- (d)
- Power Concentration of Forward Acceleration in Beta Band (9–12 Hz): We can see from Figure 19 that the average value of this feature for males is and that for females is . Most of the power of human walking motion is within the delta band (0–3 Hz), so this feature highlights that higher frequency component is comparatively more dominant in males than females for acceleration along the z-axis. Males normally walk with faster and stronger pace, which creates a more dynamic movement, increasing the percentage of power within the beta band (9–12 Hz).
- (e)
- Lower Bound of Bandwidth of Lateral Twisting Motion: We can see from Figure 20 that the average value of this feature for males is and that for females is . This feature assesses the initial frequency at which significant energy associated with twisting movement around the y-axis begins. A higher lower bound in females suggests a more frequent twisting motion around the y-axis. A longer and wider stride in males may contribute to a twisting motion of less frequency.
3.1.2. Age Estimation
- (a)
- Avg Step Interval: Based on Figure 21, we see that both male and female children from ages 0 to 10 have lower interval period, as children tend to take small steps; the standard deviation of this feature is quite high among the male children. However, for males, the interval period gradually increases until the age of 20; this is because, as they grow older and taller, their strides also become longer. Then, it decreases slightly and stabilizes until the age of 50. Then, as they grow older and their walking pace gradually slows down, the interval between positive peaks also become longer from 51 to 60 years of age. For females, the interval period gradually increases until the age of 15 and then stabilizes until the age of 30; then, it keeps on decreasing up to the age of 80. Pregnancy and other reasons may cause biological changes to females, resulting in their strides becoming smaller.
- (b)
- Range of rate of change of Vertical Swing Motion: This feature assesses the variability in the rate of change of rotational motion around the z-axis, i.e., swinging motion while walking. Based on Figure 22, we see that variability in the jerk is high for both males and females aged from 0 to 10 as children tend to have very unstable walking style and they have more swinging in their walking style. The iqr value decreases gradually with increases in age for both males and females as their walking style becomes more stable.
- (c)
- Deviation of Forward Acceleration: Based on Figure 23, we see that the MAD of accl-z is highest in the group aged 6–10 years due to their volatile walking pace. The deviation decreases gradually with increasing age for both males and females as their walking pace becomes more stable.
- (d)
- Range of Forward Tilt Motion: This feature assesses the variability in rotational speed around the x axis which basically emphasizes the importance of tilting motion in the forward and backward direction while walking. Based on Figure 24, we see that the tilting movement stays significant throughout the growing years, continuing until age 15 for both males and females. However, then it gradually goes down and remains stable until age of 80.
- (e)
- Power Concentration of Net Acceleration in Delta Band (0–3 Hz): This feature assesses the percentage of power within delta band (0–3 Hz) which is the dominant band as the walking frequency of humans is around 1 Hz. Based on Figure 25, we see that the percentage of power is lower for the group aged 6–10 years due to their erratic gait pattern. As their walking pattern stabilizes, the percentage of power increases and remains steady from ages 11 to 30 for both males and females. However, it declines progressively from ages 31 to 80, which is attributed to slower walking paces and weaker strides with advancing age.
3.2. Model Training
3.3. Classifier Models
3.4. Regression Models
3.5. Performance Evaluation of Classifiers
3.6. Performance Evaluation of Regressors
- : Residual Sum of Squares—the sum of squared differences between observed values and predicted values .
- : Total Sum of Squares—the sum of squared differences between observed values and their mean .
3.7. Identifying the Best-Performing Age Group
3.8. Finding the Optimal Position of the Sensor on the Body
4. Discussion
- Can meaningful features that reflect sex and age differences be extracted from inertial gait data?
- Can traditional machine learning models accurately predict sex and estimate age based on those features?
- Which sensor placements and demographic segments yield the most reliable results?
- (a)
- Feature Relevance and Interpretability: While Ahad et al. [25] reported strong performance using deep learning models like temporal convolutional networks and Angle-Embedded Gait Dynamic Images (AE-GDI), they emphasized that such models act as “black boxes” and offer limited interpretability. In contrast, our use of handcrafted features allows clear physiological interpretations that are useful for clinical or real-world deployment.
- (b)
- Model Performance and Generalization Behavior: Our findings demonstrate that traditional machine learning models, when trained on a carefully engineered feature set, can outperform many existing deep-learning-based solutions in both sex classification and age estimation. As highlighted in Table 8, our model achieved a sex prediction accuracy of 94.4% and an age estimation score of 0.83. This surpasses the results reported in previous studies [16,17,19], many of which relied on significantly smaller datasets or less interpretable deep models.For sex classification, the Stacking Ensemble model achieved the highest performance, followed by kNN and SVC. Interestingly, the decision boundary of SVC was the smoothest, suggesting better generalization with lower overfitting risk, whereas the more complex boundary formed by the Stacking Ensemble model likely contributed to its superior accuracy by capturing intricate decision regions—though this came at the potential cost of increased sensitivity to outliers or distribution shifts.In age estimation, the kNN Regressor yielded the best score, followed by SVR and LightGBM. However, we observed a notable performance drop in all models for the elderly age group (61–80 years). This decline is primarily attributed to the class imbalance in the dataset—a known challenge in regression with continuous targets. Our current approach does not yet incorporate reweighting or SMOTER, but we acknowledge that addressing this imbalance is critical for achieving more consistent performance across all age groups.
- (c)
- Model Robustness Across Age Groups: From our research, we have found that male and female children aged 1–10 years have very similar walking patterns. Therefore, they form the most challenging age group for accurate sex prediction, with the highest obtained accuracy being 88%. As they grow into adults, their walking patterns diverge significantly due to physiological differences between males and females. For this reason, the adult age group (16–60) is the easiest for accurate sex prediction with highest accuracy of 97%. However, as individuals enter old age (61–80), the stability of their walking pattern often deteriorates due to physical conditions and health issues. Nevertheless, the differences are significant enough for yielding a modest performance of 91%.In case of age estimation, children (1–10 years) and adults (11–60 years) are the easiest to predict accurately due to their distinct walking patterns. Children’s walking patterns are characterized by short, inconsistent strides and frequent erratic motions, whereas the walking patterns of adults demonstrate more stable and consistent strides. However, the performance significantly declines for the older age group (61–80) because of the wide variations in their walking patterns due to wide-ranging health conditions. Moreover, the imbalance in the dataset, with very few samples representing this older age group, also contributes to the reduced performance. This finding aligns with those of Van Hamme et al. [26], who noted that model accuracy in older populations tends to degrade unless specific balancing strategies (e.g., stratified sampling or segment-wise averaging) are applied.
- (d)
- Optimal Sensor Positioning: Our analysis reveals that the left and right sides of the waist just above the legs are the optimal sensor placements for predicting sex and age. Sensors positioned at these locations are able to collect more direct forces and vibrations during the leg movements of walking and capture finer details of the movements of both the legs and the waist. In contrast, sensors placed at the center of the back, being farther from the primary sources of motion, fail to capture distinct patterns related to leg movements, resulting in poor performance. This conclusion is consistent with those of both Ahad et al. [25] and Van Hamme et al. [26], who found that hip or lower trunk placements capture richer gait signals than chest or ankle sensors.
5. Limitations
- Dataset Diversity: The dataset used (the OU-ISIR Inertial Gait Database) consists of participants from a single ethnicity and geographic region. As a result, the model’s generalizability to populations with different cultural or physical traits is uncertain.
- Sensor Position Constraints: Although we compared performance across three sensor placements (center, right, and left waist), the study focused solely on the waist area. Gait dynamics may differ significantly when sensors are placed on the foot, ankle, or wrist, which were not explored here.
- Age Group Imbalance: The dataset is imbalanced in terms of age distribution; in particular, we had fewer samples from older adults (ages 61–80). This imbalance affects the accuracy of the conclusions that can be drawn for elderly populations.
- Limited Contextual Information: The model does not account for contextual or physical variables such as height, weight, health conditions, or footwear, which can significantly affect gait. Incorporating these factors could improve performance, especially in sex and age estimation.
- Activity Scope: The analysis is restricted to level-ground walking. Real-world walking patterns may vary significantly in different conditions, such as stair climbing, uneven terrain, or running, which were not included in this study.
6. Conclusions
Author Contributions
Funding
Institutional Review Board Statement
Informed Consent Statement
Data Availability Statement
Conflicts of Interest
Abbreviations
AI | artificial intelligence |
OU-ISIR | Institute of Scientific and Industrial Research (ISIR), Osaka University (OU) |
RNN | Recurrent Neural Network |
LSTM | Long Short-term Memory |
PCA | Principal Component Analysis |
SVM | Support Vector Machine |
NCA | Neighborhood Component Analysis |
MEMS | Micro-Electromechanical System |
IMU | Inertial Measurement Unit |
FFT | Fast Fourier Transform |
PSD | power spectral density |
std | standard deviation |
mad | mean absolute deviation |
iqr | inter-quartile range |
accl | accelerometer |
gyro | gyroscope |
RFE | Recursive Feature Elimination |
SVR | Support Vector Regression |
References
- Bora, N.M.; Molke, G.V.; Munot, H.R. Understanding human gait: A survey of traits for biometrics and biomedical applications. In Proceedings of the 2015 International Conference on Energy Systems and Applications, Pune, India, 30 October–1 November 2015; pp. 723–728. [Google Scholar] [CrossRef]
- Vaughan, C.L.; Davis, B.L.; O’Connor, J.C. Dynamics of Human Gait; Human Kinetics Publishers: Champaign, IL, USA, 1992. [Google Scholar]
- Murray, M.P.; Drought, A.B.; Kory, R.C. Walking Patterns of Normal Men. J. Bone Jt. Surg. Am. Vol. 1964, 46, 335–360. [Google Scholar]
- Saboor, A.; Kask, T.; Kuusik, A.; Alam, M.M.; Le Moullec, Y.; Niazi, I.K.; Zoha, A.; Ahmad, R. Latest research trends in gait analysis using wearable sensors and machine learning: A systematic review. IEEE Access 2020, 8, 167830–167864. [Google Scholar]
- Hu, K.; Wang, Z.; Wang, W.; Martens, K.A.E.; Wang, L.; Tan, T.; Lewis, S.J.; Feng, D.D. Graph sequence recurrent neural network for vision-based freezing of gait detection. IEEE Trans. Image Process. 2019, 29, 1890–1901. [Google Scholar]
- Buongiorno, D.; Bortone, I.; Cascarano, G.D.; Trotta, G.F.; Brunetti, A.; Bevilacqua, V. A low-cost vision system based on the analysis of motor features for recognition and severity rating of Parkinson’s Disease. BMC Med. Inform. Decis. Mak. 2019, 19, 243. [Google Scholar]
- Juen, J.; Cheng, Q.; Prieto-Centurion, V.; Krishnan, J.A.; Schatz, B. Health monitors for chronic disease by gait analysis with mobile phones. Telemed. E-Health 2014, 20, 1035–1041. [Google Scholar]
- Kumari, P.; Cooney, N.J.; Kim, T.s.; Minhas, A.S. Gait analysis in Spastic Hemiplegia and Diplegia cerebral palsy using a wearable activity tracking device-a data quality analysis for deep convolutional neural networks. In Proceedings of the 2018 5th Asia-Pacific World Congress on Computer Science and Engineering (APWC on CSE), Nadi, Fiji, 10–12 December 2018; pp. 1–4. [Google Scholar]
- Derawi, M.O.; Nickel, C.; Bours, P.; Busch, C. Unobtrusive user-authentication on mobile phones using biometric gait recognition. In Proceedings of the 2010 Sixth International Conference on Intelligent Information Hiding and Multimedia Signal Processing, Darmstadt, Germany, 15–17 October 2010; pp. 306–311. [Google Scholar]
- Elrefaei, L.A.; Al-Mohammadi, A.M. Machine vision gait-based biometric cryptosystem using a fuzzy commitment scheme. J. King Saud Univ.-Comput. Inf. Sci. 2022, 34, 204–217. [Google Scholar]
- Simonetti, D.; Hendriks, M.; Koopman, B.; Keijsers, N.; Sartori, M. A wearable gait lab powered by sensor-driven digital twins for quantitative biomechanical analysis post-stroke. Wearable Technol. 2024, 5, e13. [Google Scholar] [CrossRef]
- Ciprandi, D.; Zago, M.; Bertozzi, F.; Sforza, C.; Galvani, C. Influence of energy cost and physical fitness on the preferred walking speed and gait variability in elderly women. J. Electromyogr. Kinesiol. 2018, 43, 1–6. [Google Scholar]
- Singh, B.; Vo, H.; Francis, S.L.; Janz, K.F.; Yack, H.J. Changes in Gait over a 30-min Walking Session in Obese Females. Med. Sci. Sport. Exerc. 2017, 49, 555–562. [Google Scholar]
- Shafizadeh, M.; Davids, K. Wearable System Applications in Performance Analysis of RaceRunning Athletes with Disabilities. Sensors 2024, 24, 7923. [Google Scholar] [CrossRef]
- Khan, S.A.; Ahmad, M.; Nazir, M.; Riaz, N. A comparative analysis of gender classification techniques. Int. J. Bio-Sci. Bio-Technol. 2013, 5, 223–244. [Google Scholar]
- Jain, A.; Kanhangad, V. Gender classification in smartphones using gait information. Expert Syst. Appl. 2018, 93, 257–266. [Google Scholar]
- Sabir, A.T.; Maghdid, H.S.; Asaad, S.M.; Ahmed, M.H.; Asaad, A.T. Gait-based gender classification using smartphone accelerometer sensor. In Proceedings of the 2019 5th international conference on frontiers of signal processing (ICFSP), Marseille, France, 18–20 September 2019; pp. 12–20. [Google Scholar]
- Davarci, E.; Anarim, E. Hybrid architecture for gender recognition using smartphone motion sensors. In Proceedings of the 2021 29th European Signal Processing Conference (EUSIPCO), Dublin, Ireland, 23–27 August 2021; pp. 801–805. [Google Scholar]
- Meena, T.; Sarawadekar, K. Gender recognition using in-built inertial sensors of smartphone. In Proceedings of the 2020 IEEE Region 10 Conference (TENCON), Osaka, Japan, 16–19 November 2020; pp. 462–467. [Google Scholar]
- Ngo, T.T.; Makihara, Y.; Nagahara, H.; Mukaigawa, Y.; Yagi, Y. The largest inertial sensor-based gait database and performance evaluation of gait-based personal authentication. Pattern Recognit. 2014, 47, 228–237. [Google Scholar]
- Abdi, H.; Williams, L.J. Principal component analysis. Wiley Interdiscip. Rev. Comput. Stat. 2010, 2, 433–459. [Google Scholar]
- Khabir, K.M.; Siraj, M.S.; Ahmed, M.; Ahmed, M.U. Prediction of gender and age from inertial sensor-based gait dataset. In Proceedings of the 2019 Joint 8th International Conference on Informatics, Electronics & Vision (ICIEV) and 2019 3rd International Conference on Imaging, Vision & Pattern Recognition (icIVPR), Spokane, WA, USA, 30 May–2 June 2019; pp. 371–376. [Google Scholar]
- Pathan, R.K.; Uddin, M.A.; Nahar, N.; Ara, F.; Hossain, M.S.; Andersson, K. Gender classification from inertial sensor-based gait dataset. In Proceedings of the International Conference on Intelligent Computing & Optimization, Hua Hin, Thailand, 8–9 October 2020; Springer: Cham, Switzerland, 2020; pp. 583–596. [Google Scholar]
- Goldberger, J.; Hinton, G.E.; Roweis, S.; Salakhutdinov, R.R. Neighbourhood components analysis. Adv. Neural Inf. Process. Syst. 2004, 17, 513–520. [Google Scholar]
- Ahad, M.A.R.; Ngo, T.T.; Antar, A.D.; Ahmed, M.; Hossain, T.; Muramatsu, D.; Makihara, Y.; Inoue, S.; Yagi, Y. Wearable sensor-based gait analysis for age and gender estimation. Sensors 2020, 20, 2424. [Google Scholar] [CrossRef]
- Van Hamme, T.; Garofalo, G.; Argones Rúa, E.; Preuveneers, D.; Joosen, W. A systematic comparison of age and gender prediction on imu sensor-based gait traces. Sensors 2019, 19, 2945. [Google Scholar] [CrossRef]
- Antar, A.D.; Ahmed, M.; Ishrak, M.S.; Ahad, M.A.R. A comparative approach to classification of locomotion and transportation modes using smartphone sensor data. In Proceedings of the 2018 ACM International Joint Conference and 2018 International Symposium on Pervasive and Ubiquitous Computing and Wearable Computers, Singapore, 8–12 October 2018; pp. 1497–1502. [Google Scholar]
- Mallat, S. A Wavelet Tour of Signal Processing; Academic Press: Cambridge, MA, USA, 1999. [Google Scholar]
- Nussbaumer, H.J.; Nussbaumer, H.J. The Fast Fourier Transform; Springer: Berlin/Heidelberg, Germany, 1982. [Google Scholar]
- Welch, P. The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms. IEEE Trans. Audio Electroacoust. 1967, 15, 70–73. [Google Scholar]
- Spearman, C. The Proof and Measurement of Association Between Two Things; Creative Media Partners, LLC: Sacramento, CA, USA, 1961. [Google Scholar]
- Cohen, I.; Huang, Y.; Chen, J.; Benesty, J.; Benesty, J.; Chen, J.; Huang, Y.; Cohen, I. Pearson correlation coefficient. In Noise Reduction in Speech Processing; Springer: Berlin/Heidelberg, Germany, 2009; pp. 1–4. [Google Scholar]
- Guyon, I.; Weston, J.; Barnhill, S.; Vapnik, V. Gene selection for cancer classification using support vector machines. Mach. Learn. 2002, 46, 389–422. [Google Scholar]
- Hosmer, D.W., Jr.; Lemeshow, S.; Sturdivant, R.X. Applied Logistic Regression; John Wiley & Sons: Hoboken, NJ, USA, 2013. [Google Scholar]
- Cover, T.; Hart, P. Nearest neighbor pattern classification. IEEE Trans. Inf. Theory 1967, 13, 21–27. [Google Scholar]
- Burges, C.J. A tutorial on support vector machines for pattern recognition. Data Min. Knowl. Discov. 1998, 2, 121–167. [Google Scholar]
- Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar]
- Chen, T.; Guestrin, C. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; pp. 785–794. [Google Scholar]
- Ke, G.; Meng, Q.; Finley, T.; Wang, T.; Chen, W.; Ma, W.; Ye, Q.; Liu, T.Y. Lightgbm: A highly efficient gradient boosting decision tree. Adv. Neural Inf. Process. Syst. 2017, 30, 3149–3157. [Google Scholar]
- Wolpert, D.H. Stacked generalization. Neural Netw. 1992, 5, 241–259. [Google Scholar]
- Seber, G.A.; Lee, A.J. Linear Regression Analysis; John Wiley & Sons: Hoboken, NJ, USA, 2012. [Google Scholar]
- Smola, A.J.; Schölkopf, B. A tutorial on support vector regression. Stat. Comput. 2004, 14, 199–222. [Google Scholar]
- Van der Maaten, L.; Hinton, G. Visualizing data using t-SNE. J. Mach. Learn. Res. 2008, 9, 2579–2605. [Google Scholar]
Category | Sex Prediction Feature Set | Age Prediction Feature Set |
---|---|---|
Time | accl-net-positive-peak-interval-mean † | accl-net-positive-peak-interval-mean † |
accl-net-positive-peak-interval-std | accl-x-mean | |
accl-x-skewness | accl-y-mad | |
accl-y-kurtosis | accl-z-mad | |
accl-y-res-std | gyro-x-iqr | |
accl-y-positive-peak-interval-mean | gyro-y-negative-peak-val-mean † | |
accl-y-peak-interval-std | gyro-z-kurtosis | |
accl-z-negative-peak-val-mean | ||
accl-z-negative-peak-interval-std | ||
gyro-x-positive-peak-val-mean | ||
gyro-y-positive-peak-interval-mean | ||
gyro-y-negative-peak-val-mean † | ||
gyro-z-positive-peak-val-mean | ||
gyro-z-negative-peak-interval-std | ||
Frequency | accl-net-skew-freq-psd | accl-net-percentage-power-delta-band-psd |
accl-x-percentage-power-alpha-band-psd † | accl-x-percentage-power-alpha-band-psd † | |
accl-x-skewness-freq † | accl-x-percentage-power-beta-band-psd | |
accl-y-percentage-power-beta-band-psd | accl-x-skewness-freq † | |
accl-z-percentage-power-delta-band-psd | accl-y-percentage-power-alpha-band-psd | |
accl-z-percentage-power-beta-band-psd | accl-z-kurt-freq-psd | |
gyro-y-bandwidth-psd-lower-bound | accl-z-percentage-power-theta-band-psd | |
gyro-z-percentage-power-alpha-band-psd | gyro-y-percentage-power-theta-band-psd | |
gyro-y-percentage-power-gamma-band-psd | ||
gyro-z-percentage-power-theta-band-psd | ||
Jerk | - | gyro-z-jerk-iqr, gyro-z-jerk-skewness |
Correlation | accl-x-accl-y-spearman † | accl-x-accl-y-spearman † |
accl-x-accl-z-spearman † | accl-x-accl-z-spearman † | |
accl-x-gyro-z-spearman † | accl-x-gyro-x-spearman | |
accl-y-accl-z-spearman | accl-x-gyro-y-spearman | |
accl-y-gyro-z-spearman † | accl-x-gyro-z-spearman † | |
accl-net-gyro-x-spearman | accl-y-gyro-x-spearman | |
gyro-x-gyro-y-spearman † | accl-y-gyro-z-spearman † | |
gyro-x-gyro-z-spearman | accl-z-accl-net-spearman | |
accl-z-gyro-x-spearman | ||
gyro-x-gyro-y-spearman † | ||
gyro-y-gyro-z-spearman |
No. | Feature Name | Analytical Name |
---|---|---|
1 | accl-net-positive-peak-interval-mean | Avg Step Interval |
2 | accl-net-positive-peak-interval-std | Step Interval Variability |
3 | accl-x-skewness | Skewness of Lateral Acceleration |
4 | accl-y-kurtosis | Kurtosis of Vertical Acceleration |
5 | accl-y-res-std | Deviation of Vertical Acceleration |
6 | accl-y-positive-peak-interval-mean | Avg Interval Between Peaks in Vertical Acceleration |
7 | accl-y-peak-interval-std | Variability in Interval Between Peaks of Vertical Acceleration |
8 | accl-z-negative-peak-val-mean | Lower Bound of Forward Acceleration |
9 | accl-z-negative-peak-interval-std | Variability in Interval Between Peaks in Forward Deceleration |
10 | gyro-x-positive-peak-val-mean | Upper Bound of Forward Tilt Motion |
11 | gyro-y-positive-peak-interval-mean | Avg Interval Between Peaks of Lateral Twist Motion |
12 | gyro-y-negative-peak-val-mean | Lower Bound of Lateral Twist Motion |
13 | gyro-z-positive-peak-val-mean | Upper Bound of Vertical Swing Motion |
14 | gyro-z-negative-peak-interval-std | Variability in Interval Between Peaks of Vertical Swing Motion |
15 | accl-net-skew-freq-psd | Skewness of frequency of Net Acceleration |
16 | accl-x-percentage-power-alpha-band-psd | Power Concentration of Lateral Acceleration in Alpha Band |
17 | accl-x-skewness-freq | Skewness of frequency of Lateral Motion |
18 | accl-y-percentage-power-beta-band-psd | Power Concentration of Vertical Acceleration in Beta Band |
19 | accl-z-percentage-power-delta-band-psd | Power Concentration of Forward Acceleration in Delta Band |
20 | accl-z-percentage-power-beta-band-psd | Power Concentration of Forward Acceleration in Beta Band |
21 | gyro-y-bandwidth-psd-lower-bound | Lower Bound of Bandwidth of Lateral Twisting Motion |
22 | gyro-z-percentage-power-alpha-band-psd | Power Concentration of Swing Motion in Alpha Band |
23–30 | Multiple Correlation Features | Motion Correlation Metrics—capturing relationships between acceleration and rotation across the forward, lateral, and vertical axes |
No. | Feature Name | Analytical Name |
---|---|---|
1 | accl-net-positive-peak-interval-mean | Avg Step Interval |
2 | accl-x-mean | Avg Forward Acceleration |
3 | accl-y-mad | Deviation of Vertical Acceleration |
4 | accl-z-mad | Deviation of Forward Acceleration |
5 | gyro-x-iqr | Range of Forward Tilt Motion |
6 | gyro-y-negative-peak-val-mean | Lower Bound of Lateral Twist Motion |
7 | gyro-z-kurtosis | Kurtosis of Vertical Swing Motion |
8 | accl-net-percentage-power-delta-band-psd | Power Concentration of Net Acceleration in Delta Band |
9 | accl-x-percentage-power-alpha-band-psd | Power Concentration of Lateral Acceleration in Alpha Band |
10 | accl-x-percentage-power-beta-band-psd | Power Concentration of Lateral Acceleration in Beta Band |
11 | accl-x-skewness-freq | Skewness of frequency of Lateral Acceleration |
12 | accl-y-percentage-power-alpha-band-psd | Power Concentration of Vertical Acceleration in Alpha Band |
13 | accl-z-kurt-freq-psd | Kurtosis of frequency of Forward Acceleration |
14 | accl-z-percentage-power-theta-band-psd | Power Concentration of Forward Acceleration in Theta Band |
15 | gyro-y-percentage-power-theta-band-psd | Power Concentration of Lateral Twist Motion in Theta Band |
16 | gyro-y-percentage-power-gamma-band-psd | Power Concentration of Lateral Twist Motion in Gamma Band |
17 | gyro-z-percentage-power-theta-band-psd | Power Concentration of Vertical Swing Motion in Theta Band |
18 | gyro-z-jerk-iqr | Range of rate of change of Vertical Swing Motion |
19 | gyro-z-jerk-skewness | Skewness of rate of change of Vertical Swing Motion |
20–30 | Multiple Correlation Features | Motion Correlation Metrics—capturing relationships between acceleration and rotation across the forward, lateral, and vertical axes |
Model | Hyperparameters |
---|---|
LR | C = 0.1, solver = ‘lbfgs’ |
SVC | C = 10, kernel = ‘rbf’ |
kNN | n_neighbors = 3, weights = ‘distance’, metric = ‘manhattan’ |
RF | n_estimators = 100, bootstrap = False, criterion = ‘gini’ |
XGBoost | n_estimators = 500, learning_rate = 0.1, subsample = 0.7, max_depth = 10 |
LightGBM | n_estimators = 500, learning_rate = 0.1, subsample = 0.4 |
Stacking Ensemble | passthrough=True |
Base Models | SVC, kNN, RF (same as above) |
Meta Model (LightGBM) | n_estimators = 100, learning_rate = 0.1, subsample = 0.4 |
Model | Hyperparameters |
---|---|
SVR | C = 150, kernel = ‘rbf’, epsilon = 0.001 |
kNN | n_neighbors = 3, metric = ‘manhattan’, weights = ‘distance’ |
RF | n_estimators = 100 |
XGBoost Regressor | n_estimators = 300, learning_rate = 0.1, max_depth = 10, subsample = 0.8 |
LightGBM Regressor | n_estimators = 500, learning_rate = 0.1, max_depth = −1, subsample = 0.5 |
Classifier | Accuracy | AUC-ROC |
---|---|---|
LR | 0.75 | 0.82 |
SVC | 0.88 | 0.95 |
kNN | 0.90 | 0.97 |
RF | 0.83 | 0.91 |
XGBoost | 0.87 | 0.94 |
LightGBM | 0.87 | 0.94 |
Stacking Ensemble | 0.94 | 0.98 |
Regressor Model | MAE | RMSE | R2 Score |
---|---|---|---|
Linear Regression | 11.15 | 198 | 0.34 |
kNN | 3.53 | 51.1 | 0.83 |
SVR | 6.38 | 82 | 0.73 |
RF | 8.74 | 137.8 | 0.54 |
XGBoost | 7.57 | 115.5 | 0.61 |
LightGBM | 7.23 | 104.3 | 0.65 |
Author(s) | Participants | Features | Method/Model | Accuracy |
---|---|---|---|---|
Jain et al. [16] | 109 subjects | Histogram of Oriented Gradients (HOG) | Bootstrap Aggregating Classifier | - 94.44% (fast walking) - 91.07% (normal walking) |
Sabir et al. [17] | 30 males, 30 females | 38 extracted features | RNN-LSTM | 94.11% |
Davarci et al. [18] | 100 participants | Features analyzing accelerometer correlations | Hybrid model | - 85% (sitting/standing) - 83% (walking) |
Meena et al. [19] | 109 subjects | Reduced triaxial acceleration (PCA-based) | Bagging Trees | 96.3% |
Khabir et al. [22] | Not specified | 88 extracted features | SVM (sex) Decision Tree (age) | - 84.76% (sex classification) - = 0.64 (age estimation) |
Pathan et al. [23] | 1556 subjects | 84 features (using NCA) | Bagging Trees | 87.85% (sex prediction) |
This study | 744 participants | 30 features | Sex—stacking ensemble Age—KNN | 94.4% (sex prediction) - = 0.83 (age estimation) |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Ruhan, R.J.; Wahid, T.; Rahman, A.; Leshob, A.; Rab, R. Using Wearable Sensors for Sex Classification and Age Estimation from Walking Patterns. Sensors 2025, 25, 3509. https://doi.org/10.3390/s25113509
Ruhan RJ, Wahid T, Rahman A, Leshob A, Rab R. Using Wearable Sensors for Sex Classification and Age Estimation from Walking Patterns. Sensors. 2025; 25(11):3509. https://doi.org/10.3390/s25113509
Chicago/Turabian StyleRuhan, Rizvan Jawad, Tahsin Wahid, Ashikur Rahman, Abderrahmane Leshob, and Raqeebir Rab. 2025. "Using Wearable Sensors for Sex Classification and Age Estimation from Walking Patterns" Sensors 25, no. 11: 3509. https://doi.org/10.3390/s25113509
APA StyleRuhan, R. J., Wahid, T., Rahman, A., Leshob, A., & Rab, R. (2025). Using Wearable Sensors for Sex Classification and Age Estimation from Walking Patterns. Sensors, 25(11), 3509. https://doi.org/10.3390/s25113509