1. Introduction
Gait assessment is a critical research field in biomechanics and clinical diagnostics, involving the analysis of walking patterns to diagnose, treat, and monitor various conditions. Gait abnormalities, such as toe walking, characterized by the absence or limitation of heel contact with the ground during the walking cycle, can have a serious impact on physical and social well-being. Persistent toe walking is associated with musculoskeletal complications, increased risk of falls, and limitations in daily and recreational activities [
1]. This paper aims to focus on the detection of toe walking commonly observed in both children and adults. It can arise from various causes, including idiopathic factors or underlying neurological and orthopedic conditions, such as Achilles tendon shortening or stiffness in the calf muscles. The development of hardware–software systems for early detection of these abnormalities is critical for effective intervention [
2,
3]. Furthermore, in addition to clinical diagnostics, such systems can find application in rehabilitation through the production of real-time biofeedback to patients to improve therapy outcomes. Meanwhile, other areas of applications may be in the industrial field through integration into wearable devices for workplace ergonomics or the sports domain for monitoring athletic performance or inappropriate postures.
Traditional observational methods are widely used in clinical settings for their simplicity [
4]; however, they suffer from considerable subjectivity and depend heavily on the observer’s experience, whereas motion analysis systems, such as those that exploit optical motion capture technology, provide accurate kinematic data but require ad hoc structured environments and are not portable. These limitations make them inaccessible for routine diagnostics or for use in remote or resource-limited scenarios.
Motion capture systems are widely used for gait analysis, using high-speed, time-of-flight cameras and reflective markers to track movement in three dimensions. These systems provide accurate kinematic data, allowing clinicians to quantify joint angles and stride length and cadence, as well as identify step cycle deviations characteristic of toe walking. However, they require expensive equipment and controlled laboratory environments, which can limit accessibility and usability [
5]. Another technique used relies on the use of force platforms that measure ground reaction forces during walking, providing insight into the abnormal force patterns associated with forefoot loading, balance, and stability. These platforms are very accurate and valuable for kinetic analysis, but they have the limitation of being stationary and expensive, which may limit their routine use [
6].
Wearable sensors, such as accelerometers, gyroscopes, and Inertial Measurement Units (IMUs), have emerged as a portable alternative [
7,
8]. These devices measure spatiotemporal parameters such as stride length and walking speed while detecting plantar pressure distribution to identify the first contact with the toe [
9]. Although they have reduced accuracy and can be more affected by noise than vision and environmental laboratory systems, wearable sensors have the advantage of lower cost and allow continuous monitoring in both indoor and outdoor environments and can enable early detection of an abnormal condition in walking [
10,
11].
Table 1 shows recent works on gait assessment with their accuracy values. As can be seen, they achieve accuracies above 90%, indicating that they are reliable for the identification of toe walking conditions.
Surface Electromyography (sEMG) is a non-invasive technology that measures the electrical activity generated by muscles during movement. Its application in gait analysis has made significant progress in recent years. For instance, a 2023 study proposed a gait cycle-inspired learning strategy for the continuous prediction of knee joint trajectory from sEMG, highlighting the effectiveness of sEMG in predicting motor intentions prior to actual movement execution [
18]. In the rehabilitation field, the integration of sEMG with motion capture systems has improved the understanding of motor coordination and compensation mechanisms. A 2024 study examined the use of motion analysis technologies in telerehabilitation, emphasizing the importance of sEMG in patient assessment and monitoring [
19].
These advancements highlight the growing importance of sEMG in gait analysis, neural strategies, and biomechanical applications, providing more precise and reliable tools for motor assessment and rehabilitation.
Compared to video motion analysis systems, which are highly accurate but expensive, non-portable, and require significant effort for installation, sEMG offers a more practical and cost-effective alternative. Unlike observational techniques that rely on visible movement, sEMG detects muscle activation before it becomes externally observable, enabling earlier identification of gait abnormalities such as toe walking. Additionally, when compared to wearable IMU systems, sEMG provides higher accuracy, faster model learning, and valuable insights into muscle activation patterns, offering critical diagnostic information to identify the underlying causes of the problem. However, sEMG signals are intrinsically complex, characterized by high variability, noise, and dependence on sensor placement. To address these challenges, this study integrates sEMG with Continuous Wavelet Transform (CWT), a method that converts raw temporal signals into two-dimensional scalograms. As reported in [
20,
21], CWT is becoming an important tool for biomedical signal analysis. In particular, it provides significant advantages in the study of EMG signals, allowing researchers to investigate various aspects of neuromuscular function in both physiological and pathological processes within the time–frequency domain. In addition, scalograms facilitate the application of image-based Deep-Learning (DL) models, which excel at complex pattern recognition. Thus, the proposed approach combines the portability of sEMG with the accuracy of Transfer Learning (TL), offering a scalable and practical solution for gait analysis. Additionally, the TL-based software framework reduces the reliance on extensive training datasets, a common bottleneck in machine-learning applications. By leveraging TL, pre-trained models such as InceptionResNetV2 are fine-tuned for the specific task of gait detection, minimizing computational overhead and accelerating model deployment. This adaptability positions the proposed method as a viable solution for real-world applications, ranging from clinical diagnostics to wearable technologies. To the best of our knowledge, there is no relevant work on toe walking detection using sEMG wearable systems suitable for long-term monitoring. The proposed approach allows both an early identification of the problem and the acquisition of additional neuromuscular information useful for a more comprehensive investigation of the disorder in both diagnosis and rehabilitation.
The remainder of this paper is organized as follows.
Section 2 reports details on the used dataset and implemented algorithmic pipeline. The results are included in
Section 3, while the conclusions are reported in
Section 4.
3. Results and Discussion
To evaluate the effectiveness of the proposed approach, several experiments have been performed. The main algorithms were implemented using Python 3.8, with the following libraries: Tensorflow (2.10), pandas (2.0.3), scikit-learn (1.2.1), spkit (0.0.9.6.7). The hardware environment was a Dell™ Precision 7920 Rack workstation with 256 GB RAM, dual Intel Xeon Gold 5218R CPU@2.10 Ghz processors, and three NVIDIA™ RTX A2000 12 GB GPUs.
Results were analysed using the following four different metrics: accuracy (Acc), precision (Pr), recall (Re), and F1-Score, derived as:
where TP (True Positive) denotes the presence of a toe walking detected successfully by the algorithm; FP (False Positive) implies the absence of toe walking but the algorithm detects it; TN (True Negative) means toe walking is missing and the algorithm correctly does not detect it; and finally, FN (False Negative) denotes toe walking but is not detected by the algorithm. Accuracy displays the ratio of all correctly classified samples to all samples, precision refers to the model’s accuracy in finding positive occurrences, recall refers to the model’s performance in successfully matching positive occurrences using all positive occurrences, and the F1-Score impacts true positive occurrences more than precision. The performance of the proposed approach was evaluated using a 10 cross-validation procedure [
44]. Specifically, each pre-trained architecture was trained with 80% of the dataset, while the remaining 20% was used as the test set. Then, to prevent over-fitting, a validation set was generated with 20% of the training set. To reduce the simultaneous occurrence of the same samples in the training and test sets, the procedure was performed 10 times, with different training and test sets.
Table 3 reports the achieved results for each considered TL architecture. The metrics considered and the training time on the previously described hardware are included. From
Table 3, it can be seen that the modified InceptionResNetV2 architecture performed best in terms of average accuracy compared to the others. In fact, after training 150 epochs, it achieved an accuracy of 95.25%, which is higher compared to the other evaluated architectures. In terms of computational efficiency, a comparable training time can be seen for all the proposed architectures, with the exception of modified MobileNetV3, which has a lower training time of about 30–40%, but a much worse accuracy. Considering the two best performing architectures (InceptionResnetV2 and DenseNet121), the difference in training time is about 280 s, which does not represent a significant challenge for actual hardware capabilities.
For the model with the best performance (InceptionResNetV2),
Table 4 also shows the obtained performance for various considered users. For the sake of brevity, we report the considered metrics for six users with the anthropometric characteristics reported in
Table 5. As we can see from the achieved results, the model continues to perform well, maintaining a similar accuracy for all the users whose Gender, Age, Body Height, and Body Mass are quite different from each other, confirming the goodness of the proposed approach.
Figure 8,
Figure 9,
Figure 10,
Figure 11,
Figure 12 and
Figure 13 show the overall models’ performance, displaying model losses and accuracy in training and validation. From these figures, it can be seen that the validation accuracy can hardly be improved and oscillates around a robust value after 150 epochs of training, although the training accuracy constantly improves. Differences between training and validation accuracy and loss can also be noted, probably related to over-fitting problems most visible in the MobileNetV3, VGG16, and ResNet50 models, i.e., the models with the worst performance (90% lower accuracy). However, again from these figures it can be seen that the best model is still the modified InceptionResNetV2, given the smaller difference in accuracy between the training and validation phases, while MobileNetV3 is still the worst model, as can be seen in
Figure 9a, given the large difference in the two phases.
Figure 14 shows the confusion matrices of the obtained average accuracies for each considered model. Again, observing the numbers on the main diagonal, modified InceptionResNetV2 remains the best performing model, followed by DenseNet121, while, once again, MobileNetV2 performs the worst. In addition, the confusion matrices reported for each considered architecture show that, for the best models, normal and toe walking both obtain an accuracy greater than 90% and similar to each other. In particular, InceptionResNetV2 achieves an accuracy of 97% for normal walking. Finally, it is evident that the lowest accuracy is found in the toe walking classification for VGG16 and ResNet50, which is 77%.
Once the best model was identified, it was saved in “.h5” format and, to show the real generalization of the proposed approach, it was tested using the FREEEMG1000 system, as described in
Section 2.1.
Table 6 reports the obtained results for the considered metrics varying the five considered users. It can be seen that, again, the proposed approach appears to be promising, although with a small dataset. In fact, in this case, an average accuracy of about 93.4% is obtained.
Finally, three classical Machine-Learning (ML) classifiers were used to further validate the proposed approach, using both the public dataset and the dataset obtained using the BTS system. In particular, Support Vector Machine (SVM) [
45], Random Forest (RF) [
46], and K-Nearest Neighbor (KNN) [
47] were considered. The obtained results are shown in
Table 7 and
Table 8, respectively. In particular, considering the best model with the proposed approach (InceptionResNetV2) and the one with ML (RF), an increase of about 6% can be seen with the proposed approach using the public dataset and about 8% with the BTS system, which represents a considerable improvement in performance, even though it requires a significantly higher training time. However, the training is only performed once offline, and therefore the higher training time does not affect the total processing time during continuous monitoring. This ensures real-time operation in normal and toe walking classification using the previously trained model, also on embedded platforms such as Odroid N2+ [
48]. This platform has a limited cost (around EUR 200) and is also equipped with a GPU, which certainly makes it usable in real environment contexts. In support of this, some tests have been carried out on Odroid N2+, achieving an average time to obtain the classification in normal or toe walking of approximately 0.114 s by giving, as input to the model, the scalograms acquired by the two muscles in the analyzed window.
4. Conclusions
Persistent toe walking can lead to foot and ankle muscle issues, resulting in instability, pain, and an increased risk of falls. Moreover, such walking can adversely influence quality of life, making it difficult, for example, to participate in sports activities or, even worse, provoking teasing and bullying in children. So, early detection can prevent more serious health problems, such as Achilles tendon shortening, avoiding negative impacts and consequently improving quality of life. Consequently, an accurate, safe, and low-cost detection method for toe walking is very important.
In this study, a novel approach for detecting toe walking using surface electromyography on the lower limb is proposed. sEMG sensors can facilitate early detection of the problem by measuring muscle electrical activity, enabling the identification of signals prior to the corresponding movement associated with muscle activation. To reduce the complexity of the sEMG signal and enhance generalization across different devices, a TL approach is introduced. Scalograms were applied to track the data in the time–frequency domain, attenuating the sEMG signal constraints. Six different pre-trained architectures were used to distinguish toe walking from normal walking, and the best results were obtained with InceptionResNetV2, achieving an average accuracy of about 95.2% on a public dataset. The approach was also validated using a different sEMG device from the one employed in the public dataset. Specifically, participants were asked to perform normal walking and toe walking tasks, generating a small dataset with the new device. The positive results obtained demonstrate the system’s ability to generalize across different devices.
Current sEMG systems, despite their effectiveness, exhibit certain limitations, such as the challenge of precisely positioning probes on targeted muscles and the discomfort that can arise from prolonged use. To address these concerns, an innovative approach for future development could involve the creation of sensorized wearable garments—such as socks, sleeves, or bands—that seamlessly integrate the necessary electronic and software components. These advanced garments would facilitate accurate probe placement while employing hypoallergenic materials designed to minimize skin irritation, thereby enhancing overall user experience. This approach would also reduce the requirement for specialized personnel, significantly improving user convenience and comfort during extended monitoring periods. Moreover, future advancements will likely focus on building an extensive dataset through collaboration with other project partners, enriching the quality and applicability of the research. Given the overfitting observed in some of the analyzed models, such as MobileNetV3, VGG16, and ResNet50, employing strategies like Regularization or Data Augmentation—potentially using Generative Adversarial Networks—will be essential to mitigate this issue. Further exploration of additional pre-trained architectures will be pursued to determine if InceptionResNetV2 remains superior. Additionally, expanding the scope of analysis to include other walking disorders, such as heel walking, and investigating additional muscles will provide a more comprehensive understanding of sEMG applications.