Vision-based approach to knee osteoarthritis and Parkinson’s disease detection utilizing human gait patterns

Sensor-based gait acquisition for PD detection

Research studies have acquired human gait patterns utilizing sensors in the context of PD detection utilizing human gait. The system’s success depends on the quality of acquired data; hence, extreme caution must be maintained when collecting data.

Initially, the foundation of Geissler tubes in 1800 was a good method to examine the joint angles, but this led to considerable computational time (Sutherland, 2002). Some sensor-based methods (e.g., the method by Alam et al. (2017)) applied the shoe sensor-based technique utilizing the support vector machine (SVM) algorithm to detect abnormal gait patterns and achieved an accuracy of 93.10%, a sensitivity of 93.10%, and a specificity of 94.10%. Similarly, Sadeghzadehyazdi, Batabyal & Acton (2021) proposed a deep learning model in which skeletal data captured with a Kinect sensor are employed to draw spatiotemporal features and classify the gait abnormalities. The proposed model is validated on the walking gait dataset and achieves up to 90.57% accuracy.

Further, Vidya & Sasikumar (2022) proposed a hybrid deep learning model involving CNNs and long short-term memory (LSTM) to predict PD and disease severity. They utilized the freely accessible PhysioNet gait dataset, comprising vertical ground reaction force signals from three walking experiments. The foot-sensitive resistors were affixed to each foot, and gait data were collected as the individuals walked at their normal speed. Likewise, Shetty & Rao (2016) proposed a PD detection framework in which gait time-series data are employed as a feature vector followed by a classifier called the Gaussian radial basis function kernel based on the SVM. The findings indicate that seven feature vectors are suitable for detecting PD with an accuracy of 75.00%. Following on, Wahid et al. (2015) suggested employing machine learning algorithms, including random forest and SVM. They utilized quantitative gait features acquired with sensors as input to diagnose PD. Over various experiments, the spatiotemporal features commonly involved stride length, step length, step width, cadence, double support time, stance time, swing time, step time, and stride time. Their work achieved an accuracy of 92.60% utilizing random forest and 80.40% with the SVM. Li & Li (2022) also proposed machine learning classifiers that involve logistic regression, SVM, decision trees, and k-nearest neighbors (KNN) to diagnose patients and healthy controls. The input of these algorithms is the gait features recorded with force sensors. The experiment was performed with the Gait PD dataset available on PhysioNet. Wang, Zeng & Dai (2024) proposed frameworks based on machine learning algorithms to classify gait to detect PD and its severity. Their method is based on signal processing, utilizing time-series data on the vertical ground reaction force. To demonstrate the value of their method, they employed a gait database containing data on 93 patients with PD and 73 healthy subjects. The highest classification method exhibits performance of approximately 98.20% and 96.69% utilizing SVMs.

Some advanced models such as Nguyen et al. (2022) proposed transformer networks by exploiting the one-dimensional data of gait features to detect PD. These features were acquired by a sensor attached to the person’s foot. They employed dimension-reduction techniques to overcome the memory problems of the transformer models. Their suggested method outperforms the existing methods with an accuracy of 95.20% on the PhysioNet database.

As in most of these studies, sensor-based gait acquisition is employed; however, one possible limitation of this acquisition is the cost of sensors. Moreover, this method necessitates that a person cooperates with the system, such as wearing foot pressure sensors.

Sensor-based KOA detection utilizing gait

Other than PD, some wearable sensor methods help determine other conditions, such as KOA, utilizing gait (He, Liu & Yi, 2019). In this context, Kwon et al. (2020) proposed a KOA detection system in which features are extracted from radiographic images of X-rays by employing a deep learning model, namely inception-ResNet-V2. Following on, gait features such as the ankle joint, stride, and others are combined for classification using SVM, followed by KOA disease grading. Their proposed model produces good results with an F1 score of 71.00%, sensitivity of 70.00%, and precision of 76.00%. The findings of this research indicate that in addition to radiographic images, gait data play a complementary role with respect to KOA classification.

Likewise, Kotti et al. (2017) suggested a technique to assess the validity of a rule-driven strategy for comparing 47 participants with KOA versus healthy participants utilizing gait data collected from floor sensors. Similarly, Tereso, Martins & Santos (2015) exploit the use of inertial sensors to acquire the spatiotemporal parameters of gait to identify KOA abnormalities. This sensor-based modality works well yet is limited by variables that include high cost, time and energy consumption, and wearing difficulty.

Vision-based methods for KOA and PD utilizing gait

Some vision-based methods diagnose these diseases utilizing a human gait in which gait data are recorded with a camera. In this context, Kour, Gupta & Arora (2022) presented a vision-based approach to classify human walking patterns as normal, PD, and KOA. Various gait features (e.g., kinematics, motion, and other statistical methods) are extracted, and classification is performed with the KNN algorithm. Fractional-order Darwinian particle swarm optimization (PSO) is also used to increase the detection method’s accuracy and reliability for retrieving regions of interest. Their suggested technique yields 90.00% accuracy, 85.00% sensitivity, 90.44% specificity, and 89.66% precision.

Likewise, some camera-based methods, including digital and charge-coupled device cameras, are also utilized in studies, employing the conventional methods of principal component analysis (PCA), linear discriminant analysis (LDA), sequential backward selection (SBS) Hough transform, kernel principal component analysis (KPCA), and others (Chen et al., 2011). These methods employ the gait features of speed, energy, swing time, and joint angles. Similarly, Verlekar, Soares & Correia (2018) utilized a two-dimensional video-based system in which feet-related features, such as the step length and fraction of the foot, and body-related features with a human gait energy image, including the center of gravity, are employed to detect gait abnormalities. They applied a classification approach based on classic machine learning methods, such as the SVM. The findings indicate that the precision of the center of gravity features is affected by low-resolution videos.

From the above literature, the detection of PD and KOA utilizing gait patterns has primarily been exploited with sensor-based, kinematic, and handcrafted features; however, the classification of these conditions utilizing vision with deep learning has not been adequately explored. For instance, in one study (Kour, Gupta & Arora, 2022), although the vision-based approach is exploited, the authors employed conventional methods, specifically a marker-based vision approach, necessitating a proper laboratory setup and requiring more space and time to acquire the gait. A crucial research question arises here: what if we diagnose these diseases (i.e., via a joint framework for diagnosing both diseases) with deep learning methods along with a transfer learning paradigm? Moreover, what if we employed only vision data that are not high quality or resolution, disregarding other features and laboratory environmental setups?

Datasets

This research employs the KOA–PD–NM gait dataset to detect PD and KOA based on human gait (Kour, Gupta & Arora, 2020). This dataset accounts for various factors in collected gait videos, such as age, gender, and severity levels. For each disease (KOA or PD), videos from varying severity levels are provided, such as early, moderate, and severe. The dataset consists of 96 subjects, including 50 individuals with KOA, 16 with PD, and 30 healthy individuals. There are two sequences for every person, one from left to right and the other from right to left, and all videos are in the format of .mov. Individuals’ faces are not revealed in the videos and are blurred during recording. Moreover, this dataset can also be used in existing studies (Kour, Gupta & Arora, 2022, 2024; Khessiba et al., 2023; Hassine et al., 2024; Martinel, Serrao & Micheloni, 2024; Kour, Sunanda & Arora, 2022; Sousse, Grassa & Manzanera, 2023). More detailed statistics of the dataset are listed in Table 1.

Silhouette segmentation

This research exploits the gait features directly from walking captured utilizing a video camera. Hence, after acquiring videos of healthy and PD or KOA patients from dataset, we segment these silhouettes to extract the region of interest indicating the human gait cycles over several video frames. Silhouette segmentation is one of the most widely utilized gait representations to indicate gait patterns. Hence, to acquire the silhouette gait representation of a person, we fine-tuned the Mask-R-CNN (Kaiming et al., 2017) model on the Pedestrian Detection Database (PennFudan) based on the Faster-R-CNN architecture with an additional branch of segmentation. This architecture is based on two-stage object detection comprising a region proposal network for the proposal of regions and person detection and segmentation.

Spatiotemporal gait patterns from segmented videos

After silhouette segmentation for every walking sequence in the dataset, the stack of silhouettes is utilized to extract spatiotemporal gait features utilizing the VGG16, MobileNet, DenseNet, and GRU models. The fundamental rationale for employing these models is that, while the CNN and its variants are extensively employed for gait classification owing to their ability to extract informative and spatial information, they cannot investigate temporal relationships between gait cycles. In this case, the GRU model is employed to capture the temporal dependencies of the gait cycle (i.e., frame by frame in sequences). Spatiotemporal models are utilized with a transfer learning paradigm to classify gait patterns as PD, KOA, or normal. The models are described in detail below:

VGG16+GRU

The segmented video frames with dimensions of 224 × 224 × 3 are passed as input to the VGG16 (Simonyan, 2014) to extract spatial or semantic information from the segmented frames to acquire the spatiotemporal gait features. More precisely, VGG16 is a state-of-the-art CNN based on deepening the architecture of the network with small convolutional filters (i.e., 3 × 3). The VGG16 model is a much more effective convolutional network architecture that reaches state-of-the-art accuracy on ImageNet Large Scale Visual Recognition Challenge (ILSVRC) classification and localization tasks while being adaptable to other image recognition datasets. This model introduces more convolutional layers in the network by adding small filters. For instance, a stack of three convolutional layers capturing a small receptive field is equivalent to one 7 × 7 filter-based convolutional layer. This concept also introduces more nonlinearity and computational overload of parameters. We employed this architecture trained on ImageNet weights to extract features from the video frames and fine-tune the model with the transfer learning concept.

Temporal features are exploited with the GRU model, where a sequence of spatial features holding the semantic image information is passed to exploit human walking patterns over time. The GRU model is an extended version of recurrent neural networks (RNNs) and long short-term memory (LSTM) networks capable of dealing with the long-term dependencies of gait cycles, frame by frame. The distinction between GRU and LSTM is that GRU eliminates the cell state, employing the hidden state to convey information. This method comprises two gates: the reset and update gates. The working of the update gate is identical to the forget and input gates in the LSTM network, determining what information to discard and what novel information to include.

Similarly, the reset gate determines which information from the past is discarded and forgotten. Because GRUs have fewer tensor operations, they can be trained faster than LSTM networks. The proposed model employs two GRU layers comprising 16 and eight hidden units. The final feature set is passed to fully connected layers to classify the underlying walk as a normal or abnormal gait (KOA or PD). Figure 3 shows a detailed view of the proposed model for classifying normal and abnormal gait patterns.

Hyperparameters and implementation details

Hyperparameters play a major role in fine-tuning the performance of deep models. The hyperparameters of the models include the learning rate (set to 0.01) and the weight optimizer (set to Adam). The models run for 100 epochs; however, early stopping callbacks are set based on the validation loss to prevent overfitting. Likewise, a dropout rate of 0.3 is set. Frame sizes are set to 224 × 224 × 3, and the sequence length of frames is set to 100. Moreover, the implementation details include the hardware/software platforms; hence, all implementation is conducted in the Python language. Simulations are run on Google Colab with an Nvidia Tesla K80 graphics processing unit (GPU) with 12 GB of video random access memory (VRAM). These Colab notebooks are operated by a virtual machine running Ubuntu Linux as an operating system.

Evaluation metrics

The metrics employed to evaluate the model’s performance include accuracy, precision, recall, and the F1 score which are considered to be the standard metrics for classification problem in machine learning. The rationale of utilizing these metrics is to indicate the comprehensive performance of proposed model in detecting normal and abnormal gait patterns associated with KOA and PD. The following mathematical equations evaluate the performance of the model:

(2) $$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(3) $$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(4) $$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(5) $$\mathrm{F}1\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=2.\frac{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}.\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}},$$

In the above Eqs. (2)–(5), TP indicates the true positives, TN denotes true negatives, FP represents false positives, and FN indicates the false negatives.

Experiments and results

This section examines the suggested model’s performance on a comprehensive dataset. The performance is evaluated using classification accuracy, which demonstrates its ability to discriminate between normal and pathological gait patterns. As mentioned in section “Materials and Methods”, silhouettes are extracted from the raw video frames. Some samples of video frames and their corresponding segmented silhouettes are depicted in Fig. 4. Furthermore, Fig. 4 depicts silhouettes of different gait patterns, including KOA (first row), PD (second row), and normal (third row). As observed from the silhouettes, the head is blurry because it is intentionally hidden in the original frames during recording to anonymize the identities of participants for data collection. After the extraction of silhouettes, frame by frame, if these silhouettes are again combined to form the videos, it would result in segmented videos. To address the limited data problem, we augmented the training frames of the sequences with additive Gaussian noise (i.e., augmentation that does not disrupt the gait cycle patterns).

Following on, the segmented videos are input into the proposed models, including VGG16+GRU, MobileNet+GRU, and DenseNet+GRU, to acquire the spatiotemporal gait patterns of subjects walking in a video to classify their gait as abnormal (KOA or PD) and normal. Initially, we tested the algorithm with three classifications: KOA, PD, or healthy. The training and testing videos were divided with a ratio of 80:20. Moreover, we shuffled the dataset five times for each model to accurately assess the performance of the model. Table 2 presents the results of the VGG16+GRU model, revealing that the model performs very well in different runs with the random shuffling of the training and testing datasets. The highest metrics that the model achieves are 83.78% accuracy, 59.00% precision, 64.00% recall, and 61.00% F1 score. Table 2 also provides the mean ± standard deviation for all simulations. The overall values of the mean accuracy, precision, recall, and F1 score of the proposed VGG16+GRU model to detect gait abnormalities are 80.54%, 56.40%, 61.00%, and 58.40%, respectively.

After validating the results of the proposed VGG16+GRU model, the same experimental setup was employed for MobileNet+GRU. The training and testing videos were divided like in the previous scenario, with an 80:20 ratio. Furthermore, we shuffled the dataset five times for each model to correctly assess its performance. Table 3 presents the findings of the MobileNet+GRU model, demonstrating that the model works well in different runs with random shuffling of the training and testing datasets. The model’s highest metrics are 83.78% accuracy, 59.00% precision, 64.00% recall, and 61.00% F1 score. Table 3 also includes the mean × standard deviation. The low deviation in the accuracy, precision, recall, and F1 score performance metrics indicates that MobileNet+GRU exhibits good results.

The mean overall accuracy, precision, recall, and F1 score values attained by this model are about 81.08%, 57.00%, 61.60%, and 58.60%, respectively. However, when comparing MobileNet+GRU and VGG16+GRU, the results of MobileNet+GRU are slightly better than VGG16+GRU in terms of accuracy, precision, recall, and the F1 score.

As with the prior two models, DenseNet+GRU is also validated with the same experimental settings. Table 4 presents the results of the DenseNet+GRU model. The results indicate that the model performs admirably across runs with random shuffling of the training and testing datasets. The model’s highest metrics are 83.78% for accuracy, 59.00% for precision, 64.00% for recall, and 61.00% for the F1 score. The overall performance for accuracy, precision, recall, and the F1 score with mean plus or minus the standard deviation is about 80.54%, 57.00%, 61.00%, and 58.40%, respectively. If the performance of all models is contrasted, the performance of MobileNet+GRU is slightly better than that of the other two models. In addition, the training and testing plots for all three models are also presented below, indicating the accuracy and loss values over a set of epochs. The epochs were set initially to 100; however, a callback is also set to stop early if the validation loss does not improve. Figure 5 indicates that the proposed model exhibits good convergence over epochs and reaches the optimal values of accuracy and loss.

Subsequently, another experimental scenario is also implemented in which the KOA and PD classes are combined into a single abnormal class, with the other class representing healthy gait patterns, resulting in a binary classification challenge. The reason for accessing the models in this way is the limited number of videos for PD abnormalities in the dataset, as indicated in Table 1. Second, it offers another experimental test to determine the performance from a different perspective. In this case, we kept the experimental settings from the previous scenarios (i.e., the training, testing, and shuffling of datasets).

Tables 5–7 present the results of all three models. For each model, we shuffled the datasets and tested the model’s generalizability in detecting abnormal and normal gait patterns. The mean plus or minus the standard deviation of all three models was also computed. Table 5 reveals that VGG16+GRU has the highest accuracy, precision, recall, and F1 score at about 99.00% for all four metrics. Similarly, the mean plus or minus the standard deviation of the accuracy, precision, recall, and F1 score of the VGG16+GRU model is about 94.81 ± 2.32%, 94.80 ± 2.86%, 93.60 ± 3.01%, and 94.20 ± 2.48%, respectively. Likewise, the proposed MobileNet+GRU model results are also encouraging in this binary case, achieving the highest accuracy, precision, recall, and F1 score values of 98.10%, 93.00%, 96.00%, and 94.00%. The mean plus or minus the standard deviation of this model in the binary case in terms of accuracy, precision, recall, and F1 score is about 94.21 ± 3.33%, 92.20 ± 2.86%, 93.60 ± 3.44%, and 92.80 ± 3.06%. Similarly, in Table 7, DenseNet performs well, achieving an overall accuracy, precision, recall, and F1 score of about 99.00%. In this scenario, if the performance of all models is contrasted, then the overall VGG16+GRU outperforms the other two models: MobileNet+GRU and DenseNet+GRU. In addition, the training history plots of the accuracy and loss over several epochs are also plotted for all three models in Fig. 6.

Furthermore, when model performance is analyzed depending on two experimental settings, the second experimental setting (i.e., binary classification) has better results. The rationale for the weak performance in the first experimental scenario is that the model has trouble detecting abnormal gaits associated with KOA or PD. However, when combined, both are recognized as abnormal gait patterns, and the model precisely distinguishes them from normal gait. Further, the number of samples per severity level is relatively low, particularly in the case of PD. To address this, we augmented the frames of the sequences with additive Gaussian noise (i.e., augmentation that does not disrupt the gait cycle patterns), yet the model still lags. This problem can be addressed by increasing the diversity of samples in the dataset. In addition, the receiver operating characteristic curve is also depicted in Fig. 7 with a high area under the curve score of 0.89, reflecting the better ability of the model to balance TP and FP rates, indicating its effectiveness in the underlying tasks.

Discussion, analysis, and comparisons

Gait is a direct indication of one’s health, driving the present-day attempts to streamline the study of abnormal gait to help clinicians make decisions (Ortells, Herrero-Ezquerro & Mollineda, 2018). Existing studies have employed several methods to exploit gait patterns in diagnosing diseases or abnormalities. However, most existing work focuses on sensor-based gait acquisition, including the stride and step length, step breadth, and so on, whereas temporal aspects deal with the sequence of gait cycle events, such as cadence (Kour & Arora, 2019).

Despite adopting a vision-based approach, some studies have employed statistical and conventional methods to solve this problem (Kour, Gupta & Arora, 2022; Ortells, Herrero-Ezquerro & Mollineda, 2018). Nevertheless, these models have not been investigated with state-of-the-art deep learning models. A question arises: how can existing conventional methods be automated via advanced methods? Hence, one of the significant contributions of this research is to exploit gait patterns in diagnosing abnormalities utilizing deep learning models capable of extracting spatiotemporal gait features directly from human silhouettes. The proposed technique is less expensive because it utilizes low-resolution videos rather than expensive sensors, offering advantages over clinical measure-based illness detection, which is limited owing to a lack of medical data. Moreover, this method does not require a proper laboratory setup, necessitating more space and time to acquire the gait, as in existing methods (Kour, Gupta & Arora, 2022).

These contributions are justified by the good results of the proposed deep learning models in diagnosing normal and abnormal gait patterns. The proposed model extracts the spatiotemporal features. We froze the initial layers of VGG16, MobileNet, and Dense-Net trained on ImageNet to draw features and fine-tune those features via a transfer learning strategy. These state-of-the-art CNNs extract spatial information indicating semantic information from silhouette images, such as a person’s body posture and shape. During walking, the gait cycle temporal information is exploited with one of the variants of RNNs: GRU. We tested the model in various scenarios: two- and three-class categorizations of gait patterns. In both scenarios, the proposed model displays good results. Although the suggested model is based on deep learning paradigms that are fundamentally black-box, it may be made more interpretable by including attention processes, feature attribution methods (e.g., SHAP, Grad-CAM), and other explainability approaches.

To further demonstrate the worth of the proposed work, a comparison against existing methods is also conducted as shown in Table 8. In Table 8, the vision-based data are exploited with the KOA–PD–NM dataset in only two studies to classify the gait patterns as normal and abnormal. Kour, Gupta & Arora (2022) employed fractional-order Darwinian particle swarm optimization to segment regions of interest and extract gait features. Afterward, the KNN classifier was employed for the extracted feature values. Likewise, Khessiba et al. (2023) employed a graph convolutional network (GCN) in which nodes of the graph denote the body joint positions and achieve a good accuracy of 93.75%. Furthermore, unlike computationally demanding and sophisticated techniques such as PSO-based optimization and GCN-based advanced models, our method achieves great accuracy while using substantially less computational resources.

In addition to the vision-based method, other methods have also been designed in which sensors are utilized to acquire gait data, including the number of footsteps, speed, step time, stance time, swing time, cadence, swing stance ratio, and stride interval-based features. These features are input into machine learning to classify healthy and abnormal gait patterns (e.g., Wahid et al. (2015) and Balaji, Brindha & Balakrishnan (2020)). Kwon et al. (2020) employed medical data (radiographic imaging X-rays), including kinetic and kinematic gait data, to detect KOA. However, these methods (Wahid et al., 2015; Kwon et al., 2020; Balaji, Brindha & Balakrishnan, 2020) were only investigated for PD or KOA. Compared with existing studies in Table 8, the proposed work is the extended version of identifying KOA, PD, and healthy gait patterns. Furthermore, the suggested technique is a vision-based approach operating on low-resolution videos, making this method cost-effective. The validity of the research is supported by different elements such as from existing literature, it is clearly concluded that human gait is disrupted by conditions such as Parkinson’s disease (PD) and knee osteoarthritis (KOA). The study ensures the proper analysis of gait features by extracting silhouettes to acquire the spatiotemporal features from gait videos. The performance of the proposed model using different metrics indicates the worth of this study in diagnosing PD and KOA. Furthermore, the approach’s focus on automated, non-invasive analysis makes it a viable alternative to established clinical approaches, highlighting its prospects for real-world use. The practical implications of this research include early detection and screening of patients with KOA and PD disease using a cost-effective, non-invasive, and automated approach. Early assessment, monitoring of patients, and rehabilitation planning might all be greatly improved by incorporating such a cost-effective deep learning-based model into clinical practice. The model can help clinicians diagnose KOA and Parkinson’s disease by monitoring gait patterns before additional symptoms worsen. It might be particularly useful in telemedicine and remote health monitoring when patients do not have frequent access to professionals. Despite traditional gait analysis techniques, which necessitate motion-capturing systems or wearable sensors, the technique presented in this study is based simply on video analysis, making it more accessible and affordable. As a result, refining the model into a lightweight architecture that ensures consistent performance on low-resource devices is an important issue for future development. However, the trade-off between model accuracy and computing complexity remains an important consideration. In healthcare applications when patient health is at risk, accuracy takes priority over complexity, resulting in dependable and accurate diagnostic support. Following that, data privacy and security are also critical since the system uses patient gait videos, which may be handled by implementing secure data storage, anonymization techniques, and adherence to healthcare rules. Federated learning-based algorithms are also a promising future avenue for increasing framework security. The proposed model may be connected with existing diagnostic tools such as electronic health records (EHRs) and decision-support systems, providing doctors with extra guidance and insight.

Rather than focusing just on the advantages, it is critical to discuss the limitations of this study for future researchers to expand on it. Hence, one possible limitation of this work is validation on a single small-scale dataset because there are no other publicly available vision-based gait datasets for diagnosing these abnormalities (KOA or PD). Second, the KOA–PD–NM dataset utilized in this research is also limited in the number of videos, camera viewpoints, and subjects, but it serves as a key foundation for future research in gait-based classification for KOA and PD detection. Particularly, the total number of PD videos is limited and leads to class imbalance and overfitting. Although, the dataset maintains a good diversity by considering different age groups, gender, and height, but still, a more diversified dataset is required to improve the performance. Hence, in the future, we aim to collect a large publicly available vision-based gait dataset to validate the results because the model’s robustness and capacity to generalize across different patient groups would be improved with a larger and more diversified dataset. Similarly, more effective gait representations, including gait energy images or gait entropy images (Bashir, Xiang & Gong, 2009) with deep learning tactics, can also be explored. In addition, currently, the model is trained on specific labeled abnormalities i.e., KOA and PD using the supervised learning concept, however, to generalize it well on other abnormalities, incorporating self-supervised learning or few-shot learning techniques could also be a future direction. Lastly, to enhance the model’s robustness against adversarial attack-based perturbations (i.e., noise), adversarial training can be incorporated as an additional defense mechanism.