MS is a central nervous system disease that causes ataxia and balance disorders. Due to the destruction of the central nervous system, the transmission of nerves is impaired. This deterioration can cause a variety of symptoms. MS symptoms can vary from person to person. In addition, the symptoms differ depending on the region of the disease (Güner and Alsancak, 2011).

Gait disturbance, a symptom of ataxia, is one of the most common problems in MS patients. Gait disruption was identified by 75–85 % of the patients as the primary issue (Reyhan, 2022). As a result of this disease, inadequacies such as spasticity, coordination disorders, weakness, and vestibular problems occurred. This showed that patients typically walked with shorter stride lengths and increased double support phase. Fatigue is also observed in patients with MS; It can negatively affect walking parameters (distance and walking time).

In this paper, a CapsNet-based autonomous system was suggested for detecting ataxia in MS patients. Design of the CapsNet model and feature extraction are the two basic approaches followed in this study. Firstly, specific feature extraction techniques were carefully applied to extract valuable information from static plantar pressure distribution images to identify ataxia more sensitively and specifically in MS patients. CapsNet, the second approach applied, is an extended type of neural network. CapsNet is different from traditional neural networks. Because CapsNet has a hierarchical structure that represents the properties of objects and their views from different angles in different capsules. The main goal of these steps is to focus on relevant features to identify ataxia symptoms more effectively. Finally, it is seen that the CapsNet model is not employed in the literature to accurately distinguish between ataxia and healthy individuals by utilizing static pressure distribution data in patients with MS. Therefore, the proposed study is the first study aimed at detecting ataxia with the CapsNet model in people with MS disease.

The other parts of the study are organized as follows: In the second part, the methods applied in the literature were examined. In the third section, information about the design of the designed CapsNet model was presented. In chapter 4, the results obtained in the study were summarized with tables and graphs. In the last section, the obtained results were discussed in detail with their pros and cons. In the last section, suggestions were emphasized about the methods and methods to be applied in future studies for the diagnosis of ataxia symptom in MS patients.

Static plantar pressure occurs between the foot and the support surface. It is also fundamental in daily locomotor activities. Plantar pressure measurement information is crucial in the diagnosis of gait and posture and in the determination of lower extremity problems (Razak et al., 2012). In previous studies, it was stated that there may be changes in body posture and abnormal plantar pressure distribution in people with temporomandibular dysfunction (Souza et al., 2014).

The number of research on the identification of ataxia in individuals with MS using pictures comprising the static plantar pressure distribution is restricted in the literature. Furthermore, practically all of the investigations in the literature evaluated individuals who were moderately or seriously impaired (Pradhan et al., 2015). In addition, the diagnosis of ataxia was investigated using classical machine learning methods in most of the studies (Güner and Alsancak, 2011; Reyhan, 2022; Balgetir et al., 2021; Kaya et al., 2022; Salamci et al., 2022; Katmerlikaya, 2021).

Two separate studies in the literature by Praet and Louwerens, as well as Queen et al., found that neuropathic forefoot rocker-soled shoes can effectively reduce pressure between the first and fifth metatarsal heads. It has also been stated in these studies that shoe designs that reduce metatarsal pressure in the future will prevent metatarsal stress fractures in men and women (Mueller, 1999; Praet and Louwerens, 2003; Queen et al., 2010).

Summa et al. aimed to test the usability of the Kinect system to assess ataxia severity, explore the potential of clustering algorithms, and validate this system with a standard motion capture system. Gait assessment was performed on the same day with standard gait analysis and Kinect v2 in a group of young patients (mean age 13.8 ± 7.2 years). Spatial-temporal parameters of gait were analyzed and the differences between the two systems were examined through correlation and fit tests. An accuracy of 90.4 % was achieved with Support Vector Machines (SVM) (Summa et al., 2020).

In a study carried out by Holzreiter et al., artificial neural networks (ANN) and machine learning algorithms were applied to detect healthy and abnormal gait using standard gait analysis equipment (Holzreiter and Köhle, 1993). First, the data set was divided into 20 % test group, 80 % training group, then 40 % test group and 60 % training group. The highest accuracy rate of 95 % was yielded at approximately.

LeMoyne et al. aimed to facilitate the acuity of the timed 25-step walking test with the synthesis of wearable and wireless inertial body sensors and machine learning. During the timed 25-step walking test, inertial sensors mounted near the ankle joint generate a feature set for a subject with ataxia and healthy gait. With the multilayer neural network, it has provided an objective classification of gait models based on their quantitative properties. This study has yielded significant accuracy values for a relatively small number of samples. Also, for gyroscope return signals, these features yielded 74 % classification accuracy, and gyroscope deviation signal features achieved 63 % classification accuracy (LeMoyne et al., June).

In a study conducted by Procházka et al., a DL-based convolutional neural network system was used in conjunction with accelerometer data to differentiate between ataxia and normal gait. The experimental dataset includes 860 signal segments from the control set, which consisted of 16 ataxic patients and 19 normal patients with a mean age of 38.6 and 39.6 years, respectively. The classification success was compared with the results obtained with SVM, Bayes methods and standard methods including a two-layer neural network with properties estimated as relative power in selected frequency bands. The results show that proper selection of sensor locations can increase accuracy from 81.2 % for foot position to 91.7 % for spine position. Combining the input data and DL methodology with five layers increased the accuracy to 95.8 % (Procházka et al., 2021).

Another study was carried out by Dostal et al. to find the optimal sensor location for ataxia classification. Information was recorded during walking in the group consisting of 7 patients with gait ataxia in MS and 7 healthy volunteers. Each recording is divided into several short (20 s) segments, counting a total of 139 segments. The power spectral density of the accelerometer data was classified by SVM, Bayesian Analysis, K-nearest neighbor, and ANN. Accuracy is compared for different sensor locations. The highest accuracy for the different sensor positions was yielded at the head, neck and shoulders (up to 99.6 %). The feet (65–88 %) and legs (76–86 %) gave the worst results, depending on the classification method (Dostal et al., 2020).

Balgetir et al. implemented a different approach to detecting ataxia in MS patients. In this study, it was aimed to detect the symptom of ataxia using a DL-based approach with images showing the plantar pressure distribution of the patients. A total of 105 images showing the plantar pressure distribution of 43 ataxic PwMS and 62 healthy individuals were analyzed. Images are resized for models such as VGG16, VGG19, ResNet, DenseNet, MobileNet, NasNetMobile and NasNetLarge. Feature vectors were extracted from the resized images and then classification was performed using SVM, K-nearest neighbors and ANN. 10-fold cross validation was applied to increase the validity of the classifiers. The VGG19-SVM hybrid model outperformed the highest accuracy, sensitivity and specificity. (89.23 %, 89.65 %, and 88.88 %, respectively) (Balgetir et al., 2021).

In another study completed by Kaya et al., it was aimed to detect ataxia in MS patients with a DL-based approach using an image dataset containing static plantar pressure distribution. A total of 406 static bipedal pressure distribution image data for 43 ataxic MS and 62 healthy individuals were used in the study. After preprocessing, these images were arranged as input to pre-trained DL models such as VGG16, VGG19, ResNet, DenseNet, MobileNet and NasNetMobile. The performance of the proposed models was evaluated with accuracy, sensitivity, specificity and F1-score criteria. The VGG19-SVM hybrid model performed best with 95.12 % accuracy, 94.91 % sensitivity, 95.31 % specificity, and 94.44 % F1-score (Kaya et al., 2022).

Unlike the studies in the literature, this proposed study aimed to detect ataxia diagnosis in MS patients using static plantar pressure data with a DL based CapsNet approach for the first time as far as we know. The model created with the CapsNet network was trained to perform high-performance diagnosis after various preprocessing steps with images obtained from healthy and ataxia-diagnosed patients. Since the results are promising, it is emphasized that this method can be an effective method in the diagnosis of ataxia in MS patients. Moreover, the findings highlight the idea that static plantar pressure data can be applied to diagnose attacks. Furthermore, this proposed approach provides a promising basis for future research.