Artificial Intelligence Applications in the Diagnosis of Neuromuscular Diseases: A Narrative Review

The accurate diagnosis of neuromuscular diseases (NMD) is in many cases difficult; the starting point is the clinical approach based on the course of the disease and a careful physical examination of the patient. Electrodiagnostic tests, imaging, muscle biopsy, and genetics are fundamental complementary studies for the diagnosis of NMD. The large volume of data obtained from such studies makes it necessary to look for efficient solutions, such as artificial intelligence (AI) applications, which can help classify, synthesize, and organize the information of patients with NMD to facilitate their accurate and timely diagnosis. The objective of this study was to describe the usefulness of AI applications in the diagnosis of patients with neuromuscular diseases. A narrative review was done, including publications on artificial intelligence applied to the diagnostic methods of NMD currently existing. Twelve studies were included. Two of the studies focused on muscle ultrasound, five of the studies on muscle MRI, two studies on electromyography, two studies on amyotrophic lateral sclerosis (ALS) biomarkers, and one study on genes related to myopathies. The accuracy of classification using different classification algorithms used in each of the studies included in this narrative review was already 90% in most studies. In conclusion, the future design of more accurate algorithms applied to NMD with greater precision will have an impact on the earlier diagnosis of this group of diseases.


Introduction And Background
Neuromuscular diseases (NMDs) are a heterogeneous group of diseases that affect the muscle and the nerve, the neuromuscular junction, and the motor neuron, the etiology of NMD is variable, but in general there can be classified in the NMD of genetic etiology [1], and the NMD of non-genetic etiology that occur in infectious, autoimmune, inflammatory, paraneoplastic, and toxic processes [2].

Diagnosis of NMD
The diagnosis of NMD could be difficult and sometimes late, due to the variability of clinical manifestations.The initial diagnostic approach of all NMDs of both genetic and non-genetic causes is clinical, based on the symptoms presented by the patient, considering the age and the type of primary involvement of the peripheral nervous system suffered.[14] Machine Learning (ML)

TABLE 1: Domains and subdomains of Artificial Intelligence
As mentioned above, ML is one of the domains of AI, consist of a mathematical model that can improve performance on a task when exposed to data.To design ML models, algorithms are necessary, which are the sequences of instructions that a computer system must comply with to achieve the resolution of a problem.According to the complexity of the model, the strategies and functions applied vary.There are four basic methods of machine learning.The first is supervised machine learning, the second is unsupervised learning, the third is semi-supervised learning, and the last is reinforcement machine learning.[6], [7].
In the supervised machine learning, the model is provided with the correct answer in advance, therefore, it knows the question and knows the answer, that is, patterns defined between questions and answers.An example of the application of supervised AA is the automated reading of electrocardiograms [8].To develop a supervised model, labeled data, i.e., the primary data are preprocessed by identification or notation.This way it is possible to know if the data corresponds to images, text files, video.The main utilities of labeled data are obtaining more accurate predictions and better usability of the data [9].The process of developing a supervised ML model consists of training the algorithm, for which it is necessary to have three sets of data.The first is the training dataset from which the algorithm learns the optimal parameters to accomplish the task.The second is the test set, this is a dataset on which the performance of a parameterized model is based.The third group of data is the validation set, used to evaluate model performance, differs from the test set used during training to establish hyperparameters or aspects of the model architecture that can modifies to optimize model performance [6].
The methods used to develop a supervised ML model are of 2 types: regression and classification.Regression methods include linear regression, generalized linear models, penalized regression models, regression trees, and vector support regression.The classification methods use Bayesian models, vector support (SVM), k nearest neighbors and, neural networks which will discuss in detail [10].
Unsupervised machine learning usually involves the analysis of unlabeled data, under assumptions about the structural properties of the data.The data can be algebraic, combinatorial, or probabilistic.The different assumptions established by methods of dimension reduction, factor analysis, random projections, and automatic encoders or by statistical principles such as moments or by Bayesian models [6].The methods used to make unsupervised ML models are combination algorithms, self-organizing maps, and hierarchical clustering methods [10].The semi-supervised ML model that uses labeled and unlabeled data.It aims to provide a better result for prediction.Application areas of semi-supervised models between others, are machine translation, fraud detection, data labeling and text classification [11].
And, machine learning by reinforcement, this method is based on the model's ability to take actions according to patterns that are neither explicit nor predetermined, with cost and reward functions.In health care, a booster ML model would function as an agent between the medical device, the computer equipment, and the system.The model takes a particular action in a medical setting to elicit a specific reward and then uses evaluative feedback to improve its performance.Thus, reinforcement learning applied to clinical decision-making tools of treatments that require frequent modifications and for medical diagnosis to indicate the treatment initiated, indicate alerts of the treatment, or question the human decision if it does not fit the interactions expected by the model [12].
Next, a type of algorithm called artificial neural networks (NN) is explained in detail, its importance lies in the fact that it is the classification algorithms most frequently used in AI applications in medicine.An NN is a calculation algorithm that requires training, it is an analogy of the functioning of biological neurons and emulates the learning capacity of the nervous system.An NN consists of multiple simple processors or connected nodes called neurons.Each node or neuron represents the output of another neuron with and assigned value according to the degree of interconnection, thus obtaining the weighting of the input, this is necessary for the training of the algorithm [7] [13].The NN are composed of three layers or levels of nodes called neuronal layers, which resemble the layers of the cerebral cortex, the first is an input layer containing the neurons with the data from which obtains the result proposed by the model to use.The second is the hidden layer also called "black box", formed in turn by one or more layers not visible or observable, in multilayer models the output layer will correspond to the input of the next layer, and the third layer is the output or result.The connections between neurons can be forward, backward, or lateral [7].

Deep learning
Implies supervised and unsupervised ML techniques.This is what happens when an NN starts data processing and obtains basic knowledge, apart from understanding the data it also learns from it, that is, the model delves into what it learns by itself without following a list of predetermined tasks.The input layer connects forward with the hidden layer, the interaction of the data is not visible or known and will be iterative depending on the number of layers created in the model and to the right of the model is the layer of output neurons corresponding to the result.In an applied model the connections between neurons are not exactly one to one [14].

Convolutional neural networks (CNN)
Convolutional networks are a type of neural networks that use a linear function called convolution.The purpose of convolution is to extract a limited number of features from each layer, so a filter slide over the input to produce a rectified activation map for feature identification.The next layer relates to that limited number of features, the function repeats depending on the number of convolutional layers of the model until it reaches the output layer [15].The kernel function is the most frequently used application in the elaboration of CNN models, consists of the conversion of a space of few dimensions in a space of greater dimension through complex operations of the data, it is a function that quantifies the similarity between two observations in a new dimensional space [16] .

Machine Learning in Medicine
Machine learning technology allows doctors and researchers to predict which treatment and prevention strategies more accurately would be effective for a particular disease and in which groups of people those strategies will work.It requires sufficient computing power, algorithms that can learn by themselves at a high speed (e.g., deep learning) [17].In the clinical field of practice, informatic tools as Clinical Decision Support Systems (CDSS) aim to improve medical care by supporting clinical decisions, comparing a patient's characteristics with a computerized database of clinical knowledge, the output data in this case are suggestions about specific evaluations or recommendations presented to the physician or patient for decision making [18].
One of the applications of AI in ML-based image preprocessing via CNN is segmentation.It consists of delimiting the structure of the organ under study following the anatomical limits for each layer of the region of interest (ROI) did it by technicians in the anatomical tracing using specific software tools.MRI can have hundreds of layers depending on the resolution of the study images and therefore manual segmentation is a slow, subjective, and laborious process [19].The analysis of textures in muscles images allows the monitoring of NMD.For such an analysis, first-order statistical tests evaluate the distribution of gray-level frequencies of the pixel intensity histogram in a region of interest.Second-order statistical tests can be based on a co-occurrence matrix and include entropy, energy, homogeneity, dissimilarity, and correlation; And higher-order statistical tests like contrast, roughness, and occupancy can be calculated using neighboring gray-tone difference matrices, which examine the location and relationships between three or more pixels [20].
It is important to note that machine learning techniques will always have a certain error rate, although it is small.This is why it is essential to evaluate the accuracy of the model by the percentage of correct answers obtained when applying it [5].The application of AI has ethical implications, not only in the health field.This is where the concept of explainability arises also referred to as interpretability and / or transparency of the algorithms used, understood as a characteristic that allows a person to reconstruct why in a certain AI model the predictions presented have occurred [21].
The objective of this study was to describe the usefulness of AI applications in the diagnosis of patients with neuromuscular diseases.

Review
The review included studies on artificial intelligence applied to the diagnostic methods of NMD.Following, the list with specific search strategies for each of the databases and combinations of MeSH terms.After screening titles and abstracts and eliminating duplicates, a total of 12 studies were included in the present narrative review on artificial intelligence applied to currently existing NMD diagnostic methods.Two of the studies analyzed the application of AI in muscle ultrasound, 5 of the studies focused on muscle MRI, 2 studies on EMG, 2 studies of ALS biomarkers and 1 study on identification of genes with pathogenic variants related to myopathy.Table 2 summarizes characteristics of studies included in the narrative review.

Muscle imaging studies
Two of the included studies focused on muscle ultrasound and 5 of the studies on muscle MRI, then discuss each of the studies in more detail [22,23][ [24][25][26][27][28].

AI in Muscle Ultrasound
Two of the studies analyzed the application of AI in ultrasound studies [22], [23].In Marzola's study, focused on muscle segmentation for its ultrasound study, applied deep learning models with multiple learning architectures for the evaluation of NMD.Specifically, they performed the calculation of the mean grayscale value for automatic cross-sectional area in muscle ultrasound images acquired at different anatomical sites.The dataset of Marzola`s study included 3917 images of the acquired biceps brachii, tibialis anterior and medial gastrocnemii of 1283 subjects (mean age 50 ± 21 years, 729 males).The reported accuracy was 0.91 ± 0.08 for training data, 0.89 ± 0.10 for validation data, and 0.88 ± 0.12 for test data.They concluded that the applied model allowed to segment with optimal precision the cross-sectional area of the muscles evaluated by ultrasound [22].In Nodera's study, they used classification algorithms to differentiate between inflammatory myopathies and myotonic dystrophy.In the study they performed a texture analysis of muscle ultrasound images of patients with inflammatory myopathies, of these 11 patients had inclusion body myositis and 21 polymyositis/dermatomyositis with 19 patients with myotonic dystrophy (DM) type 1.The classification was made by means of an algorithm of the WEKA program.The texture study focused on the analysis of roughness with the algorithm of comparison with neighboring pixels.Average pixel values were similar in all three groups.However, they found significant differences for pixel standard deviation values, entropy histogram, gray-level length matrix and gray-level non-uniformity level.The classification accuracy obtained was 76% [23].

TABLE 2: AI Precision Results in Muscle Magnetic Resonance Images
In Chen's study, implemented a three-dimensional artificial network for segmentation of muscle MRI images.The images of 24 subjects with NMD compared with those of 19 healthy subjects.Model accuracy was assessed using pixel accuracy and Dice coefficient (CD) compared to manual methods.CD values ranged from 0.83 ± 0.17 to 0.98 ± 0.02 in thigh and from 0.63 ± 0.18 to 0.96 ± 0.02 in twins.The fat fraction accuracy values were 0.989 in thigh and 0.971 in twins [24].
Yang et al studied a collection of MRI images of 148 cases of dystrophinopathies and 284 controls with other muscle diseases, studied 2536 images of the total of 432 cases included.The region of interest was in the muscles of the right thigh.They used a deep learning model with CNNC.The accuracy of the results of three expert radiologists was compared with the result of the model.The accuracy of the model was 0.91, while the accuracy results of the 3 expert radiologists with a 95% CI were 0.80± 3, 0.84± 3 and 0.84± 3, respectively.They concluded that the performance of the applied deep learning model was comparable to that observed with experienced radiologists [25].
In Felisaz's study, on the determination of the fat fraction in muscle MRI.They applied 8 machine learning models: linear regression, collapsed regression, Lasso regression, generalized additive models, regression tree, random forest, k-nearest neighbors, support vector and mixed effect model.With them they sought to predict water content in T2 (wT2) and fat fraction by non-quantitative MRI analysis.They analyzed a dataset consisting of muscle MRI images with varying degrees of intramuscular fat replacement and edema obtained from 14 patients with facioscapulohumeral dystrophy (DFEH).To calculate the predictive performance of the model, the absolute mean error and the square mean error were used, they did not specify the percentage precision of either model.The feasibility of predicting the fat fraction parameters of quantitative MRI was demonstrated, using texture analysis and machine learning methods from conventional T2W images [26].
In Verdú-Díaz's study, they analyzed muscle MRI studies of patients with muscular dystrophies.The machine learning strategy was based on the random forest strategy, they compared the results with the concepts of 4 expert radiologists.They included in the analysis 976 MRI images of pelvic, thigh and leg muscles of patients with different muscular dystrophies with genetic confirmation, after evaluating 2000 random forest models, the final model had a diagnostic accuracy of 95.7% for 10 types of muscular dystrophy [27].
Gadermayr's study was based on muscle MRI, which included 41 participants of whom 7 had moderate myopathy characterized by small foci of fat infiltration, 13 severe myopathies characterized by severe fat infiltration of muscles and 21 were healthy controls.They used a data gain model (GAN), employed an CNN as a discriminator, and performed muscle segmentation with biomedical data segmentation applications.
For the analysis of the results, the Dice similarity coefficient was applied, whose result was 0.91 for cases of moderate myopathy and 0.88 for cases of severe myopathy [28].

AI in NMD electrodiagnosis
Two of the included studies applied AI models to electromyography studies.In Yaman's study, they classified and segmented EMG data.MUAP data was obtained in 2048 EMG samples from 7 patients with myopathy and 13 with neuropathy and 7 controls.They made the afore mentioned classification and segmentation of the electromyographic images using the functions of disaggregated wave transformation and wave packing transform, and then applied decision trees to finally combine and power the resulting algorithms.The reported classification accuracy was 97.67% [29,30].In the study by Nodera et al, extracted the characteristic patterns of 6 types of electromyographic discharges so that were recognized by automatic computerized algorithms.They collected both graphic and sound electromyographic recordings.The accuracy of the classification model was 90.4% [30].

AI for detection of amyotrophic lateral sclerosis (ALS) biomarkers
Two studies on ALS biomarkers included [31,32], Greco's study on biomarker detection in blood samples, included 726 blood samples obtained from 41 patients with amyotrophic lateral sclerosis (ALS) and 25 patients with other lower motor neuron diseases (LMND), over 10 years.A vector support machine (SMV) model was used, 692 blood samples from 35 ALS patients and 20 from LMND patients were used to construct the classification dataset, the remaining 34 samples from 6 ALS patients and 5 LMND patients were used for model evaluation.They made measurements of interleukins, lymphocyte subpopulation count, measurement of immunoglobulins A, E, G, and M, cell adhesion factors, cell growth and survival factors, and routine blood biochemistry.The most informative analytes for the entire patient group were the proportion of monocytes, IgM levels and CD3 lymphocyte count.The model's classification accuracy was 0.87 for ALS patients and 0.93 for LMND patients.The study showed the importance of immunological components in motor neuron diseases for discrimination of ALS from other lower motor neuron diseases [31].
Whereas, in Tang's study also on ALS biomarkers detection, authors analyzed ALS patient data from the international open ALS database, including results of different types of blood samples and lung function tests using an unsupervised clustering algorithm.It is striking that when analyzing the accuracy of the model they found that the univariate analysis had an accuracy of 70% compared to the unsupervised clustering model was 95%, explained by the possibility of such a method to do a multivariate analysis [32].

AI in a genetic diagnostic study of NMD
In Tran's study of identifying gene clusters for different muscle diseases, samples were analyzed 1260 patients with muscle weakness of which, 824 were cases of myopathy and 436 were controls.They performed an enrichment analysis on groups of genes.The machine learning model was based on SVM.From a total of 34099 gene expression data, they designed a multiclass classifier based on biological processes.They reported a genetic distinctive consisting of the 500 genes common to all groups, for the specific case of hereditary myopathies the accuracy of the classification was 90% [33].

Limitations
The present study has limitations, the most important is the heterogeneity of the studies of both the groups of diseases analyzed and the diagnostic methods applied.For the future, specific studies of AI for the diagnostic of each NMD are necessary to individualize the classification accuracy and thus determine the usefulness and potential application in clinical practice of the models.

Conclusions
The articles reviewed show that the use of deep learning models for the study and diagnosis of neuromuscular diseases has been increasing.
The usefulness of AI applications in imaging studies is evidenced in the pre-processing of images through segmentation, also in the analysis of the characteristics and abnormalities of the tissues studied.
The results that contrast the accuracy of the classification by the different algorithms used are greater than 90% in most studies, this contributes to the accuracy of the diagnosis of patients with NMD.
other relationships or activities that could appear to have influenced the submitted work.