Learning Spectral Fractional Anisotropy and Mean Diffusivity Features as Neuroimaging Biomarkers for Tracking White Matter Integrity Changes in Myotonic Dystrophy Type 1 Patients using Deep Convolutional Neural Networks.

Myotonic dystrophy type 1 (DM1) is a genetic neuromuscular progressive multisystem disease that results in a broad spectrum of clinical central nervous system (CNS) involvement, including problems with memory, attention, executive functioning, and social cognition. Fractional anisotropy and mean diffusivity along-tract data calculated using diffusion tensor imaging techniques play a vital role in assessing white matter microstructural changes associated with neurodegeneration caused by DM1. In this work, a novel spectrogram-based deep learning method is proposed to characterize white matter network alterations in DM1 with the goal of building a deep learning model as neuroimaging biomarkers of DM1. The proposed method is evaluated on fractional anisotropies and mean diffusivities along-tract data calculated for 25 major white matter tracts of 46 DM1 patients and 96 unaffected controls. The evaluation data consists of a total of 7100 spectrogram images. The model achieved 91% accuracy in identifying DM1, a significant improvement compared to previous methods.Clinical relevance- Clinical care of DM1 is particularly challenging due to DM1 multisystem involvement and the disease variability. Patients with DM1 often experience neurological and psychological symptoms, such as excessive sleepiness and apathy, that greatly impact their quality of life. Some of DM1 CNS symptoms may be responsive to treatment. The goal of this research is to gain a deeper understanding of the impact of DM1 on the CNS and to develop a deep learning model that can serve as a biomarker for the disease, with the potential to be used in future clinical trials as an outcome measure.

Cerebrospinal Fluid Proteomic Changes after Nusinersen in Patients with Spinal Muscular Atrophy.

Disease-modifying treatments have transformed the natural history of spinal muscular atrophy (SMA), but the cellular pathways altered by SMN restoration remain undefined and biomarkers cannot yet precisely predict treatment response. We performed an exploratory cerebrospinal fluid (CSF) proteomic study in a diverse sample of SMA patients treated with nusinersen to elucidate therapeutic pathways and identify predictors of motor improvement. Proteomic analyses were performed on CSF samples collected before treatment (T0) and at 6 months (T6) using an Olink panel to quantify 1113 peptides. A supervised machine learning approach was used to identify proteins that discriminated patients who improved functionally from those who did not after 2 years of treatment. A total of 49 SMA patients were included (10 type 1, 18 type 2, and 21 type 3), ranging in age from 3 months to 65 years. Most proteins showed a decrease in CSF concentration at T6. The machine learning algorithm identified ARSB, ENTPD2, NEFL, and IFI30 as the proteins most predictive of improvement. The machine learning model was able to predict motor improvement at 2 years with 79.6% accuracy. The results highlight the potential application of CSF biomarkers to predict motor improvement following SMA treatment. Validation in larger datasets is needed.

Cognitive Impairment Analysis of Myotonic Dystrophy via Weakly Supervised Classification of Neuropsychological Features.

The myotonic dystrophies (DM1 and DM2) are dominantly inherited disorders that cause pathological changes throughout the body. Many individuals with DM experience cognitive, behavioral and other functional central nervous system effects that impact their quality of life. The extent of psychological impairment that will develop in each patient is variable and unpredictable. Hence, it is difficult to get strong supervision information like fully ground truth labels for all cognitive involvement patterns. This study is to assess cognitive involvement among healthy controls and patients with DM. The DM cognitive impairment pattern observation is modeled in a weakly supervised setting and supervision information is used to transform the input feature space to a more discriminative representation suitable for pattern observation. This study incorporated results from 59 adults with DM and 92 control subjects. The developed system categorized the neuropsychological testing data into five cognitive clusters. The quality of the obtained clustering solution was assessed using an internal validity metric. The experimental results show that the proposed algorithm can discover interesting patterns and useful information from neuropsychological data, which will be be crucial in planning clinical trials and monitoring clinical performance. The proposed system resulted in an average classification accuracy of 88%, which is very promising considering the unique challenges present in this population.

Toward Developing Robust Myotonic Dystrophy Brain Biomarkers using White Matter Tract Profiles Sub-Band Energy and A Framework of Ensemble Predictive Learning.

The myotonic dystrophies (DM1 and DM2) are dominantly inherited disorders that cause pathological changes throughout the body and the brain. DM patients have difficulties with memory, attention, executive functioning, social cognition, and visuospatial function. Quantifying and understanding diffusion measures along main brain white matter fiber tracts offer a unique opportunity to reveal new insights into DM development and characterization. In this work, a novel supervised system is proposed, which is based on Tract Profiles sub-band energy information. The proposed system utilizes a Bayesian stacked random forest to diagnose, characterize, and predict DM clinical outcomes. The evaluation data consists of fractional anisotropies calculated for twelve major white matter tracts of 96 healthy controls and 62 DM patients. The proposed system discriminates DM vs. control with 86% accuracy, which is significantly higher than previous works. Additionally, it discovered DM brain biomarkers that are accurate and robust and will be helpful in planning clinical trials and monitoring clinical performance.

Diagnosis of Myotonic Dystrophy Based on Resting State fMRI Using Convolutional Neural Networks.

Myotonic dystrophies (DM) are neuromuscular conditions that cause widespread effects throughout the body. There are brain white matter changes on MRI in patients with DM that correlate with neuropsychological functional changes. How these brain alterations causally relate to the presence and severity of cognitive symptoms remains largely unknown. Deep neural networks have significantly improved the performance of image classification of huge datasets. However, its application in brain imaging is limited and not well described, due to the scarcity of labeled training data. In this work, we propose an approach for the diagnosis of DM based on a spatio-temporal deep learning paradigm. The obtained accuracy (73.71%) and sensitivities and specificities showed that the implemented approach based on 4-D convolutional neural networks leads to a compact, discriminative, and fast computing DM-based clinical medical decision support system.Clinical relevance- Many adults with DM experience cognitive and neurological effects impacting their quality of life, and ability to maintain employment. A robust and reliable DM-based clinical decision support system may help reduce the long diagnostic delay common to DM. Furthermore, it can help neurologists better understand the pathophysiology of the disease and analyze effects of new drugs that aim to address the neurological symptoms of DM.

Transparent Electrophysiological Muscle Classification From EMG Signals Using Fuzzy-Based Multiple Instance Learning.

Although a well-established body of literature has examined electrophysiological muscle classification methods and systems, ways to enhance their transparency is still an important challenge and requires further study. In this work, a transparent semi-supervised electrophysiological muscle classification system which uses needle-detected EMG signals to classify muscles as normal, myopathic, or neurogenic is proposed. The electrophysiological muscle classification (EMC) problem is naturally formulated using multiple instance learning (MIL) and needs an adaptation of standard supervised classifiers for the purpose of training and evaluating bags of instances. Here, a novel MIL-based EMC system in which the muscle classifier uses predictions based on motor unit potentials (MUPs) to infer muscle labels is described. This system uses morphological, stability, near fiber and spectral MUP features. Quantitative results obtained from applying the proposed transparent system to four electrophysiologically different groups of muscles, composed of proximal and distal hand and leg muscles, resulted in an average classification accuracy of 95.85%. The findings show the superior and stable performance of the proposed EMC system compared to previous works using other supervised, semi-supervised and unsupervised methods.

Electrophysiological Muscle Classification Using Multiple Instance Learning and Unsupervised Time and Spectral Domain Analysis.

OBJECTIVE: Electrophysiological muscle classification (EMC) is a crucial step in the diagnosis of neuromuscular disorders. Existing quantitative techniques are not sufficiently robust and accurate to be reliably clinically used. Here, EMC is modeled as a multiple instance learning (MIL) problem and a system to infer unsupervised motor unit potential (MUP) labels and create supervised muscle classifications is presented. METHODS: The system has five main steps: MUP representation using morphological, stability, and near fiber parameters as well as spectral features extracted from wavelet coefficients; MUP feature selection using unsupervised Laplacian scores; MUP clustering using neighborhood distance entropy consistency to find representations of MUP normality and abnormality; muscle representation by embedding its MUP cluster associations in a feature vector; and muscle classification using support vector machines or random forests. RESULTS: The evaluation data consist of 63, 83, 93, and 84 sets of MUPs recorded in deltoid, vastus medialis, first dorsal interosseous, and tibialis anterior muscles, respectively. The proposed system discovered representations of normal, myopathic, and neurogenic MUPs for each specific muscle type and resulted in an average classification accuracy of 98%, which is higher than in previous works. CONCLUSION: Modeling EMC as an instance of the MIL solves the traditional problem of characterizing MUPs without full supervision. Furthermore, finding representations of MUP normality and abnormality using morphological, stability, near fiber, and spectral features improve muscle classification. SIGNIFICANCE: The proposed method is able to characterize MUPs with respect to disease categories, with no a priori information.

A Density-Based Clustering Approach to Motor Unit Potential Characterizations to Support Diagnosis of Neuromuscular Disorders.

Electrophysiological muscle classification involves characterization of extracted motor unit potentials (MUPs) followed by the aggregation of these MUP characterizations. Existing techniques consider three classes (i.e., myopathic, neurogenic, and normal) for both MUP characterization and electrophysiological muscle classification. However, diseased-induced MUP changes are continuous in nature, which make it difficult to find distinct boundaries between normal, myopathic, and neurogenic MUPs. Hence, MUP characterization based on more than three classes is better able to represent the various effects of disease. Here, a novel, electrophysio- logical muscle classification system is proposed, which considers a dynamic number of classes for characterizing MUPs. To this end, a clustering algorithm called neighbor- hood distances entropy consistency is proposed to find clusters with arbitrary shapes and densities in an MUP feature space. These clusters represent several concepts of MUP normality and abnormality and are used for MUP characterization instead of the conventional three classes. An examined muscle is then classified by embedding its MUP characterizations in a feature vector fed to an ensemble of support vector machine and nearest neighbor classifiers. For 103 sets of MUPs recorded in tibialis anterior muscles, the proposed system had a 97% electro-physiological muscle classification accuracy, which is significantly higher than in previous works.

Automated segmentation of white matter fiber bundles using diffusion tensor imaging data and a new density based clustering algorithm.

OBJECTIVE: Robust and accurate segmentation of brain white matter (WM) fiber bundles assists in diagnosing and assessing progression or remission of neuropsychiatric diseases such as schizophrenia, autism and depression. Supervised segmentation methods are infeasible in most applications since generating gold standards is too costly. Hence, there is a growing interest in designing unsupervised methods. However, most conventional unsupervised methods require the number of clusters be known in advance which is not possible in most applications. The purpose of this study is to design an unsupervised segmentation algorithm for brain white matter fiber bundles which can automatically segment fiber bundles using intrinsic diffusion tensor imaging data information without considering any prior information or assumption about data distributions. METHODS AND MATERIAL: Here, a new density based clustering algorithm called neighborhood distance entropy consistency (NDEC), is proposed which discovers natural clusters within data by simultaneously utilizing both local and global density information. The performance of NDEC is compared with other state of the art clustering algorithms including chameleon, spectral clustering, DBSCAN and k-means using Johns Hopkins University publicly available diffusion tensor imaging data. RESULTS: The performance of NDEC and other employed clustering algorithms were evaluated using dice ratio as an external evaluation criteria and density based clustering validation (DBCV) index as an internal evaluation metric. Across all employed clustering algorithms, NDEC obtained the highest average dice ratio (0.94) and DBCV value (0.71). CONCLUSIONS: NDEC can find clusters with arbitrary shapes and densities and consequently can be used for WM fiber bundle segmentation where there is no distinct boundary between various bundles. NDEC may also be used as an effective tool in other pattern recognition and medical diagnostic systems in which discovering natural clusters within data is a necessity.

A Multi-Classifier Approach to MUAP Classification for Diagnosis of Neuromuscular Disorders.

The shapes and sounds of isolated motor unit action potentials (MUAPs) in an electromyographic (EMG) signal provide a significant source of information for diagnosis, treatment and management of neuromuscular disorders. These parameters can be analyzed qualitatively by an expert or quantitatively by using pattern recognition techniques. Due to the advantages of quantitative EMG method, developing robust automated MUAP classifiers have been explored and several systems have been developed for this purpose by now, but the accuracy of the existing methods is not high enough to be used in clinical environments. In this paper, a novel classification strategy based on ensemble of support vector machines (SVMs) classifiers in hybrid serial/parallel architecture is proposed to determine the class label (myopathic, neuropathic, or normal) for a given MUAP. The developed system employs both time domain and time-frequency domain features of the MUAPs extracted from an EMG signal using an EMG signal decomposition system. Different classification strategies including single classifier and multiple classifiers with several subsets of features were investigated. Experimental results using a set of real EMG signals showed robust performance of multi-classifier methods proposed here. Of the methods studied, the multi-classifier that uses multiple features sets and a combination of both trainable and nontrainable fusion techniques to aggregate base classifiers showed the best performance with average accuracy of 97% which is significantly higher than the average accuracy of single SVM-based classifier system (i.e., 88%).