Academic literature on the topic 'PARKINSON'S DISEASE DETECTION'

Le risque de chute provoqué par le phénomène épisodique de ‘Freeze of Gait’ (FoG) est un symptôme commun de la maladie de Parkinson. Cette étude concerne la détection et le diagnostic des épisodes de FoG à l'aide d'un prototype multi-capteurs. La première contribution est l'introduction de nouveaux capteurs (télémètres et goniomètres) dans le dispositif de mesure pour la détection des épisodes de FoG. Nous montrons que l'information supplémentaire obtenue avec ces capteurs améliore les performances de la détection. La seconde contribution met œuvre un algorithme de détection basé sur des réseaux de neurones gaussiens. Les performance de cet algorithme sont discutées et comparées à l'état de l'art. La troisième contribution est développement d'une approche de modélisation probabiliste basée sur les réseaux bayésiens pour diagnostiquer le changement du comportement de marche des patients avant, pendant et après un épisode de FoG. La dernière contribution est l'utilisation de réseaux bayésiens arborescents pour construire un modèle global qui lie plusieurs symptômes de la maladie de Parkinson : les épisodes de FoG, la déformation de l'écriture et de la parole. Pour tester et valider cette étude, des données cliniques ont été obtenues pour des patients atteints de Parkinson. Les performances en détection, classification et diagnostic sont soigneusement étudiées et évaluées<br>Freezing of Gait (FoG) is an episodic phenomenon that is a common symptom of Parkinson's disease (PD). This research is headed toward implementing a detection, diagnosis and correction system that prevents FoG episodes using a multi-sensor device. This particular study aims to detect/diagnose FoG using different machine learning approaches. In this study we validate the choice of integrating multiple sensors to detect FoG with better performance. Our first level of contribution is introducing new types of sensors for the detection of FoG (telemeter and goniometer). An advantage in our work is that due to the inconsistency of FoG events, the extracted features from all sensors are combined using the Principal Component Analysis technique. The second level of contribution is implementing a new detection algorithm in the field of FoG detection, which is the Gaussian Neural Network algorithm. The third level of contribution is developing a probabilistic modeling approach based on Bayesian Belief Networks that is able to diagnosis the behavioral walking change of patients before, during and after a freezing event. Our final level of contribution is utilizing tree-structured Bayesian Networks to build a global model that links and diagnoses multiple Parkinson's disease symptoms such as FoG, handwriting, and speech. To achieve our goals, clinical data are acquired from patients diagnosed with PD. The acquired data are subjected to effective time and frequency feature extraction then introduced to the different detection/diagnosis approaches. The used detection methods are able to detect 100% of the present appearances of FoG episodes. The classification performances of our approaches are studied thoroughly and the accuracy of all methodologies is considered carefully and evaluated

La maladie de Parkinson (MP) est un trouble neurologique causé par une diminution du niveau de dopamine dans le cerveau. Cette maladie est caractérisée par des symptômes moteurs et non moteurs qui s'aggravent avec le temps. Aux stades avancés de la maladie de Parkinson, le diagnostic clinique est clair. Cependant, dans les premiers stades, lorsque les symptômes sont souvent incomplets ou subtils, le diagnostic devient difficile et, parfois, le sujet peut rester non diagnostiqué. En outre, il n'existe pas de méthodes efficaces et fiables permettant d'établir avec certitude un diagnostic précoce du MP. La difficulté de la détection précoce est une forte motivation pour les outils d'évaluation informatisés/outils d'aide à la décision/instruments de test qui peuvent aider au diagnostic précoce et à la prédiction de la progression de la maladie de Parkinson. La détérioration de l'écriture et la déficience vocale peuvent être l'un des premiers indicateurs de l'apparition de la maladie. Selon la documentation examinée, un modèle indépendant du langage pour détecter la maladie de Parkinson à l'aide de signaux multimodaux n'a pas été suffisamment étudié. L'objectif principal de cette thèse est de construire un système multimodal indépendant du langage pour évaluer les troubles moteurs chez les patients atteints de la maladie de Parkinson à un stade précoce, basé sur des signaux combinés d'écriture et de parole, en utilisant des techniques d'apprentissage automatique. Dans ce but et en raison de l'absence d'un ensemble de données multimodales et multilingues, une telle base de données, également répartie entre les témoins et les patients atteints de la maladie de Parkinson, a d'abord été construite. La base de données comprend des enregistrements de l'écriture, de la parole et des mouvements oculaires recueillis auprès de patients témoins et de patients atteints de la maladie de Parkinson en deux phases (avec médication et sans médication). Dans cette thèse, nous nous sommes concentrés sur l'analyse de l'écriture et de la parole, où les patients atteints de la maladie de Parkinson ont été étudiés avec médication.Des modèles indépendants du langage pour la détection de la maladie de Parkinson basés sur les caractéristiques de l'écriture ont été construits ; deux approches ont été envisagées, étudiées et comparées : une approche classique d'extraction et de classification des caractéristiques et une approche d'apprentissage approfondi. Les deux approches ont permis d'atteindre une précision de classification d'environ 97 %. Un classificateur SVM multi-classes pour la détection des étapes basées sur les caractéristiques de l'écriture a été construit. Les performances obtenues n'étaient pas satisfaisantes par rapport aux résultats obtenus pour la détection de MD en raison des nombreux obstacles rencontrés.Un autre ensemble de caractéristiques acoustiques indépendantes de la langue et de la tâche a été construit pour évaluer les troubles moteurs chez les patients atteints de la maladie de Parkinson. Nous avons réussi à construire un modèle SVM indépendant du langage pour le diagnostic de la MP par analyse vocale avec une précision de 97,62%. Enfin, un système multimodal indépendant du langage pour la détection de la maladie de Parkinson en combinant l'écriture et les signaux vocaux a été mis au point, où le modèle SVM classique et les modèles d'apprentissage profond ont tous deux été analysés. Une précision de classification de 100 % est obtenue lorsque des caractéristiques artisanales des deux modalités sont combinées et appliquées au SVM. Malgré les résultats encourageants obtenus, il reste encore du travail à faire avant de mettre notre modèle multimodal de détection de la MP en usage clinique en raison de certaines limitations inhérentes à cette thèse<br>Parkinson’s disease (PD) is a neurological disorder caused by a decreased dopamine level on the brain. This disease is characterized by motor and non-motor symptoms that worsen over time. In advanced stages of PD, clinical diagnosis is clear-cut. However, in the early stages, when the symptoms are often incomplete or subtle, the diagnosis becomes difficult and at times, the subject may remain undiagnosed. Furthermore, there are no efficient and reliable methods capable of achieving PD early diagnosis with certainty. The difficulty in early detection is a strong motivation for computer-based assessment tools/decision support tools/test instruments that can aid in the early diagnosing and predicting the progression of PD.Handwriting’s deterioration and vocal impairment may be ones of the earliest indicators for the onset of the illness. According to the reviewed literature, a language independent model to detect PD using multimodal signals has not been enough addressed. The main goal of this thesis is to build a language independent multimodal system for assessment the motor disorders in PD patients at an early stage based on combined handwriting and speech signals, using machine learning techniques. For this purpose and due to the lack of a multimodal and multilingual dataset, such database that is equally distributed between controls and PD patients was first built. The database includes handwriting, speech, and eye movements’ recordings collected from control and PD patients in two phases (“on-state” and “off-state”). In this thesis we focused on handwriting and speech analysis, where PD patients were studied in their “on-state”.Language-independent models for PD detection based on handwriting features were built; where two approaches were considered, studied and compared: a classical feature extraction and classifier approach and a deep learning approach. Approximately 97% classification accuracy was reached with both approaches. A multi-class SVM classifier for stage detection based on handwriting features was built. The achieved performance was non-satisfactory compared to the results obtained for PD detection due to many obstacles faced.Another language and task-independent acoustic feature set for assessing the motor disorders in PD patients was built. We have succeeded to build a language independent SVM model for PD diagnosis through voice analysis with 97.62% accuracy. Finally, a language independent multimodal system for PD detection by combining handwriting and voice signals was built, where both classical SVM model and deep learning models were both analyzed. A classification accuracy of 100% is obtained when handcrafted features from both modalities are combined and applied to the SVM. Despite the encouraging results obtained, there is still some works to do before putting our PD detection multimodal model into clinical use due to some limitations inherent to this thesis

Le travail présenté dans ce mémoire porte sur le développement d'un système de surveillance ambulatoire pour la détection de la dyskinésie induite par la Levodopa (LID) chez les patients de la maladie de Parkinson (PD). Le système est composé d’unités de mesure inertielle (IMUs) qui recueillent des signaux de mouvement chez des sujets sains et des patients parkinsoniens. Des méthodes différentes sont évaluées pour la détection de LID avec et sans classification des activités. Les données recueillies auprès des sujets sains sont utilisées pour concevoir un classificateur d'activité fiable. Par la suite, un algorithme qui effectue la classification des activités et la détection de la dyskinésie sur les données recueillies auprès de des patients parkinsoniens est proposé. Une nouvelle approche basée sur l'analyse de réseau complexe est également explorée et présente des résultats intéressants. Les méthodes de traitement développées ont été intégrées dans une plateforme complète d’analyse nommée PARADYSE<br>The work presented in this thesis is concerned with the development of an ambulatory monitoring system for the detection of Levodopa Induced Dyskinesia (LID) in Parkinson’s disease (PD) patients. The system is composed of Inertial Measurement Units (IMUs) that collect movement signals from healthy individuals and PD patients. Different methods are evaluated which consist of LID detection with and without activity classification. Data collected from healthy individuals is used to design a reliable activity classifier. Following that, an algorithm that performs activity classification and dyskinesia detection on data collected from PD patients is tested. A new approach based on complex network analysis is also explored and presents interesting results. The evaluated analysis methods are incorporated into a platform PARADYSE in order to further advance the system’s capabilities

Introduction: Parkinson’s Disease (PD) is the second most common neurodegenerative disorder, which affects more than 6 million people worldwide. PD is a progressive multiorgan proteinopathy, also called synucleinopathy, because of the abnormal neuronal accumulation of misfolded alpha-synuclein (αS), an intrinsically disorder protein (IDP) of 140 amino acids. αS is expressed ubiquitously in the cells and has been particularly found associated with membranes in the pre-synaptic terminal, suggesting a role in the regulation of vesicle trafficking, synaptic transmission and synaptic plasticity. Environmental toxins, genetic mutations, amplification of αS gene trigger αS misfolding and aggregation in Lewy Bodies (LBs) and Lewy Neurites (LNs), typical histopathological marker in PD. This event leads to a gradual loss of dopaminergic neurons in substantia nigra pars compacta (SNpc) and the spreading of LB pathology not only in neuronal networks of the central nervous system, but also in the autonomic and peripheral nervous systems. For these reasons, PD manifest with a broad range of motor (e.g. rigidity, tremor and bradykinesia) and non-motor symptoms (e.g. anosmia, constipation, cognitive dysfunction). However, the onset of motor symptoms occurs only when about the 50% of dopaminergic neurons are lost. Therefore, identifying the early step that trigger neuronal toxicity might help in halting neurodegeneration. Misfolding, oligomerization, and fibrillization of αS are thought to be central events in the onset and progression of PD. In fact, αS oligomers acts as seeds for the nucleation dependent process of αS intracellular aggregation, and as a template for the cell-to-cell transmission of αS pathology in a prion like manner. In light of this, we developed Alpha-synuclein FRET-based Biosensors (AFBs) capable to monitor any early αS conformational changes ubiquitously in the cell or in a specific subcellular compartment such as the ER, an early site of αS aggregation, under normal and stress conditions. Methods: AFBs were designed to detect different types of αS assembly such as: - relaxed vs globular αS conformation, where both CFP and YFP are cloned respectively at the C-terminal and N-terminal of the same molecule of WT human αS (intramolecular AFB); - parallel multimeric structures based on N-N and C-C termini interactions between molecules, where both fluorophores (CFP and YFP) are cloned separately at the C-terminal of WT human αS (intermolecular AFBs CC); - antiparallel multimers based on C-N termini interactions, where CFP and YFP are cloned separately and respectively at the C-terminal and N-terminal of WT human αS (intermolecular AFBs CN). AFBs were tested in two different cell lines: SH-SY5Y human neuroblastoma cell line and an inducible cell line stably expressing αS in the ER (iSH-SY5Y). Two different FRET methodologies were used to analyses AFBs behaviour: Sensitized Emission (SE) in SH-SY5Y living cells and Acceptor Photobleaching (AP) in iSH-SY5Y fixed cells. Results: interesting outcomes arose by the analysis of AFBs in iSH-SY5Y cells with AP. AFBs provided a radical different behaviour in the ability of detecting conformational differences in αS structure. Intramolecular AFB revealed a closed conformation of monomeric αS under normal conditions, whereas intermolecular AFBs highlighted a multimerization of αS in the ER, which occurred in an antiparallel orientation of αS monomers in normal condition and with a parallel orientation under the effect of leupeptin, an inhibitor of αS degradation pathway. Discussion and Conclusions: these findings suggested that AFBs represent a valuable strategy to detect conformational changes in αS structures in physiological condition and under the effect of stress stimuli. To the best of our knowledge they are the first fluorescent probes capable of detecting different type of αS multimers in the ER, an early site of pathological αS aggregation. Therefore, AFBs represent an interesting investigative approach on αS multimerization and, in the future, might be employed as a powerful tool for the screening of therapeutic agents which aim to halt αS pathological aggregation from the early phase and to prevent the irreversible progression of PD.

Parkinson’s disease (PD). It is characterized by brief episodes of inability to step, or by extremely short steps that typically occur on gait initiation or on turning while walking. The consequences of FOG are aggravated mobility and higher affinity to falls, which have a direct effect on the quality of life of the individual. There does not exist completely effective pharmacological treatment for the FOG phenomena. However, external stimuli, such as lines on the floor or rhythmic sounds, can focus the attention of a person who experiences a FOG episode and help her initiate gait. The optimal effectiveness in such approach, known as cueing, is achieved through timely activation of a cueing device upon the accurate detection of a FOG episode. Therefore, a robust and accurate FOG detection is the main problem that needs to be solved when developing a suitable assistive technology solution for this specific user group. This thesis proposes the use of activity and spatial context of a person as the means to improve the detection of FOG episodes during monitoring at home. The thesis describes design, algorithm implementation and evaluation of a distributed home system for FOG detection based on multiple cameras and a single inertial gait sensor worn at the waist of the patient. Through detailed observation of collected home data of 17 PD patients, we realized that a novel solution for FOG detection could be achieved by using contextual information of the patient’s position, orientation, basic posture and movement on a semantically annotated two-dimensional (2D) map of the indoor environment. We envisioned the future context-aware system as a network of Microsoft Kinect cameras placed in the patient’s home that interacts with a wearable inertial sensor on the patient (smartphone). Since the hardware platform of the system constitutes from the commercial of-the-shelf hardware, the majority of the system development efforts involved the production of software modules (for position tracking, orientation tracking, activity recognition) that run on top of the middle-ware operating system in the home gateway server. The main component of the system that had to be developed is the Kinect application for tracking the position and height of multiple people, based on the input in the form of 3D point cloud data. Besides position tracking, this software module also provides mapping and semantic annotation of FOG specific zones on the scene in front of the Kinect. One instance of vision tracking application is supposed to run for every Kinect sensor in the system, yielding potentially high number of simultaneous tracks. At any moment, the system has to track one specific person - the patient. To enable tracking of the patient between different non-overlapped cameras in the distributed system, a new re-identification approach based on appearance model learning with one-class Support Vector Machine (SVM) was developed. Evaluation of the re-identification method was conducted on a 16 people dataset in a laboratory environment. Since the patient orientation in the indoor space was recognized as an important part of the context, the system necessitated the ability to estimate the orientation of the person, expressed in the frame of the 2D scene on which the patient is tracked by the camera. We devised method to fuse position tracking information from the vision system and inertial data from the smartphone in order to obtain patient’s 2D pose estimation on the scene map. Additionally, a method for the estimation of the position of the smartphone on the waist of the patient was proposed. Position and orientation estimation accuracy were evaluated on a 12 people dataset. Finally, having available positional, orientation and height information, a new seven-class activity classification was realized using a hierarchical classifier that combines height-based posture classifier with translational and rotational SVM movement classifiers. Each of the SVM movement classifiers and the joint hierarchical classifier were evaluated in the laboratory experiment with 8 healthy persons. The final context-based FOG detection algorithm uses activity information and spatial context information in order to confirm or disprove FOG detected by the current state-of-the-art FOG detection algorithm (which uses only wearable sensor data). A dataset with home data of 3 PD patients was produced using two Kinect cameras and a smartphone in synchronized recording. The new context-based FOG detection algorithm and the wearable-only FOG detection algorithm were both evaluated with the home dataset and their results were compared. The context-based algorithm very positively influences the reduction of false positive detections, which is expressed through achieved higher specificity. In some cases, context-based algorithm also eliminates true positive detections, reducing sensitivity to the lesser extent. The final comparison of the two algorithms on the basis of their sensitivity and specificity, shows the improvement in the overall FOG detection achieved with the new context-aware home system.<br>Esta tesis propone el uso de la actividad y el contexto espacial de una persona como medio para mejorar la detección de episodios de FOG (Freezing of gait) durante el seguimiento en el domicilio. La tesis describe el diseño, implementación de algoritmos y evaluación de un sistema doméstico distribuido para detección de FOG basado en varias cámaras y un único sensor de marcha inercial en la cintura del paciente. Mediante de la observación detallada de los datos caseros recopilados de 17 pacientes con EP, nos dimos cuenta de que se puede lograr una solución novedosa para la detección de FOG mediante el uso de información contextual de la posición del paciente, orientación, postura básica y movimiento anotada semánticamente en un mapa bidimensional (2D) del entorno interior. Imaginamos el futuro sistema de consciencia del contexto como una red de cámaras Microsoft Kinect colocadas en el hogar del paciente, que interactúa con un sensor de inercia portátil en el paciente (teléfono inteligente). Al constituirse la plataforma del sistema a partir de hardware comercial disponible, los esfuerzos de desarrollo consistieron en la producción de módulos de software (para el seguimiento de la posición, orientación seguimiento, reconocimiento de actividad) que se ejecutan en la parte superior del sistema operativo del servidor de puerta de enlace de casa. El componente principal del sistema que tuvo que desarrollarse es la aplicación Kinect para seguimiento de la posición y la altura de varias personas, según la entrada en forma de punto 3D de datos en la nube. Además del seguimiento de posición, este módulo de software también proporciona mapeo y semántica. anotación de zonas específicas de FOG en la escena frente al Kinect. Se supone que una instancia de la aplicación de seguimiento de visión se ejecuta para cada sensor Kinect en el sistema, produciendo un número potencialmente alto de pistas simultáneas. En cualquier momento, el sistema tiene que rastrear a una persona específica - el paciente. Para habilitar el seguimiento del paciente entre diferentes cámaras no superpuestas en el sistema distribuido, se desarrolló un nuevo enfoque de re-identificación basado en el aprendizaje de modelos de apariencia con one-class Suport Vector Machine (SVM). La evaluación del método de re-identificación se realizó con un conjunto de datos de 16 personas en un entorno de laboratorio. Dado que la orientación del paciente en el espacio interior fue reconocida como una parte importante del contexto, el sistema necesitaba la capacidad de estimar la orientación de la persona, expresada en el marco de la escena 2D en la que la cámara sigue al paciente. Diseñamos un método para fusionar la información de seguimiento de posición del sistema de visión y los datos de inercia del smartphone para obtener la estimación de postura 2D del paciente en el mapa de la escena. Además, se propuso un método para la estimación de la posición del Smartphone en la cintura del paciente. La precisión de la estimación de la posición y la orientación se evaluó en un conjunto de datos de 12 personas. Finalmente, al tener disponible información de posición, orientación y altura, se realizó una nueva clasificación de actividad de seven-class utilizando un clasificador jerárquico que combina un clasificador de postura basado en la altura con clasificadores de movimiento SVM traslacional y rotacional. Cada uno de los clasificadores de movimiento SVM y el clasificador jerárquico conjunto se evaluaron en el experimento de laboratorio con 8 personas sanas. El último algoritmo de detección de FOG basado en el contexto utiliza información de actividad e información de texto espacial para confirmar o refutar el FOG detectado por el algoritmo de detección de FOG actual. El algoritmo basado en el contexto influye muy positivamente en la reducción de las detecciones de falsos positivos, que se expresa a través de una mayor especificidad

Gastrointestinal (GI) alpha-synuclein (ASN) detection may represent a clinically useful biomarker of Parkinson's disease (PD), but this has been challenged by conflicting results of recent studies employing different immunohistochemical (IHC) methods and reporting diverse morphological patterns with variable biological interpretation. To increase sensitivity and specificity, we applied three different techniques to detect different possible conformations of ASN in GI tissue derived from biopsies of the GI tract, which were obtained from a longitudinally followed, clinically well-characterized cohort of PD subjects and healthy controls (HC) (Oxford Discovery study). With IHC, we used antibodies reactive for total (T-ASN-Abs), phosphorylated (P-ASN-Abs) and oligomeric (O-ASN-Abs) ASN; with the ASN Proximity Ligation Assay (AS-PLA), we targeted oligomeric ASN species specifically; finally, with the Paraffin-Embedded Tissue Blot (PET-Blot) we aimed to detect fibrillary conformations of ASN specifically. Optimisation and validation of the PET-Blot and PLA techniques was carried out with studies on brain tissue from subjects with ASN pathology, and these experiments were used to gain insight into morphology and distribution of different conformational variants of ASN in the brain of subjects with Lewy pathology. We specified all the detected morphological staining patterns with each technique interpreting them as pathologic or non-specific. Correlation to clinical symptoms was assessed to investigate the potential predictive or diagnostic value of specific staining patterns as biomarkers. A total of 163 GI tissue blocks were collected from 51 PD patients (113 blocks) and 21 healthy controls (50 blocks). In 31 PD patients, GI biopsies had been taken before PD diagnosis (Prodromal PD group); while in 20 PD patients biopsies were obtained after PD diagnosis (Manifest PD group). The majority of these tissues blocks were from large intestine (62%), followed by small intestine (21%), stomach (10%) and oesophagus (7%). With IHC, four ASN staining patterns were detected in GI tissue (Neuritic, Ganglionic, Epithelial, and Cellular), while two distinct staining patterns were detected with AS-PLA (cellular and diffuse signal) and with AS-PET-Blot (ASN-localised and peri-crypt signal). The level of agreement between different techniques was generally low, and no single technique or staining pattern was able to reliably distinguish PD patients (Prodromal or Manifest) from HC. Overall, our study suggests that even specific detection of ASN conformational variants currently considered pathologic was not adequate for the prediction of PD. Future studies with these or other novel techniques focusing on the upper part of the GI tract could overcome current limitations in sensitivity and specificity.

Freezing of gait (FoG) presents as a sudden and brief episode of movement cessation despite the intention to continue walking. As a common symptom of Parkinson's Disease, early detection and quantification of FoG are of great importance in clinical practice. Therefore, this thesis focuses on vision-based and pressure-based FoG detection methods, which were seldom investigated in the past. The task can be treated as a human action recognition problem. Although various deep architectures have achieved encouraging performance for general action recognition, FoG events contain fine-grained patterns, which requires novel deep architectures to learn the domain knowledge. Hence, fine-grained deep architectures are studied, and the major contributions of this thesis are as follows: 1. A graph-based neural network is proposed for vision-based FoG detection by representing a temporal video segment as a directed graph where FoG related candidate regions are the vertices. A weakly-supervised learning strategy is studied to eliminate the resource expensive annotations. 2. A graph sequence recurrent neural network is proposed to formulate long-term graph temporal patterns, which takes graph sequences of dynamic structures as inputs and characterizes FoG patterns by graph recurrent cells. 3. An adversarial spatio-temporal network is proposed to learn FoG patterns across multiple levels using footstep pressure sequences. The adversarial training scheme aims to obtain subject-independent representations to alleviate the issue of high inter-subject variance. 4. A graph fusion neural network is introduced for multimodal learning using footstep pressure maps, video recordings and their associated optical flows. Multimodal graphs are introduced by treating the encoded features of each modality as vertex-level inputs. Comprehensive experiments demonstrate the effectiveness of the proposed methods.

Inertial measurement sensing technology with the capability of capturing disease-relevant data has a great potential for improving the current clinical assessments and enhancing the quality of life in patients with neuro-developmental and neuro-degenerative diseases such as autism spectrum disorders (ASD) and Parkinson's disease (PD). The current clinical assessments can be improved by developing objective tools for the disease diagnosis and continuous monitoring of patients in out of clinical settings. To this end, it is necessary to develop automatic abnormal movement detection methods with the capability of adjusting on new patients' data in real-life settings. However, achieving this goal is challenging mainly because of the inter and intra-subject variability in acquired signals and the lack of labeled data. The research presented in this thesis investigates the application of deep neural networks to address these challenges of abnormal movement detection using inertial measurement unit (IMU) sensors with case studies on stereotypical motor movements in ASD and freezing of gait in PD patients. In this direction, this thesis provides four main contributions: i) A convolutional neural network (CNN) architecture is proposed to learn discriminative features which are sufficiently robust to inter and intra-subject variability. It is further shown how the proposed CNN architecture can be used for parameter transfer learning to enhance the adaptability of the abnormal movement detection system to new data in a longitudinal study. ii) An application of recurrent neural networks and more specifically long short-term memory (LSTM) in combination with CNN is proposed in order to incorporate more the temporal dynamics of IMU signals in the process of feature learning for abnormal movement detection. iii) An ensemble learning approach is proposed to improve the detection accuracy and at the same time to reduce the variance of models. iv) In the normative modeling framework, the problem of abnormal movement detection is redefined in the context of novelty detection and it is shown how a probabilistic denoising autoencoder can be used to learn the distribution of the normal human movements. The resulting deep normative model then is used in a novelty detection setting for unsupervised abnormal movement detection. The experimental results on three benchmark datasets collected from ASD and PD patients illustrate the high potentials of deep learning paradigm to address the crucial challenges toward real-time abnormal movement detection systems using wearable technologies.