Alzheimer's disease has become one of the biggest challenges in the healthcare system worldwide. Researches have shown that Alzheimer’s disease is the sixth leading cause of death in the United States and even the fifth leading cause among people aged 65 and older. Therefore, a screening system that can help the doctor to diagnose Alzheimer’s disease is demanded. In this thesis, we proposed a screening system based on the transcripts of speeches spoken by subjects undertaking a neuropsychology test. While most of the related studies have utilized extracted syntactic and semantic features and relied on a feature selection process, the proposed system used word vectors as the representation of a spoken speech, and Recurrent Neural Networks together with attention mechanism as the classifier. Using ten times 10-fold cross validation on an open dataset with 242 speeches samples spoken by healthy controls and 257 samples spoken by subjects with Alzheimer's disease in the USA, a mean accuracy of 0.83 is achieved in our work. And the classification of 43 healthy subjects and 43 subjects with Mild Cognitive Impairment, the model can still achieve 0.71 of accuracy. On the other hand, validate on 40 Taiwanese subjects with AD and 40 healthy Taiwanese subjects, and 30 Taiwanese subjects with MCI and 30 healthy Taiwanese subjects, accuracy of 0.89 and accuracy of 0.8 can be achieved, respectively.
A Real-time Demand-side Management System Considering User Behavior Using Deep Q-Learning in Home Area Network
In smart grids, demand-side management (DSM) has become an important topic since it can reduce the total electricity cost by smart control and rescheduling of loads, meanwhile, reduce the peak-to-average ratio (PAR) under real-time pricing policy. On the other hand, with the growth of IoT technologies, a smart home can nowadays monitor its household status and control the energy demands; besides, construct their own home area network (HAN) and build the big data database. Thanks to the growing computation ability in recent years, the machine learning skills such as reinforcement learning can be well applied into the DSM problem. However, it is hard to determine a suitable energy management strategy due to the uncertainty of user behavior and the electricity consumption. In the proposed work, a real-time multi-agent deep reinforcement learning based approach has been proposed to solve the DSM problem in HAN, and additionally to consider the user behavior to avoid disturbing user comfort; meanwhile, adaptively learns the appliance usage preference and renew the system day after. The simulation results reveal that the proposed DSM system has improved the energy efficiency in a smart home that not only reduces the electricity cost and peak value but also the PAR value.
A Digital and Automatic Screening System for Alzheimer’s Disease Based on Neuropsychological Test and Neural Network
Alzheimer’s disease (AD) and the other types of dementia have become one of the most serious global health issues and the fifth leading cause of death worldwide nowadays. Therefore, early detection of the disease in the stage of mild cognitive impairment (MCI), which is a prodromal stage of progressing to AD and mild AD, is crucial in order to improve the quality of life of the patients and to decrease the burden of their caregiver and clinicians. The aim of our study is to design a digital screening system based on the Rey-Osterrieth Complex Figure (ROCF) neuropsychological drawing test in order to assist the clinicians to detect whether the subject is MCI or AD against healthy control (HC) automatically. A data-driven deep learning approach is implemented in this work for building the screening system. An architecture of convolutional neural network is designed for pre-training and extracting useful features from the figures drawn by the subjects. The learned features are then transferred to our collected dataset for further training of the classifier in order to distinguish the patients with MCI or AD against HC. As a result, a mean area under the receiver operating characteristic curve score (AUC) of 0.913 is achieved for classifying MCI vs. HC in traditional pencil and paper based ROCF called NTUH_ROCF dataset. On the other hand, dataset that collected using digitalize graphics tablet and smart pen based which is called NTUH_D-ROCF achieved 0.950 of AUC in classifying AD vs. HC.
A Speech Assessment System for Alzheimer’s Disease Based on Neuropsychological Tests Using a Novel Feature Sequence Design and Recurrent Neural Network
In this study, we have proposed a novel Feature Sequence representation for characterizing speech data from patients with Alzheimer’s disease in a neuropsychological test scenario. An Alzheimer Disease Assessment Engine based on a bidirectional recurrent neural network with gated recurrent unit is trained to carry out the classification. Moreover, we have shown that the Feature Sequence can be generated automatically with the help of the Feature Sequence Generator, which is based on a deep convolutional recurrent neural network trained with the connectionist temporal classification loss. Cross validating with a total of 120 samples, which half of them were from the cognitive healthy subjects and the others were from the Alzheimer’s disease subjects, an area under the receiver operating characteristic curve (AUROC) score of 0.838 is achieved.
User-driven Demand-side Management in Mixed-use Community under Dynamic Sharing Price Mechanism using Particle Swarm Optimization
Demand-side management ability in a residential area has been improved a lot under the enforcement of real-time pricing mechanism. In order to utilize the energy efficiency in a residential area and improve the energy quality in whole community. In this work, we aim to integrate the residential users and commercial users into one DSM system. By integrating them, we can utilize the energy more efficiency by sharing the surplus energy in a residential area to the commercial area. The result shows that, in the residential user, the cost is 44.24% than purl residential community on average. In addition, the commercial user also reduces the energy cost. The office user reduces 22.39% and the restaurant user reduces 13.66% on average.
A Supporting System for Quick Dementia Screening Using PIR Motion Sensor in Smart Home
We proposed a supporting system that can quickly estimate the likelihood for an elder of having dementia based on 2 to 4 hours monitoring of a behavioral test done by the elder. During the test, the elder only needs to perform certain activities selected from the Instrumental Activities of Daily Living (IADL) in a smart home environment, and their movement trajectories will be extracted from motion sensors and be analyzed to find potential correlation with the indoor wandering patterns. A machine learning algorithm is selected to carry out the classification task, based on our proposed features of the wandering patterns. Two datasets are employed for performance evaluation, where the first one is 232 elders including 7 dementias, whereas the second one is collected by ourselves from a senior center, which is 30 elders including 9 dementias. It turns out that the average precision and recall for the first dataset are both up to 98.3% with Area Under the ROC Curve (AUC-ROC) being 0.846, and those for the second dataset are 89.9% and 90.0% with AUC-ROC being 0.921.
Nonparametric Activity Recognition System in Smart Homes Based on Heterogeneous Sensor Data
We proposed an activity recognition (AR) system for the elder living independently in smart homes to achieve the concept of “aging in place.” The AR model adopted by the proposed system is powerful to recognize meaningful Activities of Daily Living (ADL) by integrating heterogeneous data from both ambient and on-body sensors. Moreover, the proposed system adopts a nonparametric approach, which requires much fewer efforts from humans. The average AR precision and recall rates of this proposed system are up to 98.7% and 99.0%, which indicates its feasibility of deployment in a real-life home environment for monitoring users' ADL with promising performance.
Demand-side Management in Residential Community Realizing Sharing Economy with Bidirectional Plug-in Electric Vehicle and Renewable Energy
In smart grids, demand-side management is one of the important function since it reduce the total electricity cost of each customer. On the view point of a community, we design a fairness strategy to share Plug-in Electric vehicles' battery with neighbors to reduce the total electricity cost and peak to average ratio. In our problem formulation, each home is assumed to be connected to a renewable energy resource, be equipped with an energy storage device, and have an optional Plug-in Electric vehicles with the vehicle to grid ability, and the formulation is in terms of a multi-objective optimization cooperative game to facilitate power sharing among negihbors. In the results, if Plug-in Electric vehicles come with a lower battery state, system reduce 43.3% of electricity cost.
Contrastenhanced Effective Semisupervised Text Classification with Few Labels
Traditional text classification requires thousands of annotated data or an additional Neural Machine Translation (NMT) system, which are expensive to obtain or access in real applications. In this thesis, we present a ContrastEnhanced Semisupervised Text Classification (CEST) framework under labellimited settings without incorporating any additional NMT system or data augmentation process. First, a certaintydriven method is employed to select appropriate unlabeled data for selftraining. Then, a reliable similarity graph is proposed to induce the smoothness among data instances. Finally, the training is formulated as a“learning from noisy labels＂problem, which is then optimized accordingly. A salient feature of the formulation is that it explicitly suppresses the severe error propagation problem in conventional semisupervised learning. With solely ten labeled data per class, the performance of CEST falls within the 3.5% range of that of the fullysupervised pretrained language models finetuned on thousands of labeled data while outperforming the previous stateoftheart algorithms in the literature by 3.6%, without incorporating any additional systems.
Contrastenhanced Automatic Cognitive Impairment Detection System Embedded with Pause Encoding
As the global elderly population grows annually, healthcare systems face a burden from the rise in Alzheimer’s patients due to its high demand for treatment and early diagnosis. Therefore, research on cognitive impairment screening systems is studied widely to assist doctors in diagnosing Alzheimer’s disease. In this thesis, we propose a contrastenhanced automatic cognitive impairment screening system embedded with paused encoding based on automatic transcription. For cognitive impairment, the pause pattern in speech is a commonly studied acoustic feature that can provide more information based on which the model can make a better distinguishing judgment. Moreover, backtranslation and contrastive learning represent a better contrastenhanced model. After finetuning the transcripts embedded with pause, such a contrastenhanced model is applied to detect the patients’ cognitive impairment. To improve the applicability to the real world, our system is fully automatic, and its generated results can be shown to be explainable. We evaluate our system in two languages, English and Chinese, and both successful results demonstrate the multilingual ability of our work. In terms of quantitative evaluation of our work, our system can achieve 81% accuracy while automatically detecting Alzheimer’s disease on the public ADReSS dataset. Besides, the accuracy of our system in tackling a more challenging task of detecting mild cognitive impairment (MCI), the middle stage between healthy and Alzheimer’s, is highly promising. As for the same task of detecting MCI, on a more unstructured speech dataset, called autobiographical memory dataset collected locally, we show that the accuracy of our system can reach 71% on average.