Ambient assisted living (AAL) technology provides the opportunity for people with disabilities or chronic medical conditions to lead independent lives in their home, relying on wearable sensors and intelligent processing services for the care of their health. The concept of context-awareness is used in AAL to identify current medical conditions and potential abnormalities of a user (patient or elderly). It facilitates the decision-making tasks of health professionals in real-time through remote monitoring, thereby protecting a user from possible health-related risks. The accurate detection of user-specific anomalies, prediction of future events and real-time decision support require intelligent analysis on large biomedical data gathered from many AAL users. This necessitates the need of a scalable system with large storage and high processing capability to support multiple AAL systems simultaneously with individualized context-aware services. The work herein seeks to address such issues and develop a cloud-based context-aware system for AAL that also solves problems regarding clinical abnormalities detections and predictions by discovering personalized knowledge through using context correlations and different data mining techniques. Chronically ill patients die of various diseases from the lack of an efficient automated system having predictive ability. Traditional healthcare solutions are limited to some specific services. While learning models have been developed, they are based on generalized observations and suffer high false alert rates when uncertainties in data increase. Considering these drawbacks this research realizes the need for a self-care, predictive and protective assisted living system where a patient's personalized knowledge is developed by learning from large amounts of historical and contextual data. In an AAL system, a patient is monitored using wearable sensors. These biomedical sensors lack the processing power to perform monitoring, analysis and data-aggregation tasks, necessitating data transmission and computation at central locations. The resource-constrained nature of typical wearable sensors is factored into this research, with cloud computing features utilized to provide a real-time service. Moreover, this study focuses on the development of learning techniques to transform large biomedical data into useful knowledge using a cloud-based framework. This research is intended to explore the machine-learning methods that suit large scale data analyses and exhibit high accuracy and efficiency. We begin by designing a cloud-oriented context-aware middleware (CoCaMAAL) which has the ability to support patients of multiple AAL systems in parallel and independently. Here we introduce how different tasks such as data processing, learning, and context aggregation can be distributed in multiple cloud components by efficiently utilizing cloud resources. We then extend the capability of the model to capture individualized knowledge from big data. We realize that context-aware data of AAL systems satisfy the characteristics of big data (i.e. volume, variety and velocity) and develop a learning model using MapReduce Apriori to discover personalized knowledge by relying on historical big data and context correlations of a patient. The model is further used to find patient-specific abnormalities in the current context, acknowledging that the level of abnormality varies among individuals and in different situations. However, this model lacks the capability of predicting future changes. Therefore in the next step, we build pattern recognition models for predicting health-related changes and behavioural trends in a patient. The proposed techniques are evaluated for different patient scenarios. Moving beyond the feature of future abnormality estimation, we limited our focus only to some clinical event predictions based on the correlations of multiple vital signs (e.g. heart rate, blood pressure) considering the context as an interrelation among vital signs. We target to solve two types of problem: firstly, prediction of patient's future clinical behaviour using only his/her own data, and secondly, prediction of different clinical events of a patient using available known knowledge involving many similar patients. We aim to build these learning models with continuous prediction capability. We have used patient data from publicly available database. All the experiments are conducted using cloud platforms. The purpose of the first problem is to predict the clinical behaviour of a patient in the near future by continuously developing the knowledge from that individual's recent past data. This problem is addressed by developing a learning engine using multi-label classification. The model is patient-specific, adaptive and continuous; it can achieve multiple targets simultaneously. To resolve the second problem, given that a large number of labelled samples from many patients are available beforehand, we developed a probabilistic model using Hidden Markov Model and a static predictor model using data mining algorithms. The probabilistic model predicts approaching clinical events of a patient using temporal behaviour of six bio-signals. This model does not have any forecast margin and uses a few samples for model development. The static predictor model uses a large number of samples for training, varieties of features (e.g. wavelets, correlation coefficients and short-term statistics) and a long forecast gap. However, the prior knowledge about similar clinical conditions may not be always available. Therefore, we developed a new patient clustering method to identify similar patients based on the correlations in clinical data. We developed an unsupervised patient-specific pattern discovery method where multiple vital signs are characterized to a number of dynamic states using entropy criteria. A similarity analysis is then performed over features extracted from dynamic states of multiple patients using hierarchical clustering which can estimate similar patient subgroups. The experimental results indicate that the proposed algorithm can identify different patient subgroups with high accuracy. In a nutshell, this thesis contains problem-specific efficient learning algorithms that can be employed for context-aware sensing in an AAL system by utilizing cloud platforms. This research mostly utilized existing data mining models to solve problems on abnormality detection, prediction and knowledge discovery. The solutions addressed the scalability issues and can work on big biomedical data effectively and accurately. Therefore, the research contributions in this thesis present a scalable model to provide versatile and reliable context-aware services using proper machine-learning models. We believe that this research is a big step towards building a generic model for the AAL community and the results can inspire and be used for the development of efficient learning models for home-based monitoring.
展开▼
