Digital healthcare monitoring uses multidisciplinary sensing techniques to track diverse human data and behaviors. Machine learning can promote an individual's well-being through more efficient and accurate health status monitoring. However, challenges hinder precise monitoring, such as privacy concerns, varied subjects, diverse sensors, and different objectives. To help address these challenges, this thesis explores projects spanning various healthcare domains. Explainable and robust machine-learning solutions are proposed and tested, which include novel signal processing guidelines, innovative feature engineering methods, and pioneering deep-learning networks. These solutions contribute to the state-of-the-art in their respective healthcare domains. The first project addressed the challenge of assessing fall risk among individuals with varying levels of mobility using inertial sensors. Machine-learning models were developed and evaluated using datasets from stroke survivors and community-dwelling elders with participants of varying levels of mobility. Risk indicators were obtained through kinematics simplification that are both explainable and modifiable. These indicators considerably enhance fall risk classification performance compared to existing approaches and the conclusions align with available biomechanical evidence. In the second project, a new machine-learning architecture was created for fall detection and classification using multistatic radar sensing. This new approach (called eMSFRNet) solved the common problem of weak and varied Doppler signatures caused by line-of-sight restrictions. It is the first method that can classify among fall types using radar sensing, and yielded state-of-the-art accuracy for both fall detection (99.3%) and seven fall types classification (76.8%) tasks. In the third project, a novel combination of signal processing and a machine learning framework (named MIND) was designed to detect and forecast motor restricted and repetitive behaviors (RRBs) among children with autism spectrum disorder (ASD), using data from multiple wearable sensors. Contrary to prior beliefs that such detection or forecasting was unattainable, the novel MIND AI framework offers a comprehensive and generalizable approach. Transition behaviors were first defined and then identified, suggesting the potential to detect behavioral shifts preceding motor RRBs. The new signal monitoring quantification (MQ) guidelines minimize the impacts of inconsistent data caused by individualized sensor placements. MIND achieved 100% accuracy in detecting motor RRBs on new subjects with unfamiliar behavior types and 92.2% accuracy in forecasting motor RRBs. In conclusion, the work in this thesis showcases the pivotal contributions of robust and explainable machine learning solutions tailored for specific healthcare challenges. These contributions either solve longstanding problems in different healthcare fields or guide new research directions. The new methodologies introduced – including the MQ guidelines, modifiable fall risk indicators, and innovative deep learning models – all help to advance healthcare machine learning applications by merging accuracy with explainability. / Doctor of Philosophy / Digital healthcare monitoring uses advanced techniques to monitor a person's health and behavior. With the help of machine learning (think of this as teaching computers to think and learn), it is possible to improve health monitoring to be faster and more accurate. Still, there are important challenges to overcome, including concerns regarding personal privacy, the variety of ways in which data can be collected, and the diverse goals of each monitoring tool. This research addressed these challenges by creating and evaluating new machine learning methods for application to multiple healthcare areas. New, understandable, and powerful machine learning methods were developed, pushing the boundaries of how best to use varied technologies for monitoring. A few highlights of the research include the following. First, a method was developed to better determine if an older adult is at a higher risk of falling. The ability of the method to estimate falling risk was very strong, and superior to previously-reported methods. This new method can also explain why an individual might be at a higher risk of falling and offers suggestions on how to walk more stably. Second, a technique was created to process radar signals to detect falls and to determine the type of fall that occurred. This technique solves a long-standing problem with radar, specifically that this sensing technology often provides unclear and unstable signals. Third, a machine-learning method was constructed to identify repetitive (self-injurious) behaviors among children with autism spectrum disorder using signals from wearable sensors. This novel method can detect behaviors quite accurately, even in challenging scenarios. One notable finding was changes in normal behavior can be identified shortly before the repetitive behaviors occur. Overall, this research contributes substantially new and effective methods for healthcare and understandable machine learning solutions. These contributions help to solve challenging, ongoing problems and pave the way for future innovations. Methods such as those developed promise a future where technology can better assist in healthcare, making it more precise and understandable for everyone.

There is a sharp increase in the number of old people living alone throughout the world. More often than not, such people require continuous and immediate care and attention in their everyday lives, hence the need for round the clock monitoring, albeit in a respectful, dignified and non-intrusive way. For example, continuous care is required when they become frail and less active, and immediate attention is required when they fall or remain in the same position for a long time. To this extent, various monitoring technologies have been developed, yet there are major improvements still to be realised. Current technologies include indoor positioning systems (IPSs) and health monitoring systems. The former relies on defined configurations of various sensors to capture a person's position within a given space in real-time. The functionality of the sensors varies depending on receiving appropriate data using WiFi, radio frequency identification (RFIO), ultrawide band (UWB), dead reckoning (OR), infrared indoor (IR), Bluetooth (BLE), acoustic signal, visible light detection, and sound signal monitoring. The systems use various algorithms to capture proximity, location detection, time of arrival, time difference of arrival angle, and received signal strength data. Health monitoring technologies capture important health data using accelerometers and gyroscope sensors. In some studies, audio fingerprinting has been used to detect indoor environment sound variation and have largely been based on recognising TV sound and songs. This has been achieved using various staging methods, including pre-processing, framing, windowing, time/frequency domain feature extraction, and post-processing. Time/frequency domain feature extraction tools used include Fourier Transforms (FTs}, Modified Discrete Cosine Transform (MDCT}, Principal Component Analysis (PCA), Mel-Frequency Cepstrum Coefficients (MFCCs), Constant Q Transform (CQT}, Local Energy centroid (LEC), and Wavelet transform. Artificial intelligence (Al) and probabilistic algorithms have also been used in IPSs to classify and predict different activities, with interesting applications in healthcare monitoring. Several tools have been applied in IPSs and audio fingerprinting. They include Radial Basis Kernel (RBF), Support Vector Machine (SVM), Decision Trees (DTs), Hidden Markov Models (HMMs), Na'ive Bayes (NB), Gaussian Mixture Modelling (GMM), Clustering algorithms, Artificial Neural Networks (ANNs), and Deep Learning (DL). Despite all these attempts, there is still a major gap for a completely non-intrusive system capable of monitoring what an elderly person living alone is doing, where and for how long, and providing a quick traffic-like risk score prompting, therefore immediate action or otherwise. In this thesis, a cost-effective and completely non-intrusive indoor positioning and activity-monitoring system for elderly people living alone has been developed, tested and validated in a typical residential living space. The proposed system works based on five phases: (1)Set-up phase that defines the typical activities of daily living (TADLs). (2)Configuration phase that optimises the implementation of the required sensors in exemplar flat No.1. (3)Learning phase whereby sounds and position data of the TADLs are collected and stored in a fingerprint reference data set. (4)Listening phase whereby real-time data is collected and compared against the reference data set to provide information as to what a person is doing, when, and for how long. (5)Alert phase whereby a health frailty score varying between O unwell to 10 healthy is generated in real-time. Two typical but different residential flats (referred to here are Flats No.1 and 2) are used in the study. The system is implemented in the bathroom, living room, and bedroom of flat No.1, which includes various floor types (carpet, tiles, laminate) to distinguish between various sounds generated upon walking on such floors. The data captured during the Learning Phase yields the reference data set and includes position and sound fingerprints. The latter is generated from tests of recording a specific TADL, thus providing time and frequency-based extracted features, frequency peak magnitude (FPM), Zero Crossing Rate (ZCR), and Root Mean Square Error (RMSE). The former is generated from distance measurement. The sampling rate of the recorded sound is 44.1kHz. Fast Fourier Transform (FFT) is applied on 0.1 seconds intervals of the recorded sound with minimisation of the spectral leakage using the Hamming window. The frequency peaks are detected from the spectrogram matrices to get the most appropriate FPM between the reference and sample data. The position detection of the monitored person is based on the distance between that captured from the learning and listening phases of the system in real-time. A typical furnished one-bedroom flat (flat No.2) is used to validate the system. The topologies and floorings of flats No.1 and No.2 are different. The validation is applied based on "happy" and "unusual" but typical behaviours. Happy ones include typical TADLs of a healthy elderly person living alone with a risk metric higher than 8. Unusual one's mimic acute or chronic activities (or lack thereof), for example, falling and remaining on the floor, or staying in bed for long periods, i.e., scenarios when an elderly person may be in a compromised situation which is detected by a sudden drop of the risk metric (lower than 4) in real-time. Machine learning classification algorithms are used to identify the location, activity, and time interval in real-time, with a promising early performance of 94% in detecting the right activity and the right room at the right time.