< Back to previous page

Publication

Machine Learning for Human Health: From Activity and Vital-Sign Monitoring to Mortality Prediction

Book - Dissertation

In this dissertation, we aim at handling the problem of uncertainty in medicine and health care. This problem is at the core of clinical practice as clinicians have to make decisions on a daily basis for the good of their patients. Uncertainty in medicine and healthcare has several taxonomies, one of them classify it into three categories: personal, conceptual and technical uncertainty. In this thesis, we focus on technical uncertainty which is data related. Therefore, technical uncertainty is handled as a data analysis problem in which analytical methods (e.g. machine learning) can be used. From a clinical perspective, technical uncertainty comprises three elements: probability, ambiguity and complexity. Probability element reflects the risk and imprecision about future outcomes of a specific decision or action from a temporal perspective. For ambiguity, it reflects the imprecise predictions, conflicting information and lack of evidence. Finally, complexity reflects the multiplicity of influential factors. A possible solution to technical uncertainty in medicine is developing reliable clinical decision support systems (CDSS). The CDSS covers several categories; the most relevant one to our study is the diagnostic decision support system (DDSS). The main components of DDCS are medical data and analytical methods. Medical data can be collected using either conventional or wearable medical devices. Each option has its advantage; for example, conventional medical devices still provide the most accurate and gold-standard measurements. On the other hand, wearable medical devices provide continuous monitoring and ease of installation. For analytical methods, there is a broad spectrum of methods that can be used in the context of human-health applications. Our first candidate amongst these methods is machine learning given its high computational capabilities and efficient error performance. In this thesis, we use simple but powerful machine learning approaches. One approach is intended to be used for online and streaming analytics with wearable devices which is the localised learning approach of k nearest neighbours least squares support vector machine (kNN-LS-SVM). The other method is a linear hard margin approach to support vector machines which is used to engineer explainable features.Developing a reliable decision support system faces many challenges; these challenges are data and model-based. In this thesis, we focus on some of these challenges: class-imbalance, non-uniform data distribution, ambiguity (confusion), continuously increasing data-size, model personalisation, black-box nature, online and streaming analytics. These challenges are linked to the aforementioned elements of technical certainty. Therefore, handling these challenges can restrain the elements of technical uncertainty.This dissertation's main objective is to investigate the possibility to limit and reduce the technical uncertainty of applying machine learning approaches in human health applications. Reducing and restricting this uncertainty can be achieved by developing machine learning algorithms that can enhance the reliability of human health applications' decision-making process.Five human health applications are investigated in this dissertation: human activity recognition, thermal comfort prediction, vital signs prediction, intensive care unit (ICU) mortality prediction, and COVID-like patient mortality prediction at ICU. These applications cover the aforementioned challenges. The human activity recognition (HAR) application focuses on recognising daily life activities and postures based on accelerometer data. The application of thermal comfort prediction focuses on developing a machine learning model to predict the thermal comfort level based on a set of vital signs that can be easily measured using wearable sensors. The vital signs prediction application focuses on estimating the monitored vital signs' early warning score with a relatively high rate (every minute).Moreover, in this application, we develop a machine learning predictive model to predict the monitored vital signs for the upcoming 1, 2, and 3 hours ahead. For these three applications, all measurements are obtained using wearable sensors and the used machine learning algorithm is kNN-LS-SVM for classification and regression. Moreover, these three applications cover the challenges of class-imbalance, non-uniform data distribution, ambiguity, increasing data-size, model personalisation, online and streaming analytics.The fourth application of ICU mortality prediction focuses on engineering explainable features extracted from the monitored vital signs during patients' ICU stay. These engineered features are meant to provide clinical insight to medical staff. In this application, the used machine learning algorithm is a linear hard margin approach to SVM's which used to control the complexity of the model and assure a linear separation of the data points in the input space. The last application focuses on ICU mortality prediction of a specific profile of patients compared to the previous application which investigates multiple profiles together. The fifth application's profile of patients is pulmonary disease patients with infection and/or pneumonia, which experiences similar symptoms of COVID-19 patients. Both fourth and fifth applications are based on ICU data obtained using conventional ICU devices. Moreover, both applications cover the challenge of black-box nature and the resulting lack of interpretability and explainability.
Accessibility:Closed