| Summary: | Shifting healthcare monitoring techniques from laboratory into real-life scenarios has always been very challenging. The current shift towards the use of advanced sensors into everyday objects (e.g., smartwatches) is actively increasing the need for reliable methods and tools to analyse healthcare information acquired in real-life settings for wellbeing applications. In fact, the diffusion of wearable sensors has opened new and unexplored scenarios for Cardiovascular System (CVS) and Autonomic Nervous System (ANS) monitoring in real-life settings. As such, this thesis aims to develop methods and tools to monitor the relationship between CVS and ANS in real-life settings via biomedical signal processing and data-driven machine learning techniques, with the goal of predicting adverse healthcare events and automatically detecting the onset of unhealthy risky situations. Therefore, to investigate the relation between CVS and ANS, electrocardiogram signals and in particular Heart Rate Variability (HRV) were widely investigated in two case studies: acute mental stress detection and prediction of accidental falls in later-life via HRV. One of the main limitations of using wearable sensors for the detection of risky situations in real-life settings is the need to shorten the length of physiological signals below the standard recommendations, which may cause a loss of accuracy in the detection of adverse healthcare events. Therefore, this problem was investigated taking as an exemplar mental stress detection, which is a cogent problem for modern society and it is well-known that mental stress causes alterations in both CVS and ANS. Through a systematic review of the literature, it was demonstrated that little attention has been paid thus far to ultra-short term HRV analysis (i.e., less than 5 minutes) for mental stress detection. Consequently, four experiments were designed and carried out in real-life and in-lab environments to propose a systematic method combining both statistical and machine learning methods to select ultrashort HRV features that are reliable surrogates of 5min HRV features. As a consequence, this study proved that it is possible to automatically detect real mental stress with 1min recordings achieving accuracy rate of 88%. Another limitation of using wearable sensors is the need to improve machine learning techniques to enhance the prediction of rare events. In order to address this, an unbalanced dataset was investigated. In particular, a study was designed to apply data-driven machine learning techniques to an unbalanced dataset of ECG recordings acquired from 170 hypertensive elderly patients, of which 34 experienced an accidental fall. An experimental framework for data-driven machine learning techniques to detect rare events (i.e., falls) was developed to reduce the risk of overfitting problems in unbalanced datasets. This study was the first proving that short term HRV recordings could be used to identify future fallers with high accuracy. This research achieved novel results and significant knowledge advancement for both the investigated well-being and health problems as well as methodological techniques.
|
|---|
