Periodic Reporting for period 1 - EarlyCare (Unobtrusive Health Monitoring System for the Early Detection of Heart- and Micturition-Related Diseases)
Reporting period: 2021-04-01 to 2023-03-31
EarlyCare is a two-year project focusing on the development and validation of non-wearable, unobtrusive, and privacy-preserving health monitoring systems, namely gas and radiofrequency physiological sensors, to monitor early warning signs of Diabetes Mellitus (DM) and Congestive Heart Failure (CHF) by exploiting advanced time series analysis and explainable artificial intelligence. To achieve this overarching objective, significant advancements have been made in the following research areas:
• The first area of the research for the EarlyCare project was the development of highly sensitive gas sensors to detect urine occurrence and derive the abnormal micturition such as nocturia and pollakiuria as warning signs of both DM and CHF. In this regard, the project proposed an innovative methodology including graphene and metal oxide hybrid nanostructures along with appropriate machine learning (ML) models for the detection of these warning signs.
• The second area of the research focused on the development of deep learning (DL) and ML models to detect abnormal cardiac rhythm and abnormal respiration. The key idea was to investigate the capabilities of newly developed radiofrequency technologies together with DL/ML techniques and heart rate variability time series analysis in the recognition of arrhythmia and abnormal respiration.
• The third area of research dealt with the development of trustworthy and explainable AI (XAI) approach to simulate the detection of DM and CHF via the Belief Rule-Based Expert System.
EarlyCare is not intending to diagnose DM and CHF; it rather proposes unobtrusive solutions to detect the first apparent symptoms that are often overlooked without compromising the comfort and privacy of the users (e,g., older adults living alone).
• The first area of the research for the EarlyCare project was the development of highly sensitive gas sensors to detect urine occurrence and derive the abnormal micturition such as nocturia and pollakiuria as warning signs of both DM and CHF. In this regard, the project proposed an innovative methodology including graphene and metal oxide hybrid nanostructures along with appropriate machine learning (ML) models for the detection of these warning signs.
• The second area of the research focused on the development of deep learning (DL) and ML models to detect abnormal cardiac rhythm and abnormal respiration. The key idea was to investigate the capabilities of newly developed radiofrequency technologies together with DL/ML techniques and heart rate variability time series analysis in the recognition of arrhythmia and abnormal respiration.
• The third area of research dealt with the development of trustworthy and explainable AI (XAI) approach to simulate the detection of DM and CHF via the Belief Rule-Based Expert System.
EarlyCare is not intending to diagnose DM and CHF; it rather proposes unobtrusive solutions to detect the first apparent symptoms that are often overlooked without compromising the comfort and privacy of the users (e,g., older adults living alone).
Towards the detection of abnormal micturition (nocturia and pollakiuria), we have reviewed on the existing approaches to detect and analyze urine samples and the main volatile organic compounds released from urine during emission in order to develop highly sensitive and humidity resilient gas sensing device that will integrate the EarlyCare unobtrusive and privacy-preserving health monitoring system. The largest number of compounds identified in urine belong to the ketone group. Urine covers other chemical classes such as acids, alcohols, aldehydes, amines, N-heterocycles, O-heterocycles, sulfuric compounds and hydrocarbons. This plethora and variety of VOCs in urine is of paramount importance for the development of cross-sensitive gas sensors. Indeed, graphene and metal oxide semiconductors nanohybrids exhibit enhanced sensing performance due to the synergetic effects and complementary characteristics of each type. To build the hybrid (nano)structures, several attempts have been made in the past decade. To the best of our knowledge, this is the first Aerosol Chemical Vapour Deposition technique is investigated for such hybridization through a close collaboration with MINOS research group at Rovira i Virgili University. The morphology of the synthesized films and materials was characterized by field-effect Scanning electron microscopy, Energy Dispersive X-Ray Analysis (EDX) and transmission electron microscopy (TEM), while the gas sensing characterization has been under different temperatures and concentrations of gases and VOCs to study the optimal working temperature, sensitivity and selectivity. The selected gas sensors have been exposed to several urine samples in order to recognise later the micturition activity. In fact, the processing of gas sensors responses with appropriate feature extraction techniques that take into account the oxidizing and reducing behaviour of the sensors, permitted the development of machine learning models (Support Vector Machine) to recognise urine. After optimisation of the location monitoring system-based gas sensors in the toilet and adjustment of the developed model, we concluded that the accuracy of the later depends on the distance between the selected location and toilet bowl that must be considered during the testing of the final prototype.
In the second research area of EarlyCare, we have investigated the capability of three ML/DL techniques (i.e. Convolutional Neural Network (CNN), Random Forest (RF) and SVM) to discriminate between a) CHF, arrhythmia (Arr) and normal sinus rhythm (NSR) and b) arrhythmia and normal sinus rhythm using open accessible datasets. We have explored the original HRV and 9 different derivatives to enrich the characterization of the time series. In addition, we have applied ANOVA feature selection approach to sort the informativeness of the extracted features before applying SVM and RF. With only 20 features out of 198 extracted from 30s RR intervals, RF models have reached 92% and 96% overall accuracy for the prediction of NSR vs. Arr vs. CHF and NSR vs. Arr respectively. The same HRV time series analysis was also very useful to identify the abnormal respiration from the RF 1-min breathing wave.
In the third research area, we have addressed the predictive modeling of DM and CHF through a transparent and explainable AI framework utilizing a Belief Rule-Based Expert System (BRBES). Specifically, we analyzed the simulation of three consecutive days of expected sensor events in order to develop an early prediction system. In order to build the synthetic dataset, two referential and corresponding numerical values of antecedent nocturia and pollakiuria have been suggested for DM while two extra referential and corresponding numerical values of antecedent arrhythmia and dyspnea have been suggested for CHF. Nocturia and pollakiuria were characterized by three distinct referential values: normal, suspicious, and abnormal against two referential values for arrhythmia and dyspnea: positive and negative. 90% overall accuracy has been reached for the prediction of both DM and CHF using the BRBES model.
The EarlyCare health monitoring system that integrated the gas and radiofrequency physiological sensors along with the developed ML/DL models has been tested in Phillips-Marburg University. The successful completion of EarlyCare has paved the way towards further exploitation of the developed technologies in several national and EU research proposals in respiratory health.
In the second research area of EarlyCare, we have investigated the capability of three ML/DL techniques (i.e. Convolutional Neural Network (CNN), Random Forest (RF) and SVM) to discriminate between a) CHF, arrhythmia (Arr) and normal sinus rhythm (NSR) and b) arrhythmia and normal sinus rhythm using open accessible datasets. We have explored the original HRV and 9 different derivatives to enrich the characterization of the time series. In addition, we have applied ANOVA feature selection approach to sort the informativeness of the extracted features before applying SVM and RF. With only 20 features out of 198 extracted from 30s RR intervals, RF models have reached 92% and 96% overall accuracy for the prediction of NSR vs. Arr vs. CHF and NSR vs. Arr respectively. The same HRV time series analysis was also very useful to identify the abnormal respiration from the RF 1-min breathing wave.
In the third research area, we have addressed the predictive modeling of DM and CHF through a transparent and explainable AI framework utilizing a Belief Rule-Based Expert System (BRBES). Specifically, we analyzed the simulation of three consecutive days of expected sensor events in order to develop an early prediction system. In order to build the synthetic dataset, two referential and corresponding numerical values of antecedent nocturia and pollakiuria have been suggested for DM while two extra referential and corresponding numerical values of antecedent arrhythmia and dyspnea have been suggested for CHF. Nocturia and pollakiuria were characterized by three distinct referential values: normal, suspicious, and abnormal against two referential values for arrhythmia and dyspnea: positive and negative. 90% overall accuracy has been reached for the prediction of both DM and CHF using the BRBES model.
The EarlyCare health monitoring system that integrated the gas and radiofrequency physiological sensors along with the developed ML/DL models has been tested in Phillips-Marburg University. The successful completion of EarlyCare has paved the way towards further exploitation of the developed technologies in several national and EU research proposals in respiratory health.
EarlyCare has made a great progress in the field of Health Monitoring Systems by developing and successfully testing unobtrusive and privacy-preserving technologies in real life scenarios. The EarlyCare solution will benefit many stakeholder groups, including older adults, family, health and social care professionals, healthcare institutions, and public expenditures from public authorities and insurance companies/national health systems. The outcomes of EarlyCare project will contribute to 1) advance in the research and industrial development of unobtrusive and non-invasive solutions based on wireless technology for early disease detection; 2) improve resilience of the ageing population and increase longevity, self-sufficiency, quality of life, and healthy ageing; 3) reduce burden of formal and informal caregivers of older adults, minimise time older adults spend in hospitals and medical visits, improve awareness of the impact and benefits of eHealth technologies, and reduce costs of health and care services thanks to an optimised management of resources; and 4) improve health and care delivery, with a higher quality of data and proactivity for symptoms detection.