Heart failure (HF) is a major and growing medical and economic problem as 1–2% of the health care budget is spent worldwide on this medical condition. HF affects especially elderly, with 80% of HF related hospitalizations occurring among patients above 65 years. In addition, 70% of the elderly patients are readmitted to hospital within one year, making re-hospitalization an ongoing medical challenge.
In this context, the PerHeart project is employing Information and Communication Technology with the main goal of reducing re-hospitalization rates in HF patients by:
- 1. supporting HF patients in self-managing their disease and adhere to therapy and lifestyle changes (participatory and preventive);
- 2. providing real-time personalized feedback for HF patients and their caregivers (personalized);
- 3. elucidating specific risk factors for readmission and helping health care professionals by revealing new physiological targets or characteristic patient profiles for focused intervention, either medical or social (predictive);
- 4. providing data that can help interpretation and prediction of complex multifactorial diseases (predictive) by taking specific medical, gender, age, social and economic aspects into account (personalized).
The PerHeart ICT platform will integrate in a modular design functionalities dedicated to HF patients and their professional caregivers. The underlying artificial intelligence software will adapt to the patient’s needs while collected data in three countries (Poland, Denmark, Italy) will help elucidate specific risk factors for readmission while taking gender and socio-economic aspects into account and help interpretation and prediction of complex multifactorial disease while also providing input for focused intervention.
Transnational collaboration between multidisciplinary teams from the PerHeart countries, combined with previous experience in developing ambient assistive living solutions, will ensure a successful implementation of the project.
Activity I.1: The PerHeart platform architecture is based on previous developments of the consortium partners who have been active in the Ambient and Assistive Living (AAL) field. A modular design forms the basis of these developments. Within PerHeart we will further enhance the hardware modules and will develop an underlying data processing software which relies on artificial intelligence (AI). The collected data from sensors and devices used by the heart failure patients will be analyzed to extract risk factors and reduce readmission rates. Within this activity, we present the hardware modules and outline the AI solutions which comply with GDPR requirements.
Activity I.2: Within this activity, we have set and started to develop and reconfigure the software components of the platform. Initial tests for an implementation of Federated Learning were also performed. The platform is based on Python and the web framework Flask. The latter is a micro-framework because it is independent on additional libraries. Flask does not require an abstractization layer of the database or validation of other third-party components. The platform will be implemeted based on micro-services. Each micro-service has it own database and a APIs for intercommunication. The APIs allow operations such as Create, Read, Update, Delete for the exchanged data.
Activity I.3: PerHeart webpage is maintained at: www.citst.ro/projects/perheart/. Communication and management will be facilitated by the use of the OpenProj framewrk. This is an open-source desktop project management application similar to Microsoft Project. OpenProj has a easy to use interface and provides Gantt and PERT charts. Openproj allows to create projects and break them up into steps and milestones and each may have its own set of details that allows you to delegate each millstone or step as tasks for teams.
Activity II.1 - Design of the platform architecture: The architecture of the PerHeart platform is based on the previous development of the partners involved in Ambient and Assistive Living (AAL) projects. They are based on a modular design that will be complemented within PerHeart with software based on artificial intelligence (AI). This will help the data collected by the sensors and devices integrated in the platform to be analyzed in order to establish the risk factors that lead to the readmission of patients with heart failure. The modules that go into the components of the PerHeart platform are presented along with the software components selected so far.
Activity II.2 - Platform integration: In this activity were developed and integrated a first version of the medicine box, user interface (for purchasing medical parameters and saving them in the database), as well as the stability of calendar functionality and identifying an open-source solution, which will be adapted and integrated in the next stage.
Activity II.3 - Development of the AI module: In this activity the AI module based on federated learning was developed. The solution proposed in the PerHeart system is a generalization (or can be seen as an extension) of the basic federated learning algorithm. This involves a grouping of customers, depending on the data distribution, to perform the training distributed on data that are IID (distributed identically and independently) within each group. Thus, convergence can be faster and, in addition, accuracy can be improved.
Activity II.4 - Laboratory testing. Implementation optimization: In this activity the AI module was evaluated using public data sets: CIFAR-10, MNIST and HAPT. The accuracy obtained on the FedAvg algorithm, after 500 epochs, is approx. 80%, and in the CIFAR-10 experiment on IID data, the accuracy is 17.5% lower. On the other hand, CIFAR10-nonIID-2 got much closer to FedAvg. This may indicate the potential of the proposed model. In the case of HAPT-4, the results obtained are comparable to those of FedHealth, especially considering that no additional development or training is done for customization. By adding regularization, accuracy can be increased.
Activity II.5 - End-user classification: This activity outlined the inclusion and exclusion criteria that were taken into account when selecting end-users for pilot tests. The classification was made based on the cognitive and functional assessment of the use of the Katz Index of Independence in Activities of Daily Living (ADL) and the Mini-Mental State Examination (MMSE).
Activity II.6 - Dissemination: Conferences, workshops: Within this activity, projects were presented within Be Health 2021 and one paper was published at an international conference with ISI indexed procedures. The project page is maintained in English and Romanian.
Activity III.1 - Development of the AI module: In this activity the FedHealth framework was used on a non-standard sleep dataset which is composed of data from multiple sources. The dataset was built using the ISRUC-SLEEP dataset as well as data taken from different sensors such as EmfitQS (3 users), Oura Ring (one user) and Samsung Galaxy Watch Active (one user). From this dataset, pulse values (being common to all data) measured during sleep were used. A neural network similar to U-Net was used to learn the patterns in the pulse evolution of the users. To learn the sleep pattern, we also used transposed convolution layers to help with the reconstruction of the patterns. The developed model was subjected to a series of tests that demonstrated that it is capable of recognising users' sleep patterns. Furthermore, it was shown that using U-Net and a federated learning framework it is possible to build a personalized model based even from a non-standard dataset.
Activity III.2 - Laboratory testing. Implementation optimization: In this activity we have continued to improve the the smart pillbox developed by CITST in the PerHeart project. We have extended the user interface by developing the application associated with the pillbox. We have also developed and integrated a personal calendar which allows the insertion and listing of the user’s events. The pillbox is modular, with detachable compartments for easy cleaning and replacement. Each compartment is illuminated with a LED that lights up to indicate to the user which compartment must be used. The pillbox is designed and 3D printed. The number of compartments as well as their size can be adjusted according to user’s preferences which will be assessed during pilot tests.
Activity III.3 - Technical aspects for the preparation and implementation of the pilot studies: In order to prepare and implement the pilot studies, several sets of the PerHeart platform have been prepared, manuals for pairing and using medical devices have been produced as well as a presentation to show users how to use the PerHeart platform interface. Several replicas of the PerHeart platform consisting of a dedicated Android tablet on which the PerHeart software and related devices (glucometer, blood pressure monitor, scale, Xiaomi bracelet, medicine box, etc) were installed and tested. Data acquisition and transmission for health monitoring is done through a software application installed on the user's tablet. Therefore, the health monitoring devices in a set have to be connected via Bluetooth with the users’ tablets. This pairing is likely to be lost during the pilot studies and, the user manuals also contain the necessary instructions to recreate the pairing.
Activity III.4 - Development of local and common databases: The Perheart_db database is developed in MongoDB. Three local databases were developed in this phase (one for each country where the pilot will be carried out: Romania, Denmark and Poland). Each local database consists of the following entries: users IDs, health_data, calendar_data, mi_band_data, pillbox_data, pillbox_compartments. The anonymized data from all users are collected in a common database with the same structure as the local ones. This is used to run AI algorithms and create models that have high accuracy (minimum 90%). Models created using the common database can be transferred and used in local databases. In this way a better performance of the created models will be achieved.
Activity III.5 - Mining tool development and data analysis: In cadrul acestei activitati au fost prezentate rezultatele proiectului in cadrul mai multor evenimente de diseminare precum Be Health 2022 si au fost publicate 4 lucrari la conferinte internationale cu proceedings in curs de indexare ISI. Proiectul este diseminat prin mediile profesionale si prin pagina web mentiute de catre partenerul CITST in limbile engleza si romana.
Activity III.6 - Dissemination: conferences, workshops, exhibitions: Under this activity we have presented the project results at several dissemination events such as Be Health 2022. In addition, 4 papers were published at international conferences with proceedings undergoing ISI indexing. The project is disseminated through professional media and through the webpage maintained by the CITST partner in English and Romanian.