Heart Failure (HF) is a prevalent and fatal clinical syndrome that affects the quality of life of millions of people worldwide. Between 17% and 45% of patients suffering from HF die within the first year and the remaining die within 5 years. Furthermore, those patients have a high risk of rehospitalization, their associated healthcare costs are huge, and the higher the life expectancy, the higher the disease's prevalence. HF symptoms commonly include shortness of breath, excessive tiredness, and leg swelling which may be worsened with decompensation, and thus displacement to medical centers represents a handicap for such individuals. Remote monitoring technologies provide a feasible solution that allows earlier decompensation identification and better adherence to lifestyle changes and medication. Although telemonitoring by smartphones showed the potential to reduce both the frequency and the duration of HF hospitalizations, there was no association with the reduction of all-cause mortality. Thus, it indicates there is a need to look for more effective and precise methodologies. In recent years, the use of wearable devices that allow daily monitoring of patient's physiological data combined with Artificial Intelligence (AI) has shown immense potential in predicting cardiovascular-related diseases, their adverse events and patient's health status, including that of patients with HF.
Vitalera has implemented a cloud platform and an alarm-based system for remote monitoring of patients that delivers health alarms when a patient's biomedical measurement is out of a predefined range. The platform relieves clinicians and caretakers of going through each patient's data to check for anomalies, accelerating the decision-making process and reducing hospital consultations. However, the system is creating many straightforward alarms that are finally being discarded after evaluation by the medical professional. In the present project, we propose to run an observational study in order to create a huge dataset with patients' clinical data that will contain annotations regarding the relevance of each alarm. With the annotated dataset we will be able to implement and train Machine Learning (ML) models that will improve the remote monitoring system and its alarm-based system by making it more robust, trustworthy and reliable.
This study is being conducted in the framework of a European project promoted by the European Innovation Council (EIC). An earlier version of the platform was validated in a study conducted in 2020 at Hospital de Torrevieja focused on HF. The rationale for this study is in line with vitalera's goal of incorporating artificial intelligence tools to optimize the digital platform. While this study is focused on the creation of a diverse and labeled dataset and on the development of artificial intelligence event-prediction algorithms, a forthcoming second study will focus on the validation of the algorithms to assess their clinical effectiveness.
This is an observational study involving a European network of hospitals. The study consists of continuous remote patient monitoring using vitalera's digital platform and the supplied devices (tensiometer, wearable, scale and oximeter). For 6 months, a total of 500 patients suffering from HF will have their physiological constants monitored.
Patients will be included in the study based on the eligibility criteria and must complete the informed consent provided. Each hospital will decide when to include their patients according to their particular clinical practice (either in the process of discharge planning or during the first follow-up visit, i.e.. 1 or 2 weeks after discharge). The recruitment period is defined as 6 months. That means patients will be incorporated into the study from its start until the sixth month. The last subject included in the study will then finish the study after one year from the first day of the study. Medical professionals from each hospital will be in charge of recruiting the participants. The recruitment rate is specific for each hospital, and it may vary depending on the month.
There is no power calculation associated with the study since the main objective of the study is to gather a dataset in order to train ML models. Once the algorithms are developed, model performance in terms of accuracy will be evaluated by means of C statistic, the area under the receiver operating characteristic curve, and creation of a calibration plot. Furthermore, the models will be evaluated in terms of fairness and potential bias using metrics including statistical parity, group fairness, equalized odds and predictive equality.
