Exame Cardíaco Confiável e Pervasivo com Múltiplos Sensores

The primary aim of this project is to develop a series of innovative AI solutions for assessing the heart's electromechanical activity at point-of-care and providing low-cost, non-invasive methods for continuous monitoring. These solutions will rely on analyzing physiological signals, such as the phonocardiogram, electrocardiogram, and photoplethysmogram, obtained from patients using affordable hardware like modern multimodal digital stethoscopes and wearable devices. The proposed solutions will be capable of detecting general cardiac anomalies and specifically identifying and characterizing two cardiovascular diseases: aortic stenosis and pulmonary hypertension. These diseases were chosen based on their societal impact and the advantages of low-cost screening and monitoring through non-invasive sensing methods. The project's development will focus on the following objectives: 1- Collecting a multimodal dataset of non-invasive cardiovascular measurements in a clinical setting, involving both control and clinical populations. This objective aims to advance the current state-of-the-art in non-invasive cardiac sensing by addressing the lack of large, well-annotated datasets containing synchronous measurements of heart activity from three modalities. Existing datasets typically offer single-modality measurements or, at most, two modalities with limited dimensions. 2- Proposing novel multimodal algorithms capable of effectively incorporating domain knowledge into deep learning approaches. Current data-driven solutions for cardiac signal analysis often require extensive datasets to generalize effectively to new data. However, in cases of limited annotated data, such solutions struggle to be applied practically in real-world clinical environments. This objective seeks to leverage the inherent relationships among the three considered sensing modalities and their sources to explicitly incorporate physiological priors into data-driven methods, thus improving their generalization capabilities. 3- Introducing methods to quantify the reliability of the outputs generated by the developed deep learning models. This objective aims to offer a refined and dependable assessment of the confidence levels associated with the outputs of the proposed deep learning models, by proposing new metrics for quantifying uncertainty in data-driven models, which leverage the available multimodal input. 4- Developing and deploying an early-stage prototype that integrates the proposed AI-based features into a device capable of collecting PCG, ECG, and PPG signals using multimodal information. Existing multimodal stethoscopes on the market do not effectively combine the synchronous collection of PCG, ECG, and PPG signals with multimodal AI solutions for signal processing and clinical information extraction. By taking a comprehensive approach to non-invasive cardiac "imaging" with affordable hardware and cutting-edge deep learning solutions, this project aims to significantly enhance cardiovascular disease screening and monitoring capabilities in both point-of-care and remote patient monitoring scenarios.