SenSE: Multimodal Biosensors and Data driven Methods for Explainable Analytics for a Proactive approach to Heart Failure Care

充血性心力衰竭影响着近600万美国人，每年确诊67万例。心力衰竭是美国住院、再入院和死亡的主要原因之一，也是最昂贵的疾病综合征之一。这一高费用的主要部分与在医院处理心力衰竭失代偿发作有关。这些反复住院治疗降低了心力衰竭患者的生活质量，使他们无法过上富有成效和充实的生活。不断上升的成本，越来越多的老年成人慢性疾病需要新的预测，个性化和主动的方法来心血管健康。心力衰竭管理的标准护理依赖于容易观察到的症状，如体重增加和呼吸困难。不幸的是，由于这些症状出现在心力衰竭失代偿过程的后期，住院后才进行干预。在本提案中，我们寻求一种主动的护理方法，该方法由无创多模态传感器系统的创新支持，并与机器学习模型相结合，用于评估心力衰竭失代偿风险，并支持干预措施，以防止心力衰竭患者住院。该项目的目的是(a)设计、制造和验证一个易于使用的传感器贴片，该贴片结合了四种关键模式来评估心脏和肺功能：心电图（ECG）。生物射频（RF）、生物阻抗和地震心动图（SCG） (b)学习潜在变量模型，利用电子健康记录（EHR）的上下文信息将传感器测量与发生失代偿性心力衰竭事件的风险联系起来，以及(c)开发可解释的深度学习模型，将电子健康记录数据与多模态传感器数据结合起来，用于风险预测和指导治疗。传感器贴片的设计将探索新技术，将来自广泛频段的信号集成到单个柔性板上，在功率预算下自主运行。该项目开发的用于ECG、SCG、Bio- RF和Impedance的联合传感器模型将提供与心血管健康相关的非侵入性测量的见解，以前只有植入传感器和导管等侵入性方法才能使用。这些心脏健康的非侵入性测量将用于开发基于学习的数据融合模型，用于推断量化为失代偿风险的潜在健康状态。该项目将产生可解释的深度学习模型，将多模态电子病历数据与多模态传感器数据相结合，以早期发现补偿状态并指导医疗干预。这些模型将考虑到患者数据在时间上的稀疏和非均匀抽样，并采用多模态嵌入的学习来提高可解释性。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Congestive Heart Failure affects nearly six million Americans, with 670,000 diagnosed annually. Heart failure is one of the leading causes of hospital admission and readmission and death in the United States and one of the costliest disease syndromes. A major portion of this high cost of care is related to managing episodes of heart failure decompensation in the hospital. These recurring hospitalizations reduces the quality of life of heart failure patients, preventing them to lead productive and fulfilling lives. Ever rising costs, growing population of aging adults with chronic conditions necessitate new predictive, personalized and proactive approaches to cardiovascular health. The standard care to heart failure management relies on readily observable symptoms such as weight gain and labored breathing. Unfortunately, because these symptoms appear late in the course of heart failure decompensation, intervention is applied after hospitalization. In this proposal we pursue a proactive approach to care supported by innovations in noninvasive multimodal sensor systems paired with machine learning models for assessing the risk of heart failure decompensation and supporting interventions to prevent hospitalization in heart failure patients. The aims of this project are (a) Design, fabrication and validation an easy to use sensor patch that combines four key modalities to assess cardiac and lung function: Electrocardiogram (ECG). Bio Radio Frequency(RF), Bio-Impedance, and Seismocardiogram( SCG) (b) Learning of latent variable models for linking sensor measures to the risk of developing decompensated heart failure events with contextual information from electronic health records (EHR), and (c) Development of interpretable Deep Learning models for combining EHR data with multimodal sensor data for risk prediction and guiding therapy. The design of the sensor patch will explore new techniques integrating signals from a wide range of frequency bands into a single flexible board operating autonomously under a power budget. The joint sensor models developed in this project for ECG, SCG, Bio- RF and Impedance will provide insights into the noninvasive measures related to cardiovascular health previously only available to invasive methods such as implanted sensors and catheterizations. These non-invasive measures of cardiac health will be used to develop a learning based data fusion model for inferring latent health status quantified as decompensation risk. The project will result in interpretable deep learning models for combining multimodal EHR data with multimodal sensor data for early detection of compensated state and guiding medical interventions. These models will account for the sparse and non-uniform sampling of patient data in time, and employ learning of multi-modal embeddings for interpretability.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.