Developing a platform for deep phenotyping of heart failure with preserved ejection fraction using raw, widely-available, multi-modality data and artificial intelligence algorithms

项目摘要/摘要

PROJECT SUMMARY / ABSTRACT The pathophysiologic heterogeneity underlying heart failure with preserved ejection fraction (HFpEF) is poorly- understood and is a major barrier to effective HFpEF treatments, necessitating bold new approaches to HFpEF phenotyping. The long-term goal is to reduce the substantial morbidity and mortality of HFpEF by enabling better detection, understanding and treatment of its phenotypic subtypes. The overall objectives of this application are to (i) develop machine learning algorithms that can sequentially detect HFpEF then identify HFpEF phenotypes using widely-available data; then (ii) validate this detection-phenotyping approach in a large cross-University of California (UC) cohort. The central hypothesis is that machine learning can algorithmically extract physiologically- valuable information from raw multi-modality data to phenotype HFpEF. The rationale is that algorithms to reliably phenotype HFpEF will provide both the framework to investigate phenotype-specific mechanisms and therapies, and the method by which to identify target patients. The first aim will develop algorithms that can reliably detect HFpEF using widely-available electrocardiogram (ECG) data. Neural network algorithms will be trained using ECG data to discriminate HFpEF from heart failure with reduced ejection fraction and patients without heart failure. For the second aim, a novel machine learning architecture will be developed to extract maximal information from multiple diagnostic modalities simultaneously. This architecture will then be used to train algorithms to identify and phenotype HFpEF with widely-available data: ECGs, echocardiograms (echo) and specific electronic health record (EHR) data elements. Once reproducible HFpEF phenotypes are identified using our multi-modal neural network phenogrouping approach, we will characterize physiologic differences between identified phenotypes. The third aim will construct a cross-UC heart failure/HFpEF cohort to externally validate these multi-modal HFpEF algorithms and the identified HFpEF phenotypes. The cross-UC heart failure/HFpEF cohort will be updated regularly and designed to support future prospective multi-center studies. The research proposed in this application is innovative, in the applicant’s opinion, because it develops a novel algorithmic approach to extract maximal information from widely-available data in multiple modalities simultaneously, to more closely mimic how physicians triangulate information to make diagnoses. The proposed research is significant because applying this algorithmic approach to HFpEF is expected to provide a critical phenotypic framework, through which current and future HFpEF therapies can be tested and administered, and which will also support future investigations into underlying disease mechanisms. Ultimately, establishment of reproducible HFpEF phenotypes, and the ability to identify them with widely-available data, would dramatically shift the management and research paradigms in HFpEF, enabling the targeting of phenotype-guided therapies in a precision medicine approach.