Development and Validation of a Machine Learning-based Phenomapping Tool to Identify Acute Decompensated Heart Failure Patients with Diuretic Resistance
Ambarish Pandey, M.D., M.S.C.S. (team leader); Matthew W. Segar, M.D., M.S.; Duwyanne Willett, M.D.; Wilson Tang, M.D.; Javed Butler, M.D., M.P.H., M.B.A.
This machine learning-based semi-supervised sub-phenotyping schema incorporates multiple disease-related factors among patients with acute decompensated heart failure (ADHF) to derive three phenogroups according to their diuretic efficiency. The phenomapping approach has identified key biological factors that underlie diuretic resistance and provide a tool to identify such patients for further phenotyping with proteomic and metabolomic analysis. The phenomapping tool will also help identify patients with a high probability of diuretic resistance for enrollment in clinical trials aimed at testing novel approaches to decongestion and symptom relief among patients with ADHF.
A Deep Learning Classifier for Assessing Diastolic Dysfunction Severity in HFpEF
Partho P Sengupta, M.D., D.M. (team leader); Ambarish Pandey, M.D., M.S.C.S.; Nobuyuki Kagiyama, M.D., Ph.D.; Matthew W. Segar, M.D., M.S.; Naveena Yanamala, Ph.D.
This is a rigorously validated and well-documented semi-supervised deep learning Solution for phenotyping patients with heart failure with preserved ejection fraction (HFpEF). The resulting classifier automatically integrates multidimensional echocardiographic data to grade the severity of diastolic dysfunction and segregates patients into subgroups with predominant cardiac versus extracardiac phenotypic patterns of HFpEF. The deep learning classifier offers a viable solution to overcome the limitations of the existing clinical standards for accurately characterizing the burden of diastolic dysfunction in HFpEF.
Intelligently Characterizing Patient Hemodynamic Phenotypes for Advanced Heart Failure in the ESCAPE Trial Using Learned Multi Valued Decision Diagrams
Josephine Lamp (team leader); Yuxin Wu; Steven Lamp; Lu Feng, Ph.D.; Sula Mazimba, M.D., M.P.H.
This machine learning Solution presents an open source risk stratification and phenotyping tool for hemodynamic phenotypes in advanced heart failure. The tool provides a hemodynamic risk score that takes in single point of care measures as input, and uses a diverse set of features (including invasive and composite hemodynamics and other clinical measures) to return a score of 1 (< 10% chance) to 5 (> 40% chance), indicating the probability of a specified outcome, such as mortality, rehospitalization, or readmission in 30 days.
LEveraging Aggregated Data From Open Sources in Acute Heart Failure (LEAD-AHF)
David Kao, M.D. (team leader); Pardeep Jhund, M.B., Ch.B., Ph.D.; Jennifer L. Hall, Ph.D.
This Solution proposes a harmonized dataset, the Precision Medicine Platform, that will enable researchers to construct virtual patient cohorts that can be used to support a wide range of phenotyping analyses. In accordance with recent calls to improve data sharing, the Solution will be reusable and will support the creation of virtual cohorts and extraction of datasets from acute heart failure clinical studies. The Solution will provide these publicly available datasets in a consistent and transparent manner for secondary analyses.
Application of Machine Learning in Identifying Heterogeneity of Treatment Effects in Patients with Heart Failure and Preserved Ejection Fraction
Rishi J Desai, Ph.D., M.S. (team leader); Sebastian Schneeweiss, M.D., Sc.D.; Muthiah Vaduganathan, M.D., M.P.H.
This Solution proposes to evaluate two novel machine learning-based predictive modeling approaches for grouping the HFpEF populations in order to identify subgroups that are most likely to derive benefit with certain treatments. The goal is to address key limitations of the traditional ‘one variable at a time’ subgroup analysis methods. Successful application of advanced machine learning approaches for identifying factors leading to differential treatment effects in heart failure with preserved ejection fraction (HFpEF) could be crucial in generating hypotheses regarding the role of certain etiologic factors in putative treatment benefits. This could facilitate planning and design of future investigations of targeted treatments for this heterogenous disease.