Prediction of Heart Failure Onset using Multimodal Data Analysis, Deep Learning and Commercial Wearables

Prediction of Heart Failure Onset using Multimodal Data Analysis, Deep
Learning and Commercial Wearables
Project Summary/Abstract
Research: Heart failure is one of the leading causes of mortality and drivers of healthcare costs in the United
States. By 2030, the number of heart failure patients is projected to reach 8 million. If we could predict who will
develop heart failure, this would create an opportunity to improve patient experiences and outcomes by
initiating earlier behavioral and therapeutic interventions. Electronic health records (EHR) contain information
that can be used to predict heart failure before its onset. However, the existing models lead to a large number
of false positive predictions, limiting their clinical utility. The PI proposes to augment the EHR data with
electrocardiogram (ECG) and heart rate variability (HRV) features to improve the accuracy of predicting the
onset of heart failure 12 months in advance. The three modalities of data (EHR, ECG and HRV) will be
analyzed using deep learning methods, including novel techniques proposed by the PI. The models will be
developed and validated retrospectively using patient data available at Michigan Medicine. The second aim of
the proposal is to increase the impact of this research by replacing the clinically measured ECG and HRV with
those obtained by consumer wearables such as smart watches. A prospective cohort of patients will wear a
wearable device for seven days, which will allow the PI to determine whether the collected information
(intermittent ECG, continuous HRV derived from photoplethysmography, and actigraphy), combined with EHR,
can provide clinicians with a more effective tool to identify which patients are at risk of heart failure. While this
approach will benefit a larger population of patients, it will still be limited to those with past medical history. To
further expand the impact of this research to those who wear consumer wearables but have no previous
medical history, a limited model that depends only on the information gathered by the wearable device will be
evaluated. Thus, the outcomes of this study will include multiple models targeting various populations, such as
those with and without prior medical history. Candidate / Career Development: Dr. Sardar Ansari is a computer
scientist and statistician with expertise in biomedical signal processing, machine learning, and medical
wearable devices. His past research experience includes analysis of ECG signal to improve detection of
cardiac arrhythmias and reduce false alarms in intensive care units; detection and removal of noise and motion
artifacts in biomedical signals such as ECG and bioimpedance; prediction of hemodynamic decompensation
using HRV; and detection of hemorrhagic shock, intradialytic hypotension, and low cardiac index using
wearable technology. This award will allow Dr. Ansari to acquire needed additional training in cardiovascular
physiology and heart failure pathophysiology through mentorship, didactic training, attending workshops and
scientific meetings, and clinical exposure, preparing him for an independent career focused on developing
diagnostic and clinical decision support tools for cardiovascular medicine.

Grant Number: 5K01HL155404-05
NIH Institute/Center: NIH
Principal Investigator: Sardar Ansari