Modeling Day-Long ECG Signals to Predict Heart Failure Risk with Explainable AI

TL;DR

The feasibility of deep learning to model 24-hour continuous ECG data, capturing paroxysmal events and circadian variations essential for reliable risk prediction is highlighted, making it a promising tool for HF risk prediction.

Abstract

Heart failure (HF) affects 11.8% of adults aged 65 and older, reducing quality of life and longevity. Preventing HF can reduce morbidity and mortality. We hypothesized that artificial intelligence (AI) applied to 24-hour single-lead electrocardiogram (ECG) data could predict the risk of HF within five years. To research this, the Technion-Leumit Holter ECG (TLHE) dataset, including 69,663 recordings from 47,729 patients, collected over 20 years was used. Our deep learning model, DeepHHF, trained on 24-hour ECG recordings, achieved an area under the receiver operating characteristic curve of 0.80 that outperformed a model using 30-second segments and a clinical score. High-risk individuals identified by DeepHHF had a two-fold chance of hospitalization or death incidents. Explainability analysis showed DeepHHF focused on arrhythmias and heart abnormalities, with key attention between 8 AM and 3 PM. This study highlights the feasibility of deep learning to model 24-hour continuous ECG data, capturing paroxysmal events and circadian variations essential for reliable risk prediction. Artificial intelligence applied to single-lead Holter ECG is non-invasive, inexpensive, and widely accessible, making it a promising tool for HF risk prediction.

Figures (16)

• Figure 1: Perspective clinical scenario for using the DeepHHF score. The patient undergoes a Holter ECG examination for regular indications, such as suspected arrhythmia or syncope. The ECG recording is then processed as an opportunistic analysis by the DeepHHF model and outputs a heart failure (HF) risk score. Patients identified as being at moderate or high risk could be directed toward preventive actions, such as additional screening with brain-type natriuretic peptide (BNP) testing or echocardiography. The icons were created with BioRender.com.
• Figure 1: Study cohort characteristics. Holter examinations are divided into train and test sets and by whether they belong to the positive class (HF) or negative class (non-HF).
• Figure 1: Probability of the comorbidities described in Main Table \ref{['tab1']} divided into the classification groups. A threshold yielding a specificity of 90% was used, correspondingly to the high-risk threshold in Main Figure \ref{['fig_results_model']}c.
• Figure 2: Cohort definition and experimental settings. The flow diagram shows patient inclusion and exclusion criteria, the definition of the two classes with respect to heart failure diagnosis (HF) or lack thereof (non-HF) within five years of the Holter ECG recording, and the partition between training and test sets.
• Figure 2: HF label verification by examining variables of prescribed medications associated with HF treatment. Comparison is between four periods with respect to interval between prescription and diagnosis. The medication groups are detailed in Extended Data Table \ref{['tab_medications']}: Mineralocorticoid receptor antagonists (MRA); Sodium-glucose co-transporter-2 (SGLT2) inhibitors (DapaEmpa); Angiotensin-converting enzyme inhibitors (ACE-I); Angiotensin II receptor blockers (ARB); Angiotensin receptor-neprilysin inhibitors (ARNI); Diuretics (exc. MRA); and Beta-adrenergic blocking agents (BBlockers).