Detecting cardiac pathologies via machine learning on heart-rate variability time series and related markers

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Cardiac pathologies significantly impact patient health and require accurate diagnostic tools.
• Analyzing heart rate variability from long-term recordings offers insights into cardiac function.
• Machine learning approaches show promise in interpreting complex physiological data for disease detection.

Purpose of the Study:

• To develop and validate statistical algorithms for inferring cardiac pathologies from 24-hour Holter recordings.
• To compare two distinct algorithmic approaches for classifying cardiac health status.
• To achieve high accuracy in distinguishing healthy patients from those with specific cardiac conditions.

Main Methods:

• Statistical analysis of heart-beat series to extract 49 established heart variability markers.
• Application of principal component analysis for marker selection and data augmentation.
• Training a multi-layer feed-forward neural network for patient classification.
• Construction and analysis of patient similarity networks based on heart-beat series.

Main Results:

• The neural network model achieved up to ~85% accuracy in classifying patients with atrial fibrillation or congestive heart failure.
• Network analysis revealed distinct emergent properties capable of discriminating between healthy and pathological patient groups.
• Both algorithmic approaches demonstrated strong agreement in their diagnostic capabilities.

Conclusions:

• Statistical algorithms and machine learning, applied to 24-hour Holter data, provide a robust method for diagnosing cardiac pathologies.
• The study validates the effectiveness of analyzing heart rate variability markers and patient similarity networks for cardiac health assessment.
• These findings support the potential of data-driven approaches in improving cardiovascular diagnostics.