Machine Learning
New Model Diagnoses Heart Conditions Accurately With 95% Fewer Parameters
Researchers develop BEAT-Net, a biomimetic neural network that mimics cardiologists' reasoning to interpret ECG signals more efficiently and transparently than large foundation models.
A team of researchers has introduced BEAT-Net, a novel approach to automated ECG analysis that combines biological insight with deep learning efficiency. Rather than forcing neural networks to rediscover the structure of heartbeat signals from scratch, the system incorporates domain knowledge directly into its architecture, achieving competitive diagnostic accuracy while using a fraction of the computational overhead required by existing methods.
The Problem with Current Approaches#
Traditional deep learning models treat electrocardiogram recordings as generic time-series data or images, ignoring the specialized physiological knowledge that cardiologists apply. This signal-agnostic approach forces models to waste capacity learning patterns that clinicians already understand, leading to three key limitations: inefficient use of training data, poor performance when evaluated on different datasets, and decision-making processes that don't align with clinical reasoning.
How BEAT-Net Works#
The system is organized into four specialized processing stages that mirror how a cardiologist interprets an ECG:
- QRS Tokenization: The model first identifies and segments heartbeats around the QRS complex, converting continuous signals into meaningful heartbeat units rather than arbitrary time windows
- Morphological Analysis: A Word Encoder extracts shape-based features from individual heartbeats
- Spatial Normalization: A Spatial Operator handles variations across the multiple leads of an ECG
- Temporal and Global Reasoning: Temporal operators and a Transformer-based encoder build understanding of patterns across heartbeats and the full recording
Efficiency and Performance#
When tested on three major datasets (PTB-XL, CPSC2018, and CSN), BEAT-Net achieved an AUC of 0.924—nearly identical to conventional CNN approaches that reached 0.925 AUC. The breakthrough comes in resource efficiency: the model contains only 0.7 million parameters compared to 2.06 million for baseline CNNs, a 95 percent reduction. It also outperformed a 39.5-million-parameter foundation model called HeartLang on morphological classification tasks, reaching 0.901 AUC versus HeartLang's 0.832 AUC.