📄 medicine

Clinical Evidence Morphology--Rhythm for ECG (CEMR-ECG): a traceable evidence framework for imbalanced heartbeat classification

The paper introduces CEMR-ECG, a traceable decision-layer framework that integrates rhythm, morphology, and prototype evidence with backbone classifiers to significantly improve macro-averaged F1 scores and address class imbalance in AAMI-style ECG heartbeat classification across multiple datasets.

Original authors: Chen Xu

Published 2026-07-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chen Xu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Needle in a Haystack"

Imagine you are hiring a team of detectives (the computer models) to find rare, dangerous criminals (abnormal heartbeats) in a city of millions of innocent citizens (normal heartbeats).

In the world of heart monitoring, normal beats are the citizens, and abnormal beats (like those from a heart attack or irregular rhythm) are the rare criminals. Because there are so many more normal beats, standard computer models get lazy. They just guess "Normal" for almost everything. This gives them a high "accuracy" score (they are right 99% of the time), but they miss the dangerous criminals entirely.

Furthermore, when these models do make a guess, they just give a final score like "85% chance this is a crime." They don't explain why. If a doctor sees a high score, they don't know if the computer saw a specific shape, a weird rhythm, or just guessed.

The Solution: The "Expert Consultant" (CEMR-ECG)

The authors created a new system called CEMR-ECG. They didn't try to replace the detectives (the underlying computer models). Instead, they added a specialized consultant who sits between the detective and the final verdict.

Think of the original computer model as a Generalist Detective. It looks at a heartbeat and says, "I think this is a Ventricular Ectopic beat."

The CEMR-ECG Consultant is a Forensic Specialist. Before the final verdict is written, this specialist checks five specific, named pieces of evidence:

The Rhythm: How far apart are the heartbeats? (Like checking the timing of a clock).
The Shape: What does the heartbeat wave look like? (Like checking a fingerprint).
The Comparison: Does this lead (wire) look different from the other lead?
The Change: How fast is the wave changing? (Like checking the speed of a car).
The Prototype: Does this look like a known "criminal" pattern we have seen before?

How It Works: The "Safety-Checked" Decision

Once the Generalist Detective and the Forensic Specialist have their opinions, CEMR-ECG combines them using a special rulebook called BioAdaptive Decoding.

Here is the clever part: The system doesn't just blindly trust the specialist. It has a Safety Guard.

If the data is strong and clear, the system lets the specialist's evidence boost the score for the rare beats.
If the data is weak or the rare beat is extremely rare (like finding only 23 examples in a whole database), the Safety Guard says, "Hold on, we don't have enough proof to change the verdict." It prevents the system from making wild guesses just to look good.

This ensures the system is traceable. If it flags a beat as dangerous, you can look at the report and say, "Ah, it flagged this because the Rhythm was off and the Shape matched a known pattern," rather than just seeing a mysterious black-box number.

The Results: Catching More Criminals Without False Alarms

The authors tested this "Consultant" on 20 different types of computer models (from simple math models to complex AI) using three famous heart databases.

The Win: In almost every single test, adding the CEMR-ECG consultant helped the models find more of the rare, dangerous beats.
The Score: The ability to correctly identify all types of beats (including the rare ones) jumped by about 11.5 percentage points.
The Balance: It didn't just find more rare beats by making a mess of normal ones. It actually improved the detection of the rare "Supraventricular" and "Ventricular" beats while keeping the "Normal" beats correct.

One Catch: The system is smart about what it can handle. In one database where there were only 23 rare "Fusion" beats, the system correctly said, "I can't reliably find these because there aren't enough examples to learn from." It didn't force a guess; it stayed conservative.

What This Paper Does Not Claim

It is important to stick to what the paper actually says:

It is not a new heart monitor: It is a software layer that improves existing computer models.
It is not ready for patient care yet: The authors explicitly state this is a "retrospective" study (looking at old, public data). They say it cannot be used on real patients right now. It needs more testing in real-world hospitals and with real doctors before it can be used to make medical decisions.
It doesn't replace the doctor: It is a tool to help computers make better, more explainable guesses, but the final decision for patient care requires human validation.

Summary

Think of CEMR-ECG as a smart filter you can attach to any heart-beat computer program. It forces the computer to look at specific, named clues (like rhythm and shape) before making a final call. This makes the computer much better at spotting rare, dangerous heartbeats and explains why it made that choice, all while having a safety switch to prevent it from guessing when the evidence is too weak.

Technical Summary: Clinical Evidence Morphology–Rhythm for ECG (CEMR-ECG)

Problem Statement
Automatic electrocardiogram (ECG) heartbeat classifiers often struggle with class imbalance, where normal beats dominate the distribution, leading to missed rare beats (supraventricular, ventricular, and fusion). Furthermore, standard deep learning models often provide limited decision evidence beyond final scores, making it difficult to interpret why a specific beat was classified as such. Existing approaches typically focus on improving the waveform encoder, reshaping training data, altering loss functions, or providing post-hoc explanations. There is a gap for a traceable decision-layer framework that inserts fixed, named clinical evidence between a backbone's probability output and the final classification, specifically designed to recover rare classes without compromising the interpretability of the decision process.

Methodology
The authors propose CEMR-ECG, a traceable evidence framework that operates as a decision-layer interface rather than replacing the underlying waveform backbone. The framework is designed to be compatible with any AAMI-style heartbeat classifier that outputs a probability vector over the standard classes: Normal (N), Supraventricular ectopic (S), Ventricular ectopic (V), Fusion (F), and Unknown (Q).

Evidence Representation:
For each annotated beat, a fixed 362-dimensional evidence vector ( $z_i$ ) is constructed from five specific clinical evidence groups:
- RR Rhythm: Pre- and post-R–R intervals, local/global medians, and ratios.
- Morphology: 26 local morphology summaries per retained lead.
- Inter-lead Waveform Differences: Pairwise differences between two retained leads.
- Shape-Derivative Summaries: Resampled waveforms and first-difference traces.
- Prototype Evidence: Cosine similarity and distance metrics against class prototypes (computed from the training split).
Architecture and Fusion:
- Backbone Interface: A candidate backbone (machine learning, ECG-specific deep learning, or general time-series models) provides a raw probability vector ( $p^B_i$ ).
- Evidence Encoder: A shared Extremely Randomized Trees encoder maps the evidence vector $z_i$ to an evidence probability vector ( $p^E_i$ ).
- Adapter: The framework forms a normalized weighted geometric pool of the backbone and evidence probabilities using a parameter $\alpha$ (selected from a grid via internal validation).
- BioAdaptive Decoder: A bounded decoder applies class multipliers derived from training prevalence (capped to prevent over-inflation of minority recovery) and a power parameter $\gamma$ . Six candidate policies (e.g., S-recovery, F-recovery, conservative-default) are evaluated.
- Selection Logic: The adapter weight ( $\alpha$ ) and decoder policy ( $k$ ) are selected based on internal validation performance (maximizing M-F1(4) with tie-breaking). Crucially, "guard conditions" prevent the decoder from applying aggressive recovery if validation support is insufficient (e.g., if F-class gain is based on fewer than 10 validation beats) or if accuracy drops significantly.
Experimental Design:
The study evaluated 20 diverse backbones (11 machine learning, 4 ECG-specific deep, 5 time-series) across three public databases (MIT-BIH, INCART, SVDB) using five random seeds, resulting in 300 paired runs. The primary endpoint was the four-class macro-averaged F1 (M-F1(4)) over N/S/V/F. A support-aware three-class metric (M-F1(3-NSV)) was used as an interpretive endpoint for datasets with sparse F-beat support.

Key Contributions

Traceable Evidence Layer: A fixed, named morphology–rhythm evidence layer interposed between any AAMI-compatible backbone and the final decision, enabling computational traceability of decisions to specific clinical evidence components.
Validation-Bounded BioAdaptive Decoding: A decoder that uses prevalence-derived multipliers and guard-limited adaptation rather than test-tuned thresholds, ensuring that minority recovery is stable and bounded by internal validation support.
Comprehensive Paired Evaluation: A rigorous evaluation across 20 backbones and three databases (300 runs) demonstrating that the framework improves performance without requiring retraining of the underlying waveform encoders.

Results

Performance Gains: CEMR-ECG improved the primary M-F1(4) metric in all 60 dataset–backbone summaries and 297/300 seed-level runs.
- Mean M-F1(4) increased by 11.48 percentage points (from 48.32% to 59.80%).
- Mean Accuracy increased by 6.81 percentage points (from 86.32% to 93.13%).
Class-Specific Recovery:
- S (Supraventricular): F1 increased by 17.53 percentage points (28.11% to 45.64%).
- V (Ventricular): F1 increased by 15.83 percentage points (66.94% to 82.77%).
- F (Fusion): Recovery was support-dependent. MIT-BIH showed a +19.62 point gain; INCART showed +5.58 points. In SVDB, where only 23 F beats existed, F recovery was not pursued due to lack of validation support, and the framework remained conservative.
Component Analysis: Ablation studies indicated that RR rhythm was the primary driver of improvement (removing it caused a 5.06 point drop in M-F1(4)). Morphology and prototype components provided traceability and secondary evidence but contributed less to the aggregate performance gain when removed individually.
Score Quality: The framework improved Brier scores (0.2313 to 0.1172) and Expected Calibration Error (0.0975 to 0.0689) across the decision interfaces.
Overhead: The combined prediction time was approximately 1.57 times that of the raw backbone, with a median of 0.475 seconds per 10,000 test beats.

Significance and Claims
The paper claims that CEMR-ECG provides a reusable, computationally traceable decision layer that enhances imbalanced heartbeat classification without replacing the waveform encoder. Its significance lies in:

Interpretability: Decisions can be linked to named evidence components (RR rhythm, morphology, etc.) rather than opaque scores.
Robustness to Imbalance: It achieves significant gains in minority class recovery (S and V) while maintaining high accuracy for normal beats, without relying on test-set tuning.
Generalizability: The framework is architecture-agnostic, improving performance across diverse machine learning and deep learning backbones.

Limitations and Scope
The authors explicitly state that this is a retrospective computational study using public, de-identified databases.

The framework is not a new waveform classifier but a decision-layer interface.
Claims are limited to the supported N/S/V axes; F-beat recovery is strictly dependent on data support.
No patient-care claims: The study does not include prospective validation, workflow testing, or integration with upstream QRS detectors in a clinical setting. The authors emphasize that prospective, patient-level, and workflow-oriented validation is required before any clinical implementation.