💻 computer science

FedCardioGuard: Quality- and Uncertainty-Aware Personalized Federated Learning with Rare-Class Generalization Audits for ECG Classification

FedCardioGuard is a quality- and uncertainty-aware personalized federated learning framework for ECG classification that significantly improves macro-F1 scores and calibration on non-IID data through multi-scale encoding and adaptive fusion, though it currently struggles with rare-class recognition due to cold-start and score transfer limitations.

Original authors: Sareh Saberi, Abbas Karimi, Faraneh Zarafshan, Javad Akbari Torkestani

Published 2026-10-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sareh Saberi, Abbas Karimi, Faraneh Zarafshan, Javad Akbari Torkestani

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

Heart disease remains one of the leading causes of premature death worldwide, and the electrocardiogram, or ECG, is one of the most accessible tools doctors have to detect dangerous irregularities in heart rhythms. Modern computers can now learn to recognize these patterns directly from the raw electrical signals of the heart, often matching the skill of human specialists. However, a significant problem arises when these computer systems are used to monitor patients in the real world. Heartbeat shapes vary wildly from person to person due to differences in body chemistry, how sensors are placed, and the quality of the recording equipment. Furthermore, the dangerous heartbeats that doctors need to catch are often extremely rare and appear in only a few patients, while normal beats are common in everyone. If a computer system is trained on data that mixes all these different patients together, it often fails to learn the specific quirks of any single individual, leading to missed diagnoses or false alarms.

To solve this, researchers are turning to a method called federated learning. Instead of gathering all patient data into one central computer, which raises serious privacy concerns, this approach keeps the data on the patient's own device. The computer model travels to the data, learns from it, and sends back only a summary of what it learned, leaving the private heart signals behind. The goal is to build a smart system that respects privacy while still learning from many different people. The challenge is that because every patient is different, the data is messy and uneven. Some patients have thousands of normal beats but only a handful of dangerous ones, while others have the opposite. A standard learning system often gets confused by this imbalance, ignoring the rare but critical signals in favor of the common ones.

A team of researchers has developed a new system called FedCardioGuard to address these specific difficulties. This system is designed to work on a patient-by-patient basis, treating each person as a unique learning partner rather than just a source of data points. The system pays close attention to two things that often get overlooked: the quality of the heart signal being recorded and how certain the computer is about its own answers. If a signal is noisy or unclear, the system learns to trust it less. If the computer is unsure about a diagnosis, it is programmed to admit uncertainty rather than guessing, which is crucial for medical safety. The researchers tested this approach using a well-known database of heart recordings, treating each of the twenty-two patients in the test group as a separate client. They ran the system through ten different training scenarios to ensure the results were consistent and not just a lucky accident.

The results showed that this new approach worked significantly better than previous methods at identifying heart rhythms across the board. The system achieved a score of 0.556 in correctly identifying all types of heartbeats, a notable improvement over the next best method, which scored 0.514. More importantly, the system was much better at protecting the patients who were hardest to diagnose. In the worst-case scenario for any single patient, the new system improved its ability to correctly identify heartbeats by nearly fifty percent compared to older techniques. It also became much better at knowing when it was right. The system's confidence scores, which tell doctors how much to trust a specific reading, were far more accurate, reducing the error in confidence from a high level down to a very low one. This means doctors can rely more on the computer's certainty, knowing that when the system says it is sure, it is likely correct.

However, the researchers were careful to point out that the system is not yet a finished product ready for every hospital. While it excelled at identifying the most common types of irregularities and the most dangerous ones, it still struggled with a specific, less common type of abnormal beat. In the tests, the system correctly identified this rare beat only about eight percent of the time, falling far short of the target needed for safe clinical use. The researchers tried several different adjustments to fix this, such as changing how the computer weighs different types of beats or how it learns from rare examples, but none of these changes solved the problem without making other parts of the system worse. In fact, some adjustments that improved the detection of rare beats made the system less trustworthy overall, causing it to become overconfident in wrong answers.

Because of this lingering weakness, the researchers decided not to test the system on a second, larger set of patient data. They kept that second set completely locked away, refusing to look at it until the system can reliably detect all types of heartbeats, including the rare ones, without losing its accuracy on the common ones. This decision highlights a key principle in their work: a system that is good on average is not good enough if it fails on the specific patients who need help the most. The study serves as a strong proof of concept that personalized, privacy-preserving learning can work better than traditional methods, but it also clearly defines the limits of what is currently possible. The system successfully learned to balance the needs of different patients and to recognize when it was unsure, but the final hurdle of reliably catching every single type of rare heart rhythm remains unsolved. The researchers view their work as a validated framework for future development rather than a completed medical tool, emphasizing that the path forward requires solving the specific problem of rare beat detection before the technology can be trusted in the real world.

Technical Summary of FedCardioGuard

Problem Statement
The paper addresses the challenge of deploying dependable, patient-specific electrocardiogram (ECG) classification systems in a federated learning (FL) environment. Existing centralized deep learning approaches fail to account for data privacy and the natural fragmentation of medical data across institutions. Conversely, standard federated learning methods often struggle with three specific compounding issues in the context of ECG arrhythmia detection:

Patient-level Non-IID Data: Data distributions vary significantly across patients due to physiological differences, electrode placement, and recording conditions.
Severe Class Imbalance: Clinically relevant ectopic beats (e.g., supraventricular and fusion beats) are rare and unevenly distributed, leading to poor recognition of minority classes.
Unreliable Confidence and Edge Constraints: Neural networks often produce miscalibrated confidence scores under distribution shifts, and edge devices (wearables) have strict memory and latency budgets.

Current solutions typically address these issues in isolation (e.g., personalization without signal quality awareness, or uncertainty estimation without federated constraints), leaving a gap for a unified framework that handles noisy acquisition, heterogeneous distributions, and resource constraints simultaneously.

Methodology: FedCardioGuard
The authors propose FedCardioGuard, a quality- and uncertainty-aware personalized federated learning framework designed for a "patient-as-client" architecture. The system operates without centralizing raw ECG signals and integrates six key components:

Quality-Gated Lightweight Encoder:
- A multi-scale 1D CNN encoder processes ECG morphology and rhythm features.
- Instead of scaling raw waveforms (which distorts physiology), a Signal Quality Index (SQI) module modulates latent features via affine gates.
- SQI values also weight the local loss contribution, ensuring low-quality segments contribute less to training.
Quality- and Reliability-Aware Aggregation:
- The server aggregates model updates not just by sample count, but by a composite weight considering effective sample support, measured signal quality, and update reliability (distance from the median update).
- Updates are clipped to prevent outliers from dominating the global model.
Uncertainty-Weighted Prototype Personalization:
- The framework employs a local-global prototype mechanism. Local class prototypes are computed but are subject to shrinkage toward robust global prototypes if local evidence is sparse or uncertain.
- A reliance coefficient ( $\lambda_{k,c}$ ) dynamically balances local and global prototypes based on class support and estimated uncertainty (via Monte Carlo dropout).
- This prevents patients with few rare-class examples from learning unreliable, noisy decision boundaries.
Calibrated Selective Prediction:
- The system uses a gated mixture of parametric and prototype heads.
- Predictive uncertainty is quantified using entropy and stochastic disagreement.
- A selective prediction layer allows the model to "abstain" (defer to clinical review) if confidence is low or uncertainty is high, ensuring that only reliable predictions are output.
Edge Deployment Strategy:
- Compression (pruning and INT8 quantization) is applied strictly after training and calibration are frozen.
- This ensures that resource constraints are evaluated independently of the learning process, with validation checks on degradation.
Locked Evaluation Protocol:
- The study utilizes the MIT-BIH Arrhythmia Database with a strict patient-disjoint split (DS1 for development, DS2 locked).
- The evaluation uses 22 patient-clients, 10 independent random seeds, and 15 federated rounds to ensure statistical robustness.

Key Contributions
The paper claims the following methodological contributions:

Integrated Framework: It is the first to couple signal quality gating, uncertainty-aware prototype shrinkage, and calibrated selective prediction within a patient-centric FL pipeline.
Lightweight Architecture: A multi-scale encoder that modulates features based on SQI without altering raw signal amplitudes.
Robust Aggregation: A server-side rule that weights clients by support, quality, and reliability rather than raw volume.
Personalization Mechanism: A class-conditional shrinkage strategy that protects against overfitting to sparse local rare-class data.
Rigorous Audit: A "locked" development study design where every component maps to a reproducible ablation, and the external test set (DS2) remains unopened throughout the development phase.

Results
The study reports results based on the DS1 development set (22 patients) across 10 seeds:

Primary Performance: FedCardioGuard achieved a macro-F1 of 0.556 ± 0.049, significantly outperforming baselines FedProto (0.514), Ditto (0.488), and FedPer (0.251). The improvement over FedProto was 0.042 (8.2% relative).
Statistical Significance: The Friedman test confirmed significant differences ( $p = 2.23 \times 10^{-5}$ ), with paired Wilcoxon tests (Holm-corrected) showing FedCardioGuard was superior to all baselines.
Calibration: The framework reduced Expected Calibration Error (ECE) from 0.245 (FedProto) to 0.089 without altering hard predictions, significantly improving reliability.
Robustness: The "support-aware worst-client F1" improved by 49.7% over FedProto (from 0.198 to 0.296), indicating better performance on the most difficult patients.
Limitations: Despite improvements, the system struggled with the rare S-class (supraventricular) beats, achieving a recall of only 0.082. The F-class recall was 0.465.
Ablation Findings: Subsequent "Stage 2" and "Stage 3" interventions (attempting to fix rare-class recall via resampling or contrastive learning) failed to meet pre-specified promotion gates (Recall-S $\ge$ 0.40, ECE $\le$ 0.05) and often degraded calibration or worst-client performance.

Significance and Claims
The authors position FedCardioGuard not as a finalized clinical classifier, but as a validated development framework.

Validation of Design: The results support the hypothesis that combining prototype personalization with uncertainty-aware shrinkage and calibration improves macro-F1 and lower-tail robustness simultaneously.
Modesty on Claims: The paper explicitly states that the current results are "development evidence" because the external DS2 partition was never opened. The failure to resolve rare-class (S-class) recognition is identified as the principal unresolved limitation.
Governance: The study emphasizes a rigorous governance model where model selection is strictly controlled to prevent overfitting to the development set. The failure of subsequent interventions to meet all gates reinforces the difficulty of the rare-class problem in federated settings.
Conclusion: The work demonstrates that a patient-centric, quality-aware, and uncertainty-calibrated approach is superior to standard FL baselines for ECG, but it also highlights that solving rare-class generalization in this setting remains an open research challenge requiring further representation-level innovation.

Technical Summary of FedCardioGuard

More like this