Abstract
Inter-patient variability presents a significant challenge for automated electrocardiogram (ECG) analysis, as heart signal morphology varies significantly between individuals. This paper proposes Personalized Generative Adversarial Networks (PGANs) to synthesize patient-specific synthetic data. By tuning a general generative model with a small amount of unlabeled data from a specific patient, the system generates artificial heartbeats that reflect the patient's unique cardiac signature. These synthetic examples are used to augment training sets for deep learning classifiers, leading to substantial improvements in arrhythmia detection accuracy without requiring extensive manual labeling by clinicians.
Introduction: The Challenge of Inter-Patient Variability
Electrocardiograms (ECG) are essential for monitoring cardiac health, yet automating their interpretation remains difficult due to the unique "electrical language" of every patient's heart. Factors such as age, physical condition, and electrode placement cause a "normal" heartbeat in one individual to resemble an arrhythmia in another. While deep learning models like LSTMs perform well on general populations, they often fail when applied to new individuals whose cardiac morphology differs from the training set.
To achieve high accuracy, models ideally require thousands of labeled heartbeats from the specific patient being monitored. However, manual labeling by cardiologists is time-consuming and impractical for every new patient. This research addresses the "data sparseness" problem by using generative models to create personalized, labeled training data from unlabeled observations.
Methodology: From Generalist to Specialist
The authors propose a two-stage generative approach to bridge the gap between general medical knowledge and individual patient needs.
1. ECG GAN (The Generalist)
Initially, a GAN is trained on a large, diverse population. To ensure the generated signals are medically plausible, the model incorporates a specialized Mean Square Error (MSE) loss based on cardiac morphology. It specifically optimizes for the five key turning points of a cardiac cycle: the P, Q, R, S, and T waves.
2. PGAN (The Specialist)
To personalize the model, the authors utilize a few minutes of unlabeled ECG data from a new patient. A pre-trained general classifier identifies beats likely belonging to specific arrhythmia classes. The unique PQRST features are then extracted from these patient-specific beats to "nudge" the GAN's loss function. The resulting PGAN generator produces synthetic heartbeats that combine the rhythm of a disease with the unique morphology of the specific patient.
Experimental Setup
The study utilized the MIT-BIH Arrhythmia Database, which includes 48 half-hour ECG records and approximately 109,000 labeled heartbeats. The evaluation focused on three clinically significant arrhythmia types:
- Supraventricular ectopic beats (S)
- Ventricular ectopic beats (V)
- Fusion beats (F)
Synthetic data generated by the PGAN was used to augment the training set for a patient-specific LSTM classifier, which consists of two layers with 512 neurons each.
Results and Discussion
The PGAN-augmented models significantly outperformed standard LSTMs and non-personalized GAN approaches. By forcing the generator to respect the PQRST wave structure, the authors ensured the synthetic data remained medically relevant rather than just visually similar.
Key performance metrics included:
- Ventricular Ectopic Beats (V): 0.99 AUC
- Fusion Beats (F): 0.95 AUC
- Supraventricular Ectopic Beats (S): 0.85 AUC
The results indicate that adding synthetic personalized data consistently improves performance. However, there is a "sweet spot" for the number of synthetic samples; adding an excessive amount can eventually introduce noise, though the initial gains are substantial.
Conclusion and Future Work
PGANs demonstrate that generative models can serve as sophisticated data augmentation tools, effectively personalizing medical AI using only unlabeled data. This approach reduces the burden on clinicians while improving diagnostic accuracy.
Limitations: * Cold Start: The method requires at least some instances of an arrhythmia to appear in the unlabeled data for the PGAN to learn the specific morphology. * Computational Cost: Training a separate GAN per patient is more resource-intensive than a single global model.
Future research may extend this framework to other physiological signals, such as EEG for brain activity, where individual variability similarly limits the effectiveness of "one-size-fits-all" models.
Summary updated