A Novel Approach to Create Synthetic Biomedical Signals Using BiRNN

Introduction

Automated medical diagnosis systems rely heavily on large, high-quality biomedical signal databases. However, creating or accessing these databases can be time-consuming, expensive, and limited by patient privacy concerns. Researchers often encounter incomplete or unbalanced datasets as well. To address these issues, this paper proposes a novel approach to generate synthetic biomedical signals using a Bidirectional Recurrent Neural Network (BiRNN). The system aims to create realistic data that reflects the characteristics of genuine patient signals without disclosing any sensitive medical information.

The approach is designed to produce multiple types of time-series biomedical signals (ECG, EEG, ballistocardiogram, photoplethysmogram, and respiratory impedance) and includes an optional preprocessing step to remove noise. This synthetic-data generation pipeline empowers researchers and clinicians to augment real-world databases for machine learning model training, testing, and validation.

Methodology

1. Data Preprocessing (Optional)

Certain biomedical signals (e.g., ECG, PPG) require denoising before segmentation. This paper applies a Discrete Wavelet Transform (DWT) to attenuate baseline wander and electrical noise. When the raw data are sufficiently clean, this step is omitted.

2. Segmentation and Normalization

The signals are segmented according to class labels (e.g., different arrhythmia types in ECG). To manage variable lengths within each class, a resampling procedure is used to standardize the segment lengths. A Z-score normalization step removes offset and amplitude scaling issues.

3. Bidirectional RNN (BiRNN) for Signal Generation

The core generative model is a BiRNN, which captures temporal dependencies both forward and backward along the signal. Each segment is paired with a slightly perturbed version of itself – one as the training input and one as the desired output. The network iterates over these sequences to learn a mapping from “noisy input” to “clean output,” thereby creating realistic synthetic signals.

Once trained, the BiRNN can generate new synthetic segments that mirror real physiological patterns. Parameters such as the hidden layer size, number of layers, and activation functions are tuned for optimal replication of the original data distribution.

4. Statistical Filtering

The paper employs Pearson’s Correlation Coefficient to verify the morphological similarity between each synthetic segment and the corresponding original segment. Any generated signal failing to surpass a correlation threshold of 0.90 is discarded, ensuring that the final synthetic dataset closely resembles the real signals in waveform and key features.

Key Results

Multi-Signal Generation: The BiRNN method was successfully used to generate ECG, EEG, BCG, PPG, and respiratory impedance signals, demonstrating broad applicability.

High Correlation: The correlation coefficient between original and synthetic signals exceeded 0.90 in most cases, indicating strong morphological similarity.

Low Error Rates: Quantitative metrics such as Percent Root Mean Square Difference (PRD) and RMSE were significantly lower compared to other generative methods (e.g., certain GAN-based models).

Enhanced Classifiers: Classification experiments (e.g., ECG arrhythmia classification, EEG open/closed-eye detection) using synthetic data combined with real data achieved high accuracy, illustrating the practical utility of the generated signals.

Scalability: By iteratively reusing the synthetic dataset as “new” training data, the system can repeatedly double the amount of available training data, continuously refining model performance.

Conclusion

This paper presents an effective pipeline to synthesize diverse biomedical signals. The BiRNN-based approach learns from real datasets and outputs realistic signals while preserving patient privacy. Rigorous statistical filtering ensures that synthetic signals closely resemble the morphology of actual physiological data.

Experimental results show low error metrics and strong correlation coefficients, surpassing many alternative generative models. Moreover, the synthetic data have been validated by training machine learning classifiers for arrhythmia and EEG-based state recognition, underscoring their practical value.

Future research includes expanding the BiRNN architecture for highly varied or lengthy signals, examining privacy safeguards, and adding domain-specific constraints (e.g., clinically significant rhythms or events). Nonetheless, the current study demonstrates that BiRNN-generated synthetic biomedical signals are a valuable and scalable resource for modern data-driven healthcare.