TL;DR
This paper introduces CorGAN, a novel framework using Convolutional Generative Adversarial Networks (GANs) and Convolutional Autoencoders to generate realistic synthetic healthcare records. CorGAN effectively captures complex correlations within EHR data, producing synthetic data that performs comparably to real data in downstream machine learning tasks while offering enhanced privacy.
What Problem Does This Paper Address?
The increasing adoption of Electronic Health Records (EHRs) has unlocked vast potential for AI-driven healthcare advancements. However, accessing and utilizing this sensitive data for research is severely hampered by privacy concerns. While de-identification methods exist, their long-term effectiveness and vulnerability to information leakage are increasingly questioned. This has led to a growing interest in Synthetic Data Generation (SDG) as a privacy-preserving alternative. The challenge lies in creating synthetic EHR data that is not only realistic and privacy-preserving but also captures the intricate correlations between various medical features, which are crucial for accurate downstream analysis.
Methods
The core of this work is the CorGAN (Correlation-Capturing Generative Adversarial Network) architecture. CorGAN combines two key components:
- Convolutional Autoencoders (CAEs): A pre-trained CAE is used to bridge the gap between continuous outputs from the GAN generator and the discrete nature of EHR data. The CAE's decoder transforms the generator's continuous latent space into discrete, interpretable values. This indirect method addresses the common challenge GANs face in generating discrete data directly.
- Convolutional GANs (CGANs): Instead of standard Multilayer Perceptrons (MLPs), CorGAN employs 1D Convolutional Neural Networks (CNNs) for both the generator and discriminator. This is a crucial innovation, as CNNs are adept at capturing local correlations and patterns, which are vital for understanding relationships between adjacent medical features in sequential EHR data.
The training process involves a min-max game between the generator and discriminator, aiming to produce synthetic data that the discriminator cannot distinguish from real data. For discrete outputs, the continuous values from the decoder are rounded to 0 or 1.
Datasets
The authors evaluated CorGAN on two publicly available datasets:
- MIMIC-III Dataset: This dataset, containing medical records of nearly 46,000 patients, was used for experiments with binary discrete variables. Specifically, ICD-9 codes were extracted and represented as fixed-size vectors.
- UCI Epileptic Seizure Recognition Dataset: This dataset, characterized by continuous EEG recordings, was used for experiments with continuous variables. It features 179 features and 11,500 samples, with the primary task being seizure activity classification.
Key Insights and Ideas
- CNNs for EHR Correlation: The paper strongly advocates for using CNNs over MLPs for EHR data. This is because CNNs are inherently better at capturing local dependencies and correlations between adjacent features, which are prevalent in medical records (e.g., sequences of diagnoses or lab tests).
- Bridging Discrete and Continuous: The use of a pre-trained autoencoder's decoder to transform continuous GAN outputs into discrete EHR features is an effective strategy for handling the mixed data types common in healthcare.
- Privacy-Preserving Realism: CorGAN demonstrates that it's possible to generate synthetic data that is both realistic enough for machine learning tasks and offers a good level of privacy protection.
Key Results
- Data Quality: CorGAN generated synthetic data that performed comparably to real data in various machine learning settings, including classification and prediction tasks.
- Dimension-wise Probability: CorGAN showed superior performance in matching the dimension-wise probability distributions of real data compared to baseline models like DBM, VAE, and medGAN.
- Dimension-wise Prediction: For discrete data (MIMIC-III), CorGAN achieved the smallest F1-score difference (0.021 ± 0.045), indicating that models trained on its synthetic data performed very similarly to models trained on real data.
- Binary Classification: For continuous data (UCI Epileptic Seizure Recognition), CorGAN achieved the highest averaged AUROC (0.92 ± 0.012) and AUPRC (0.41 ± 0.015) among the generative models, suggesting better synthetic data quality for classification tasks.
- Privacy Assessment: The membership inference attack results indicated that CorGAN provides an acceptable level of privacy, especially as the amount of synthetic data increases relative to known real data. Increasing synthetic data volume significantly decreased precision in attack success, making it harder for an adversary to identify real records.
Limitations and Future Work
- Temporal Information: While the paper highlights CorGAN's ability to capture correlations, it acknowledges that for the MIMIC dataset experiments, temporal information was ignored. The authors suggest that CorGAN can be effectively extended to handle longitudinal records, where capturing temporal characteristics is crucial.
- Privacy Thresholds: The privacy assessment involved setting a threshold for cosine similarity. Further research could explore more sophisticated privacy metrics and attack models.
- Dataset Specificity: While evaluated on two datasets, the generalizability of CorGAN to a wider range of EHR data complexities and formats could be further investigated.
Summary updated