PaperlayerView exact revision on arXiv

Source-linked AI summary

Boosting Deep Learning Risk Prediction with Generative Adversarial Networks for Electronic Health Records

Zhengping Che, Yu Cheng, Shuangfei Zhai, Zhaonan Sun, Yan Liu

arXiv:1709.01648v1cs.LGstat.ML

TL;DR

Limited accurate EHR labels constrain deep-learning risk prediction. The paper introduces ehrGAN and a semi-supervised CNN framework to generate plausible labeled records and augment training, improving classification across two clinical datasets over several baselines.

  • Problem

    Healthcare EHRs have heterogeneous structure and limited accurate labels, while deep-learning methods generally require massive labeled datasets.

  • Method

    The framework combines an EHR-specific adversarial generator, ehrGAN, with CNN-based prediction and semi-supervised data augmentation.

  • Results

    The framework produces realistic samples and improves classification performance over strong baselines across two real clinical datasets.

  • Takeaways & Limitations

    Generated EHR data and learned representations can support boosted onset-risk prediction when labeled training data are limited.

Abstract

from arXiv · show

The rapid growth of Electronic Health Records (EHRs), as well as the accompanied opportunities in Data-Driven Healthcare (DDH), has been attracting widespread interests and attentions. Recent progress in the design and applications of deep learning methods has shown promising results and is forcing massive changes in healthcare academia and industry, but most of these methods rely on massive labeled data. In this work, we propose a general deep learning framework which is able to boost risk prediction performance with limited EHR data. Our model takes a modified generative adversarial network namely ehrGAN, which can provide plausible labeled EHR data by mimicking real patient records, to augment the training dataset in a semi-supervised learning manner. We use this generative model together with a convolutional neural network (CNN) based prediction model to improve the onset prediction performance. Experiments on two real healthcare datasets demonstrate that our proposed framework produces realistic data samples and achieves significant improvements on classification tasks with the generated data over several stat-of-the-art baselines.

I. INTRODUCTION

Deep learning could improve healthcare, but EHR heterogeneity and scarce accurate labels limit its application. The paper addresses this gap with ehrGAN-based augmentation and semi-supervised risk prediction.

  • EHRs support data-driven healthcare, including clinical tasks such as phenotyping, risk prediction, imaging analysis, and clinical language processing.
  • Heterogeneous, irregular, noisy, and incomplete EHRs make healthcare modeling harder than modeling clean data.
  • Accurate clinical labels are limited because expert diagnosis and patient labeling are time-consuming, especially for rare diseases and conditions.
  • Semi-supervised learning combines unlabeled or augmented data with relatively small labeled datasets to improve performance.
  • Prior semi-supervised EHR work mainly addresses clinical text, with limited application to structured quantitative EHR data.
  • The paper proposes ehrGAN and a semi-supervised framework to augment EHR data and improve risk prediction with limited labels.

II. RELATED WORK

Prior healthcare deep-learning models are largely supervised, while GAN-based semi-supervised methods have mostly targeted vision or NLP. The paper extends this paradigm to temporal EHR onset prediction.

  • Existing EHR deep-learning models use architectures including MLPs, RNNs, state-space models, and CNNs for complex temporal and clinical patterns.
  • Most prior EHR models rely mainly on fully labeled data, although supervised information is expensive and limited.
  • Earlier semi-supervised healthcare studies primarily address clinical text, structured quantitative EHR learning, graph classification, or healthcare forum question answering.
  • GANs train a generator to synthesize dataset-like samples while a discriminator distinguishes real from generated samples.
  • GAN-based semi-supervised frameworks combine generated samples, discriminator representations, and classifier outputs for classification.
  • The proposed paradigm belongs to data-augmentation methods for semi-supervised learning.
  • The framework addresses the difficulty of adapting GANs to EHRs and using them for onset prediction, areas previous models did not directly cover.

III. THE PROPOSED METHOD

The proposed method combines a temporal CNN risk predictor with ehrGAN, a modified GAN designed for EHR generation. Generated records and learned representations support semi-supervised onset prediction.

  • The framework contains a CNN baseline, an EHR-specific modified GAN, and a data-augmented semi-supervised prediction scheme.
  • The prediction model applies 1D temporal convolution and max-over-time pooling to patient event embeddings.
  • Multiple temporal filter lengths capture dependencies at different levels before pooling the most important features.
  • The original GAN trains generator and discriminator networks through an iterative minimax game involving real and generated samples.
  • For EHRs, the generator and discriminator use one-dimensional convolutional and deconvolutional neural networks.

1) Discriminator:

ehrGAN modifies GAN generation for EHRs by learning transitions from real records to neighboring synthetic samples. Its objective balances adversarial realism with reconstruction closeness.

  • 1) Discriminator: The discriminator adopts the prediction model and replaces its top layer with a sigmoid output for real-sample probability.
  • 2) Generator: A conventional generator maps latent noise to synthetic samples but cannot directly provide class-aware information for semi-supervised learning.
  • 2) Generator: VCD-based ehrGAN learns a transition distribution p(˜x|x), generating neighboring samples from real training records.
  • 2) Generator: The generator mixes encoded real-record representations with noise through a binary mask before decoding synthetic samples.
  • 2) Generator: The objective combines adversarial loss with reconstruction error, while ρ controls closeness between synthetic and corresponding real samples.
  • 2) Generator: The VCD design is intended to avoid mode collapse and preserve local data-manifold structure useful for classification.
  • 2) Generator: ehrGAN training uses alternating stochastic-gradient optimization with generator updates performed five times per discriminator update.

3) Training techniques:

The ehrGAN optimization alternates generator and discriminator updates until convergence, using sampled noise, binary masks, and training data in nested iteration loops.

  • Training repeats the optimization procedure for enough iterations until convergence.
  • Each outer iteration performs k inner generator-update steps.
  • Generator updates use sampled noise variables and binary mask vectors with gradient ascent on the discriminator response to generated data.
  • Discriminator updates use training data, noise variables, and binary mask vectors with gradient descent.

C. Semi-supervised Learning with GANs

SSL-GAN uses the learned ehrGAN transition distribution to augment labeled training data and incorporates the augmented samples into classifier training through a weighted loss.

  • SSL-GAN uses the learned transition distribution p(x̃|x) to perform data augmentation.
  • The training loss combines binary crossentropy on original samples with a weighted contribution from augmented samples.
  • The weight µ reflects the ratio between the numbers of training data and GAN-augmented data.
  • The method assumes generated samples from a well-trained generator are likely to remain aligned with the source sample’s class.

IV. EXPERIMENTAL RESULTS

Experiments use longitudinal EHR data from heart-failure and diabetes cohorts to evaluate realistic generation and onset-prediction performance under defined cohort and temporal-processing settings.

  • Datasets and settings: The experiments use two real clinical datasets from heart-failure and diabetes cohorts.
  • Datasets and settings: The source database contains 218,680 patients, 14,969,489 observations, and 14,690 unique medical events collected between 2011 and 2015.
  • Datasets and settings: Heart Failure includes 3,357 case patients and 6,714 controls, while Diabetes includes 2,248 case patients and 4,496 controls.
  • Datasets and settings: Diagnosis and medication features are grouped into disjoint 90-day windows, and datasets are split into training, validation, and test sets at a 7:1:2 ratio.
  • Datasets and settings: The embedding dimension is 200, and ehrGAN is trained only on the training subset.
  • Evaluation: Table I presents the prediction-performance comparison using classification accuracy and AUROC.

B. Risk Prediction Comparison on Basic Models

The CNN baseline performs strongly across the two prediction tasks, while ehrGAN-generated data are evaluated for realism before augmentation. Generated records reproduce original length, feature-frequency, and comorbidity patterns across cohorts.

  • CNN is among the best methods for Heart Failure and significantly outperforms baselines for Diabetes.The improvement is attributed mainly to learned embeddings in Heart Failure and CNN structure in Diabetes.
  • GRU and LSTM perform well but do not surpass CNN.
  • Generated datasets have length distributions similar to the original datasets.
  • Generated data preserve frequencies for the 100 most frequent features and reproduce comorbidity clusters among the 20 most frequent diagnosis features.
  • The generated models capture occurrence patterns across case cohorts while retaining cohort-specific frequency differences.

D. Evaluation of the Boosted Model

The evaluation compares ehrGAN-based SSL-GAN with supervised, randomized-label, and standard semi-supervised baselines under varying labeled-data settings. SSL-GAN consistently improves over CNN-BASIC and CNN-RAND, approaches CNN-FULL, and outperforms the standard SSL baselines.

  • The evaluation includes CNN-BASIC, CNN-FULL, CNN-RAND, SSL-SMIR, LGC, and the proposed SSL-GAN.These approaches compare training with labeled subsets, additional data, randomized labels, standard semi-supervised methods, and ehrGAN-based augmentation.
  • SSL-GAN consistently beats CNN-BASIC and CNN-RAND across the evaluated labeled-data settings.CNN-RAND performs worse than CNN-BASIC because its additional labels are randomized.
  • 0.8574 accuracy and 0.9075 AUROC are achieved by SSL-GAN on HF50, versus 0.8096 accuracy and 0.8784 AUROC for CNN-BASIC.
  • SSL-GAN improves 3% −4% over the baseline in both measurements on Dia50 and Dia67.
  • SSL-GAN is about 2% lower in AUROC than CNN-FULL on Dia50 and Dia67, with smaller margins on HF50 and HF67.
  • SSL-SMIR and SSL-LGC achieve performances similar to CNN-BASIC, while SSL-GAN easily beats them.

E. Selections of Parameters

Parameter selection determines whether generated EHR samples improve prediction: moderate transition and augmentation settings help, whereas excessive or poorly labeled synthetic data can reduce performance. Across the tested settings, ρ = 0.1 and µ = 0.6 are identified as effective choices for SSL-GAN.

  • The effectiveness of ρ: ρ = 0.1 is optimal, while ρ = 0 hurts performance and ρ = 1 disrupts training because generated samples may not preserve labels.The transition distribution at ρ = 0.1 generates samples sufficiently different from training data to improve learning.
  • The effectiveness of µ: µ controls the ratio of real labeled data to GAN-augmented data in the semi-supervised prediction framework.The experiments vary µ while holding ρ fixed to assess how much augmented data should support supervised learning.
  • The effectiveness of µ: µ = 0.6 gives SSL-GAN its best prediction performance when ρ = 0.1, after which performance decreases slightly.The comparison uses prediction AUROC scores across different µ values and includes a fully labeled-data setting.
  • The effectiveness of µ: For fully labeled data, prediction performance continues improving as µ increases, unlike SSL-GAN at its optimal ρ setting.In SSL-GAN, adding more augmented data beyond the best setting does not provide further benefit.
  • The effectiveness of µ: Accuracy shows a trend similar to the AUROC-based parameter comparison.The reported similarity extends the observed parameter-selection pattern beyond AUROC.

V. CONCLUSION

The paper presents ehrGAN and a semi-supervised augmentation framework for deep-learning risk prediction on EHR data. Experiments on two datasets show improved generalization and prediction performance over strong baselines, while future work calls for broader clinical comparisons and richer clinical structure.

  • Conclusion: The proposed ehrGAN generates realistic EHR samples by mimicking real data and learns a latent space that continuously encodes plausible samples.The model uses adversarial training for generation and then supports downstream data augmentation.
  • Conclusion: Semi-supervised learning uses the learned generative model to augment data and boost risk prediction performance.The framework combines generation with prediction rather than treating ehrGAN only as a standalone data model.
  • Conclusion: Experiments on two datasets show improved generalization power and prediction performance compared with strong baselines.The conclusion reports a cross-dataset improvement without specifying a single metric value.
  • Future work: Future work includes comprehensive clinical comparisons with domain experts and improved clinical interpretation or structural information.The authors also identify extensions involving newer GANs, boosted learning techniques, and other healthcare applications.
Loading 1709.01648v1…