PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep EHR: A Survey of Recent Advances in Deep Learning Techniques for Electronic Health Record (EHR) Analysis

Benjamin Shickel, Patrick Tighe, Azra Bihorac, Parisa Rashidi

arXiv:1706.03446v2cs.LGstat.ML

TL;DR

EHR data are heterogeneous and increasingly used for clinical-informatics tasks, motivating methods that can learn useful representations from them. The paper surveys deep-learning architectures, technical approaches, and applications, finding an emerging field with broad use across EHR analysis and continuing opportunities for unified representations.

  • Problem

    EHR data are heterogeneous, while existing clinical-text and structured-data analyses have often required manual feature engineering or separate representations.

  • Method

    The paper surveys deep-learning techniques for EHR analysis and inference, covering architectures, representation learning, clinical applications, and future directions.

  • Results

    The surveyed studies apply deep learning across information extraction, patient representation, outcome prediction, and computational phenotyping tasks.

  • Takeaways & Limitations

    Unified representations that combine different patient-data types are identified as an important future direction for patient and disease modeling and related tasks.

Abstract

from arXiv · show

The past decade has seen an explosion in the amount of digital information stored in electronic health records (EHR). While primarily designed for archiving patient clinical information and administrative healthcare tasks, many researchers have found secondary use of these records for various clinical informatics tasks. Over the same period, the machine learning community has seen widespread advances in deep learning techniques, which also have been successfully applied to the vast amount of EHR data. In this paper, we review these deep EHR systems, examining architectures, technical aspects, and clinical applications. We also identify shortcomings of current techniques and discuss avenues of future research for EHR-based deep learning.

I. INTRODUCTION

EHR adoption and data availability have expanded, creating opportunities for secondary clinical-informatics use. This survey focuses on deep-learning techniques tailored to heterogeneous EHR data and their applications.

  • EHR systems contain demographic, diagnostic, laboratory, medication, imaging, and clinical-note data used for clinical-informatics tasks.
  • Traditional EHR analysis commonly used logistic regression, support vector machines, and random forests before deep-learning methods became prominent.
  • Deep learning has been applied to EHR data because it constructs hierarchical features and captures long-range dependencies, often with less preprocessing and feature engineering.
  • The survey reviews EHR-specific deep-learning techniques, technical aspects, and clinical applications rather than health informatics broadly.
  • Publication searches through August 2017 examined EHR-related deep-learning studies across application areas and technical methods.
  • EHR data combine numerical, datetime, categorical, free-text, and derived time-series representations.

III. MACHINE LEARNING OVERVIEW

Deep learning builds representations from data using neural-network architectures rather than relying exclusively on manually engineered features. The overview introduces common architectures, optimization, and software tools used in deep EHR research.

  • Supervised learning maps inputs x to outputs y, whereas unsupervised learning seeks properties of the input distribution.
  • Deep learning learns data-driven features and hierarchical representations instead of depending entirely on hand-crafted features.
  • Artificial neural networks consist of interconnected nodes arranged in layers, with hidden-unit weights updated during training.
  • ANN weights can be optimized by minimizing negative log likelihood with regularization, using backpropagation to reduce overfitting and improve generalization.
  • Open-source deep-learning tools include TensorFlow, Theano, Keras, Torch, PyTorch, Caffe, CNTK, and Deeplearning4j.
  • The survey covers supervised MLP, CNN, and RNN models alongside unsupervised autoencoder and restricted Boltzmann machine architectures.

A. Multilayer perceptron (MLP)

MLPs use fully connected forward networks to transform inputs through nonlinear hidden layers, while CNNs impose local connectivity and shared filters to extract features from structured data.

  • Multilayer perceptron (MLP): An MLP contains multiple hidden layers whose neurons are fully connected to the next layer, with data flowing in one direction.
  • Multilayer perceptron (MLP): Each MLP hidden unit applies a nonlinear activation to a weighted sum of outputs from the preceding layer.
  • Multilayer perceptron (MLP): Adding hidden layers is expected to produce increasingly abstract representations through successive nonlinear transformations.
  • Convolutional neural networks (CNN): CNNs model spatially or locally structured inputs by extracting features from local patches or signal segments.
  • Convolutional neural networks (CNN): Convolution applies a weighting function or kernel across an input to produce feature maps, followed typically by pooling or subsampling.
  • Convolutional neural networks (CNN): CNN filters are smaller than the input and reused across it, enabling sparse interactions, parameter sharing, and relatively few parameters.

C. Recurrent neural networks

RNNs are designed for sequentially ordered data by carrying information through a recurrent hidden state. Gated variants regulate information flow and support longer-term dependencies.

  • RNNs are suited to sequential data such as time series and natural language, unlike CNNs, which are suited to clear spatial structure.
  • At each time step, an RNN updates its hidden state using the current input and the previous hidden state.
  • After processing a sequence, the final hidden state contains information from its preceding elements.
  • LSTM and GRU models replace standard recurrent units with cells containing recurrence loops and gates that control information flow.
  • Gated RNNs have shown benefits for modeling longer-term sequential dependencies.

D. Autoencoders (AE)

Autoencoders learn lower-dimensional representations by encoding inputs and reconstructing them, and can be stacked hierarchically for richer EHR representations. Related unsupervised architectures, including RBMs, learn representations through probabilistic modeling and can also be stacked.

  • D. Autoencoders (AE): Autoencoders encode input x into a lower-dimensional representation z and decode z to reconstruct an approximation x̃.Training minimizes reconstruction error ∥x − x̃∥, making z more reliable as reconstruction improves.
  • D. Autoencoders (AE): After training, the innermost hidden-layer activations serve as encoded representations that retain important derived dimensions through nonlinear transformations.This makes autoencoders similar to PCA and SVD while supporting nonlinear dimensionality reduction.
  • D. Autoencoders (AE): Stacked autoencoders train hidden layers independently, feeding the first layer’s embeddings z into a second autoencoder.This hierarchical construction extends representation learning across multiple encoding stages.
  • D. Autoencoders (AE): Restricted Boltzmann machines model the probability distribution of input data using stochastic, energy-based representations.RBMs have no connections among visible units or among hidden units, and can be hierarchically stacked into deep belief networks.
  • D. Autoencoders (AE): The reviewed deep EHR applications include information extraction and use both private institutional datasets and public resources such as MIMIC and i2b2.The review notes that many applications rely on private datasets, while several studies use freely available or public clinical-note datasets.

A. EHR Information Extraction (IE)

Deep EHR information extraction addresses structured concept, temporal event, relation, and abbreviation extraction from difficult unstructured clinical notes. The broader representation-learning literature converts sparse codes and patient histories into vectors used for predictive tasks, but proprietary datasets limit direct comparison.

  • A. EHR Information Extraction (IE): Clinical EHR information extraction comprises single concept, temporal event, relation, and abbreviation extraction.These subtasks target structured information that is difficult to obtain from unstructured clinical notes.
  • A. EHR Information Extraction (IE): RNN architectures including LSTMs, GRUs, and Bi-LSTMs were evaluated for clinical concept sequence labeling against CRF baselines.The task assigns clinically relevant tags to words, covering medication and disease concepts such as dosage, route, indication, and severity.
  • A. EHR Information Extraction (IE): Word embeddings trained from ICU text, Wikipedia, and medical literature improved abbreviation expansion, with combined background sources producing the greatest accuracy.The task is challenging because more than 197,000 unique medical abbreviations may each have multiple explanations.
  • A. EHR Information Extraction (IE): Deep learning studies evaluate extraction tasks mainly with precision, recall, and F1, while abbreviation expansion uses accuracy.Results across studies are not directly comparable because some rely on proprietary datasets.
  • B. EHR Representation Learning: Representation-learning methods project sparse medical codes and patient histories into vectors using embeddings, autoencoders, and temporal neural networks.Examples include stacked autoencoders for generalized disease prediction, skip-gram patient vectors for heart-failure onset prediction, and concatenated temporal vectors in DeepCare.

Methods of evaluation for EHR representation learning

EHR representation-learning studies evaluate embeddings either through downstream prediction or directly through coherence, similarity, relatedness, and qualitative visualization. Evaluation remains task-dependent, and proprietary datasets prevent direct comparison across some studies.

  • Methods of evaluation for EHR representation learning: Many representation-learning studies assess learned representations through auxiliary classification tasks, assuming better prediction reflects more robust clinical representations.Reported metrics vary by task, including AUC and precision@k.
  • Methods of evaluation for EHR representation learning: Direct representation evaluation uses measures such as concept coherence, medical concept similarity, and medical relatedness.These methods are used when studies do not include secondary classification tasks and no agreed-upon metric exists.
  • Methods of evaluation for EHR representation learning: Qualitative analysis commonly complements quantitative evaluation by inspecting embedded concepts or patients with t-SNE visualizations and heatmap clusters.The analysis subjectively evaluates similarity between representations in vector space.
  • Methods of evaluation for EHR representation learning: Results from studies sharing tasks or metrics are not directly comparable when they use proprietary datasets.Dataset differences constrain cross-study interpretation even when evaluation setups appear similar.

C. Outcome Prediction

Deep EHR outcome prediction spans static and temporal tasks, using learned representations with diverse neural architectures to forecast diagnoses, interventions, readmissions, complications, and other patient outcomes.

  • Outcome prediction types: Outcome prediction includes static one-time predictions and temporal forecasts based on sequential encounters or specified future intervals.Examples include heart failure prediction from a single encounter and disease-onset prediction within the next six months.
  • Static outcome prediction: The best heart failure model was a standard MLP using embedded patient vectors, outperforming variants trained on raw categorical codes.This comparison directly contrasts learned representations with unprocessed code features.
  • Static outcome prediction: A three-layer autoencoder produced patient representations that improved generalized disease prediction over raw patient features.DeepPatient incorporated clinical codes and clinical-note topics into the representation framework.
  • Static outcome prediction: DeepPatient improved precision@k over raw features for every reported k across ICD9-based disease-diagnosis prediction windows.Other static systems combined learned patient vectors with logistic regression or SVM classifiers for risk and disease prediction.
  • Temporal outcome prediction: GRU, LSTM, CNN, MLP, and modified recurrent architectures were applied to temporal diagnosis, intervention, readmission, pain, complication, mortality, and classification tasks.These systems modeled event sequences, temporal matrices, static-plus-dynamic features, or missingness in clinical time series.
  • Temporal outcome prediction: Deepr exceeded bag-of-codes and logistic-regression baselines for unplanned-readmission prediction, while large input time gaps did not reduce its accuracy.GRU-D networks also improved AUC on two ICD-9 classification and mortality-prediction tasks.

Methods of evaluation for EHR outcome prediction

Outcome-prediction studies primarily use standard classification metrics, but proprietary datasets limit direct comparison across otherwise similar tasks and evaluations.

  • Evaluation metrics: Evaluations commonly report AUC, accuracy, precision, recall, and F1 score across outcome-prediction tasks.The surveyed applications use these metrics for heart-failure prediction, diagnosis classification, readmission, analgesic response, and risk stratification.
  • Cross-study comparison: Results are not directly comparable across studies because many use proprietary datasets, even when tasks and metrics overlap.Dataset differences are identified as a limitation of cross-study interpretation.

D. Computational Phenotyping

Computational phenotyping uses deep learning to discover data-driven disease structure or improve classification within existing definitions, using unsupervised and supervised strategies.

  • Overview: Computational phenotyping seeks richer, data-driven illness descriptions from abundant clinical records rather than relying only on manual disease boundaries.Deep learning is used to discover latent relationships and hierarchical concepts from raw clinical data.
  • Overview: The field divides into discovering new subtypes and improving phenotype classification under existing disease definitions.New-subtype discovery is largely unsupervised, whereas classification improvement is supervised and easier to validate quantitatively.
  • New Phenotype Discovery: Autoencoders support new phenotype discovery by enforcing lower-dimensional representations that encourage latent-structure discovery.Beaulieu-Jones and Greene used a single-layer denoising autoencoder on binary clinical descriptors, with t-SNE visualizations showing clearer label boundaries during training.
  • New Phenotype Discovery: DeepPatient used a deeper autoencoder with diagnoses, medications, procedures, laboratory tests, and LDA-derived clinical-text topics from 704,587 real patient records.It therefore used heterogeneous real-world EHR inputs rather than only simulated clinical descriptors.
  • New Phenotype Discovery: CNNs and RBMs represented patient data as temporal or code-based matrices to extract phenotypes, identify significant patterns, and examine cluster changes over time.Cheng et al.’s four-layer CNN used convolution, max pooling, and softmax prediction; slow-fusion variants performed best over baselines.
  • Improving Existing Definitions: Supervised phenotyping used LSTM models with replicated targets and auxiliary objectives, while ontology-informed Laplacian regularization incorporated structured medical knowledge.The LSTM approach outperformed logistic-regression and MLP baselines, and its LSTM-plus-MLP ensemble performed best.

Methods of evaluation for computational phenotyping

Computational phenotyping is evaluated through external prediction tasks and standard classification metrics, while privacy constraints shape the availability of clinical-text datasets.

  • Evaluation metrics: Phenotype quality is often evaluated indirectly through external prediction tasks using AUC, accuracy, precision, recall, and F1 score.These metrics are applied across simulated disease prediction, chronic-disease onset, and other phenotype-classification tasks.
  • Cross-study comparison: Phenotyping results are not directly comparable across studies because proprietary datasets differ despite shared tasks or evaluation metrics.This limitation restricts straightforward cross-study benchmarking.
  • Data constraints: Clinical notes contain personal health information, requiring released notes to remove identifiers such as names, identification numbers, locations, and dates.These privacy requirements complicate public release of useful clinical datasets.
  • Data constraints: Clinical-text de-identification studies evaluate word-level PHI-category prediction with precision, recall, and F1 score on MIMIC and i2b2 notes.RNN variants outperformed traditional CRF baselines, while location-based PHI was the most difficult category to detect.

VI. INTERPRETABILITY

Deep EHR interpretability remains difficult because deep models learn complex representations, yet several approaches inspect activations, constrain representations, or visualize learned clusters to make them more understandable.

  • Deep models are difficult to interpret because their internal layers do not naturally convey meaning to human observers.This lack of transparency is especially consequential when predictions affect clinical decisions and treatments.
  • Maximum Activation: Maximum-activation analysis identifies inputs that produce the strongest hidden-unit responses, helping assign importance to raw input features.Several reviewed EHR studies adopted this tactic from image processing.
  • Constraints: Non-negativity, sparsity, and ontology-based smoothness constraints expose interpretable disease groups, visit patterns, activating inputs, and medical-concept structure.Med2Vec uses non-negative representations, while eNRBM adds non-negative weights and ontology-distance regularization.
  • Constraints: Sparsity constraints in Lasko et al.’s autoencoder yielded first-layer features interpretable as functional element detectors such as uphill or downhill signal ramps.The associated weights were learned from uric acid sequences for phenotype discovery.
  • Qualitative Clustering: PCA followed by t-SNE enables qualitative inspection of hidden-weight clusters, including separability between case and control phenotypes.Other studies similarly projected clinical-event, patient, and medical-object representations into two dimensions for comparison.
  • Mimic Learning: Interpretable Mimic Learning trains a gradient boosting tree to reproduce a deep network’s probability predictions from raw patient data.The tree serves as a more interpretable surrogate for the deep model’s outputs.

VII. DISCUSSION AND FUTURE DIRECTION

Deep EHR research is an emerging field centered on representation learning for heterogeneous clinical data, with future work targeting richer text processing, unified patient representations, privacy-preserving data sharing, reproducible benchmarks, and interpretability.

  • Most surveyed deep EHR papers were published within the past two years, indicating that the field remains emerging.
  • Data Heterogeneity: EHR representation learning is difficult because patient data spans free text, medical codes, demographics, time series, and medication measurements.This heterogeneity contrasts with the more homogeneous inputs typical of image and natural-language processing systems.
  • Data Heterogeneity: Current work largely uses a divide-and-conquer strategy, separately representing medical codes, patient encounters, and clinical trajectories.Code-based methods include embeddings, clustering, autoencoders, CNNs, GRUs, LSTMs, and recurrent neural networks.
  • Data Heterogeneity: Code-based representations are tractable but omit real-valued laboratory, infusion, and vital-sign measurements.Future research is expected to process these measurements directly rather than relying primarily on billing-oriented controlled vocabularies.
  • Clinical Text: Clinical notes remain an open problem because their unstructured content varies with author shorthand, ordering preferences, and writing style.Deep learning has advanced extraction of concepts, events, entities, relations, abbreviations, and temporal expressions, but further text-based research remains available.
  • Unified Representation: Unified representations that combine all patient-data types could support patient and disease modeling, trajectory prediction, and intervention recommendation.
  • Patient De-identification: Large data requirements and privacy policies make patient de-identification important for transferring information across hospitals and institutions.Reviewed RNN and bidirectional-LSTM approaches improved over traditional lexical-feature de-identification models.
  • Benchmarks: Private institutional datasets limit transparency and reproducibility because there are few universally agreed reference benchmarks for verifying claimed improvements.The authors identify this as a barrier to future model development and a cause of slow advancement.
Loading 1706.03446v2…