Source-linked AI summary
Deep learning for temporal data representation in electronic health records: A systematic review of challenges and methodologies
Feng Xie, Han Yuan, Yilin Ning, Marcus Eng Hock Ong, Mengling Feng, Wynne Hsu, Bibhas Chakraborty, Nan Liu
TL;DR
Temporal EHR data pose challenges for modeling the whole time series. This review summarized those challenges and deep learning solutions, finding that the challenges are addressed to some extent while several challenges remain.
Problem
Variation between encounters creates challenges for modeling the whole temporal EHR time series.
Method
The review summarized temporal EHR challenges and discussed how deep learning solutions could help overcome them.
Results
Four major challenges were identified: data irregularity, sparsity, heterogeneity, and model opacity; these challenges were addressed to some extent.
Takeaways & Limitations
Deep learning techniques show promising results for temporal EHR representation, but comprehensive solutions remain needed.
Takeaways & Limitations
The study sought to understand the current state of the field, and several challenges remain despite promising deep learning results.
Abstract
from arXiv · showhide
Objective: Temporal electronic health records (EHRs) can be a wealth of information for secondary uses, such as clinical events prediction or chronic disease management. However, challenges exist for temporal data representation. We therefore sought to identify these challenges and evaluate novel methodologies for addressing them through a systematic examination of deep learning solutions. Methods: We searched five databases (PubMed, EMBASE, the Institute of Electrical and Electronics Engineers [IEEE] Xplore Digital Library, the Association for Computing Machinery [ACM] digital library, and Web of Science) complemented with hand-searching in several prestigious computer science conference proceedings. We sought articles that reported deep learning methodologies on temporal data representation in structured EHR data from January 1, 2010, to August 30, 2020. We summarized and analyzed the selected articles from three perspectives: nature of time series, methodology, and model implementation. Results: We included 98 articles related to temporal data representation using deep learning. Four major challenges were identified, including data irregularity, data heterogeneity, data sparsity, and model opacity. We then studied how deep learning techniques were applied to address these challenges. Finally, we discuss some open challenges arising from deep learning. Conclusion: Temporal EHR data present several major challenges for clinical prediction modeling and data utilization. To some extent, current deep learning solutions can address these challenges. Future studies can consider designing comprehensive and integrated solutions. Moreover, researchers should incorporate additional clinical domain knowledge into study designs and enhance the interpretability of the model to facilitate its implementation in clinical practice.
4 Saw Swee Hock School of Public Health, National University of Singapore,
The section lists affiliations with the Saw Swee Hock School of Public Health and the National University of Singapore.
- The listed affiliation is the Saw Swee Hock School of Public Health at the National University of Singapore.
- The institution is located in Singapore.
- The National University of Singapore is identified as the institutional affiliation.
7 Department of Statistics and Data Science, National University of Singapore,
The section lists affiliations with National University of Singapore units and Health Services Research Centre, Singapore Health Services.
- The listed affiliations include the Department of Biostatistics and Bioinformatics at Duke University.
- The Health Services Research Centre is identified as part of Singapore Health Services.
- The Programme in Health Services and Systems Research is listed among the affiliations.
1. Introduction
The introduction explains why temporal EHR data matter for secondary uses, while highlighting representation challenges that motivate deep learning and this systematic review. It frames the review around technical challenges, proposed methods, and remaining research needs.
- Temporal EHR data preserve longitudinal patient information and may support risk stratification, disease prognosis, and chronic disease management.Examples include chronic kidney disease prediction and adverse drug event detection.
- Conventional approaches often aggregate repeated measurements, losing sequential information and temporal dynamics.The introduction identifies accounting for time-series temporality as an important research question.
- Temporal EHR data pose challenges involving irregularity, heterogeneity, sparsity, and model opacity.These challenges arise from complex structure and unevenly distributed clinical events.
- Deep learning architectures such as RNNs, LSTMs, and GRUs have been proposed for temporal EHR representation.The introduction describes these architectures as potentially suited to temporal data and reports their use in many applications.
- The review addresses a gap by systematically consolidating deep learning methods and examining the challenges of temporal EHR representation.It evaluates selected studies from the perspectives of primary challenges and methodologies addressing them.
- Usability and transferability remain open challenges for future research.
2. Methods
The review used a systematic search and screening process to identify methodological studies on deep learning for temporal structured EHR data. Included studies were categorized by technical challenge and analyzed across time-series characteristics, methodology, and implementation.
- The review searched PubMed, EMBASE, IEEE Xplore, ACM Digital Library, and Web of Science, supplemented by selected conference proceedings.The conference search included NeurIPS, IJCAI, and AAAI proceedings.
- The search targeted deep learning studies on temporal EHR representation published from January 1, 2010, to August 30, 2020.
- Studies were screened under PRISMA, with methodological papers included and reviews, duplicates, static-data studies, unstructured-data studies, and non-novel application papers excluded.Two reviewers screened studies independently and resolved ambiguous cases through discussion with a third researcher.
- The review identified four main technical challenges: data irregularity, data sparsity, data heterogeneity, and model opacity.
- Included papers were evaluated by time-series nature, methodology, and model implementation.Extracted information covered time-series components, sparse-code representation, methods, addressed challenges, architectures, datasets, metrics, and comparators.
- The review defined temporal EHR data as patient-specific sequences whose patients may have different numbers of time steps.Each time step includes sparse medical concepts such as laboratory results and vital signs.
3. Results
The review included 98 papers and characterized their publication trends, temporal variables, architectures, and deep-learning approaches to major temporal EHR challenges. Most studies used encounters as time steps, while irregularity, heterogeneity, sparsity, and opacity motivated specialized representations and interpretability mechanisms.
- Selection process: 1,421 records were identified initially; after removing 495 duplicates, 926 underwent title and abstract screening.
- Selection process: 246 records were excluded for irrelevance to EHR, 243 for using unstructured data, and 98 papers were ultimately included.
- Publication patterns: The volume of included articles increased significantly from 2010 to 2020, with ACM, KDD, IEEE, and biomedical informatics venues among the prominent publication outlets.
- Study characteristics: Data irregularity was the most frequently studied challenge, diagnosis was the most common temporal variable, and LSTM was the most widely adopted architecture.
- Study characteristics: 88 of 98 studies used encounters such as episodes, visits, or admissions as the time step, while others used fixed time windows.
- Methodological approaches: Deep-learning solutions modeled irregular intervals directly as inputs or customized neural-network architectures, while attention mechanisms represented variable importance for interpretation.
4. Discussion
This review consolidates deep learning methods for temporal EHR representation and frames their progress around four challenges: irregularity, sparsity, heterogeneity, and opacity. It also identifies persistent concerns about data quality, generalizability, interpretability, and clinical applicability.
- Deep learning solutions address temporal representation through sequential architectures, attention mechanisms, transfer learning, data augmentation, and alternative frameworks such as Transformer and MLP-Mixer.Transfer learning can use large public datasets before conditioning models on local data, while Transformer relies on self-attention and MLP-Mixer avoids convolutions and attention.
- The review identifies data irregularity, data sparsity, data heterogeneity, and model opacity as four major challenges in temporal EHR analysis.
- Nearly one hundred deep learning methods have been proposed for temporal EHR data, with LSTM the most widely adopted architecture among reviewed studies.LSTM was used in 35 studies; CNN and RNN were used in 9 and 18, respectively.
- Data limitations remain substantial because most studies analyze only one dataset, while only ten use two or more datasets, raising questions about transferability and generalizability.MIMIC was the most frequently used dataset, appearing in 36 of 98 included studies.
- Interpretability remains problematic because post-hoc explanations may reflect artifacts rather than knowledge derived from the data, limiting usability in healthcare settings.The review contrasts these approaches with intrinsically interpretable models that clinicians can understand more naturally.
- The review cannot recommend an overall best method because included studies differ in preprocessing, parameter tuning, and clinical tasks, and it focuses on structured temporal EHR data.It also does not summarize all clinical applications or report the performance of deep learning solutions.
5. Conclusion
The review comprehensively examined major issues in representing temporal EHR data and evaluated state-of-the-art deep learning solutions. It concludes that current methods address these challenges to some extent, while future work should improve transferability, clinical knowledge integration, and interpretability.
- The review comprehensively examined primary issues in analyzing temporal EHR data.
- It presented state-of-the-art deep learning solutions for temporal EHR data representation.
- Current solutions address various challenges arising from temporal EHR data representation to some extent.
- Future research may focus on model transferability and incorporating clinical domain knowledge into study design.
- Future research may also enhance model interpretability to facilitate clinical implementation.
Abbreviation:
The abbreviation list defines neural-network architectures, evaluation metrics, statistical methods, and clinical measures used in the paper.
- ANN denotes Artificial Neural Network, while CNN denotes Convolutional Neural Network and RNN denotes Recurrent Neural Network.
- AUC denotes Area Under Curve, MAE denotes Mean Absolute Error, MSE denotes Mean Squared Error, and RMSE denotes Root Mean Squared Error.
- ROC denotes Receiver Operating Characteristic Curve, PRC denotes Precision Recall curve, and PPV denotes Positive Predictive Value.
- SIRS denotes systemic inflammatory response syndrome, and SOFA denotes Sequential Organ Failure Assessment.
Appendix
The appendix presents database-specific search strategies combining deep-learning terms, temporal-representation terms, and electronic-health-record terms.
- The search strategies combine deep learning or neural-network terms with CNN, RNN, and LSTM variants.
- Representation-related search terms include embedding, representation, time series, sparse, temporal, concept, sequential, and attention.
- EHR-related search terms include electronic health record, EHR, electronic medical record, and EMR variants.
- The Web of Science strategy expresses the three concept groups through repeated TS fields.