Deep learning for temporal data representation in electronic health records: A systematic review of challenges and methodologies

1. Introduction

• Temporal EHR data can reveal longitudinal patterns for risk stratification, disease prognosis, chronic disease management, and adverse drug event detection [3, 4].The review describes structured EHR data as repeated measurements of patients’ conditions over time.Source: §1 · p. 3
• Although temporal EHR data can improve predictive-model discrimination, conventional methods often aggregate time series and lose sequential information [8, 9].The paper notes that temporal information is not often fully utilized because conventional regression and machine-learning methods struggle to extract patterns from repeated variables.Source: §1 · p. 3
• Complex structure and unevenly distributed clinical events create key challenges for temporal EHR modeling: irregularity, heterogeneity, sparsity, and model opacity.The review frames accounting for time-series temporality as an important research question.Source: §1 · p. 3
• The review consolidates deep-learning methods for temporal data representation and evaluates studies by the challenges addressed and their methodologies.It responds to a gap in prior EHR reviews, which did not provide an in-depth systematic account of temporal-data challenges and solutions.Source: §1 · p. 4
• Usability and transferability remain open challenges for future research.The review identifies them as potential topics for further study.Source: §1 · p. 4

2.1 Search strategy and data sources

• The literature search covered five databases—PubMed, EMBASE, IEEE Xplore, ACM Digital Library, and Web of Science—and also searched relevant articles in several prestigious computer science conference proceedings.Source: §2.1 · p. 4

2.2 Inclusion and exclusion criteria

• Following PRISMA, the review included English-language methodological papers applying deep learning to temporal EHR data and excluded reviews, duplicates, irrelevant studies, non-novel applications, and static or unstructured-data studies.Two reviewers screened studies independently and resolved ambiguous selections through consensus.Source: §2.2 · p. 5

2.3 Data extraction, synthesis, and analysis

• The review grouped temporal EHR analysis challenges into data irregularity, data sparsity, data heterogeneity, and model opacity.Included papers were also assessed by time-series nature, methodology, and model implementation.Source: §2.3 · p. 5
• The review extracted time-series components and sparse-code representation, methods and neural-network architectures, plus applications, datasets, evaluation metrics, and comparators.For example, the listed dataset and metric examples were MIMIC [25] and AUROC.Source: §2.3 · p. 5

2.4 Definition of temporal EHR data

• Temporal EHR data were defined at the patient level, with potentially different numbers of time steps across patients and each step represented as a pair including sparse medical-concept data types.Examples of concepts include laboratory test results and vital signs.Source: §2.4 · pp. 5–6

3.1 Selection process and results overview

• In this review, an initial search yielded 1,421 papers; after 495 duplicates were removed, a total of 98 papers were eventually included in the review.Source: §3.1 · p. 6
• Among the included studies, data irregularity was the most frequently studied challenge (n=37), diagnosis the most common temporal variable (n=61), and LSTM the most widely adopted architecture (n=35).Most studies (n=88) used encounters as their time step.Source: §3.1 · p. 6

3.2 Challenges and deep learning solutions

• The review identifies data irregularity, data sparsity, data heterogeneity, and model opacity as four major challenges in temporal EHR data.Source: §3.2 · p. 13

3.2.1 Data irregularity in temporal EHR

• Unequal intervals between EHR encounters complicate time-series modeling but may carry clinical information, as shorter gaps can indicate more frequent examinations and worsening patient condition.Proposed methods use irregular intervals as inputs and tailor neural architectures to time distribution; T-LSTM, for example, incorporates interval information into its gate architecture.Source: §3.2.1 · p. 13
• Fixed-interval resampling can create many missing values, while carry-forward imputation may bias estimates by assuming values remain unchanged.Traditional imputation may also miss temporal patterns; some deep learning approaches instead exploit temporal dependencies or informative missingness.Source: §3.2.1 · pp. 13–14

3.2.3 Data heterogeneity in temporal EHR

• Deep learning can capture complex temporal dynamics for patient similarity and jointly learn multiple clinical outcome prediction tasks.The review describes patient phenotyping and clinical outcomes as two forms of heterogeneity; conventional methods may miss complex phenotypes or be ineffective for multiple outcomes.Source: §3.2.3 · pp. 16–17
• RNN-based multitask frameworks share layers while using task-specific layers for outcomes, and clinical prediction benchmarks were proposed to compare multitask with single-task learning [141].Sharing layers can save computing resources and extract useful temporal information.Source: §3.2.3 · p. 17

3.2.4 Opacity in modeling temporal EHR

• Deep models’ depth and module complexity make temporal EHR behavior difficult to understand, prompting post hoc explanations or calls for inherently interpretable models.Source: §3.2.4 · p. 17
• GBT was used to imitate GRU and achieved superior performance and good interpretability when extracting feature importance.Source: §3.2.4 · p. 17
• Attention can highlight predictive time points, visits, and variables, with larger weights indicating greater importance.Source: §3.2.4 · p. 18
• Other attention variants incorporate feature occurrence frequency to capture infrequent temporal variables or learn time-decay factors to interpret disease progression and changing future-visit risk.Source: §3.2.4 · p. 18
• Hierarchical attention can process temporal EHR sequences while explaining the importance of variables or events.Two-level attention models distinguish medical events from visits, integrating local and global time information.Source: §3.2.4 · pp. 18–19

Discussion

• Nearly 100 deep learning methods were proposed over the last decade, and their number continues to grow, underscoring their potential for temporal EHR analysis.The review also finds that deep learning can represent temporal data while efficiently retaining sequential information.Source: §4 · p. 19
• Only ten included studies used two or more datasets, raising transferability questions; the review identifies freely accessible worldwide large-scale EHR databases as a resource for research.MIMIC was used in 36 of 98 included papers; correction of outliers is mentioned among data collection and processing improvements.Source: §4 · pp. 19–20
• Che et al. demonstrated that newly generated data were of appropriate quality, while Zhang et al. used MetaPred, a novel representation framework based on longitudinal data, to take advantage of limited data.Transfer learning has the potential to relieve healthcare data shortages and improve model generalizability.Source: §4 · p. 20
• Post-hoc explanations may reflect learned artifacts rather than data-derived knowledge, raising usability concerns, while clinicians prefer inherently interpretable models [143, 159, 160].The review suggests integrating deep learning with intrinsically interpretable methods to facilitate real-world validation.Source: §4 · pp. 20–21
• CNN (n=9), RNN (n=18), and LSTM (n=35) were the most common architectures; Transformers use self-attention, while MLP-Mixer offers a simpler alternative for temporal EHR data.Transformers show potential for temporal EHR representation [101, 105]; MLP-Mixer uses separate MLPs for per-location features and spatial information.Source: §4 · p. 21
• The review did not assess all clinical applications or report model performance, could not recommend an overall best method, and excluded unstructured data and preprints.Its scope was methodological and limited to structured temporal EHR data; it notes that preprint exclusion may have missed new evidence while preserving peer-reviewed results.Source: §4 · p. 21

• The review covers primary issues in temporal EHR analysis and state-of-the-art deep learning solutions, which address representation challenges to some extent.Source: §5 · p. 22
• Future research may focus on improving model transferability.Source: §5 · p. 22
• Future research may incorporate clinical domain knowledge into study design and enhance model interpretability to facilitate clinical implementation.Source: §5 · p. 22

Abbreviation:

• Model abbreviations include ANN (Artificial Neural Network), CNN (Convolutional Neural Network), GNN (Graph Neural Network), LSTM (Long Short-Term Memory), and TCN (Temporal Convolutional Network).The glossary also defines Bi-LSTM, Bi-RNN, GRU, RNN, TANN, and other model terms.Source: Abbreviation · pp. 35–36
• Evaluation abbreviations include AUC (Area Under Curve), F1 score, MAE (Mean Absolute Error), NMI (Normalized Mutual Information), PPV (Positive Predictive Value), and RMSE (Root Mean Squared Error).Other defined measures include sensitivity, specificity, and ROC and PRC curves.Source: Abbreviation · pp. 35–36
• The glossary also defines methods and clinical terms, including BOW (Bag of Words), PCA (Principal Component Analysis), SVM (Support Vector Machine), SIRS, and SOFA.It includes additional methods, scores, and measures such as Hawkes Process, Modified Early Warning Score, and TPR (True Positive Rate).Source: Abbreviation · pp. 35–36

Appendix

• The PubMed query uses three AND-linked groups covering deep-learning methods, temporal representation, and EHR terminology.The groups include “deep learning,” “neural network,” “CNN,” “RNN,” and “LSTM”; “embedding,” “time series,” “temporal,” “sequential,” and “attention”; and “electronic health record,” “EHR,” and “EMR” variants.Source: Appendix · p. 37

IEEE Xplore

• The query paired deep-learning terms with representation and time-series terms, then with electronic-health-record terms.Terms included “deep learning,” “neural network,” CNN, RNN, LSTM, “embed*,” “representation,” “time series,” “sparse,” “temporal,” “concept,” “sequential,” “attention,” and EHR/EMR variants.Source: IEEE Xplore · p. 37

ACM Digital Library

• The ACM Digital Library query grouped “deep learning,” neural network, deep, CNN, RNN, and LSTM with OR.Source: ACM Digital Library · p. 37
• The ACM query paired those terms with OR-grouped embedding, representation, time-series, sparse, temporal, concept, sequential, and attention terms, then EHR synonyms.Source: ACM Digital Library · p. 37

Web of Science

• Web of Science used TS fields to OR-group deep learning, neural network, deep, CNN, RNN, and LSTM terms.Source: Web of Science · p. 37
• Its TS search also OR-grouped embedding, representation, time-series, sparse, temporal, concept, sequential, and attention terms with EHR synonyms.Source: Web of Science · p. 37

Embase

• The Embase query grouped “deep learning,” neural network, deep, CNN, RNN, and LSTM with OR.Source: Embase · p. 37
• The Embase query paired those terms with OR-grouped embedding, representation, time-series, sparse, temporal, concept, sequential, and attention terms, then EHR synonyms.Source: Embase · p. 37