The Taxonomies, Training, and Applications of Event Stream Modelling for Electronic Health Records

Mingcheng Zhu; Yu Liu; Zhiyao Luo; Tingting Zhu

eess.SPMar 14, 2026

The Taxonomies, Training, and Applications of Event Stream Modelling for Electronic Health Records

Authors: Mingcheng Zhu, Yu Liu, Zhiyao Luo, Tingting Zhu

Organizations: University of Oxford, UK

Abstract

The widespread adoption of electronic health records (EHRs) enables the acquisition of heterogeneous clinical data, spanning lab tests, vital signs, medications, and procedures, which offer transformative potential for artificial intelligence in healthcare. Although traditional modelling approaches have typically relied on multivariate time series, they often struggle to accommodate the inherent sparsity and irregularity of real-world clinical workflows. Consequently, research has shifted toward event stream representation, which treats patient records as continuous sequences, thereby preserving the precise temporal structure of the patient journey. However, the existing literature remains fragmented, characterised by inconsistent definitions, disparate modelling architectures, and varying training protocols. To address these gaps, this review establishes a unified definition of EHR event streams and introduces a novel taxonomy that categorises models based on their handling of event time, type, and value. We systematically review training strategies, ranging from supervised learning to self-supervised methods, and provide a comprehensive discussion of applications across clinical scenarios. Finally, we identify open critical challenges and future directions, with the aim of clarifying the current landscape and guiding the development of next-generation healthcare models.

Figures & tables

	Shchur et al., (2021)	Amirahmadi et al., (2023)	Zólyomi et al., (2025)	This Work
Modelling Taxonomy
Time Modelling	✓		✓	✓
Type Modelling	✓	✓	✓	✓
Value Modelling		✓	✓	✓
Training Strategy
Supervised	✓	✓		✓

Table 1. Comparison of existing reviews on EHR event stream. ✓indicates coverage.

Figure 1. The flow diagram of the literature searching strategy.

Figure 2. Illustration of the event stream and multivariate time series representations of EHR. (a) Raw EHR Data: Clinical events spanning multiple heterogeneous types (e.g., vital signs, lab tests, medications, procedures, and diagnoses) are recorded at irregular, continuous timestamps. (b) Event Stream: All extracted clinical events are arranged chronologically into a sequence, with co-occurring events assigned an arbitrary relative order. (c) Multivariate Time Series: The data is converted into a structured grid by aligning variables to fixed time intervals (e.g., hourly windows). To reduce sparsity, features with a high missing rate are filtered out, and missing data for the retained variables are imputed, resulting in a dense, regularly sampled matrix.

Figure 3. Overview of event stream modelling. This overview categorises the field into three primary components, each highlighted using a distinct colour. Event stream modelling details the modelling methods for different components of an event. Training strategy outlines different learning paradigms for event stream models, including supervised and self-supervised methods. The application focuses on the different healthcare scenarios, including critical care, emergency department, hospital and secondary care, and primary care.

Figure 4. Illustration of self-supervised learning methods for EHR event stream modelling. The four main approaches are: (a) Next-token prediction, where the model learns to predict the subsequent event in the sequence given the preceding events. (b) Masked reconstruction, where events in the input sequence are deliberately hidden (masked), and the model is trained to reconstruct the complete sequence. (c) Skip-gram (

n=1

), where the model uses a single target event (e.g.,

t_{2}

) to predict its surrounding context events (e.g.,

t_{1}

and

t_{3}

). (d) Contrastive learning, which trains the model to maximise the embedding similarity between an original sequence and a positive pair that are generated via event swapping of co-occurring events, while minimising the similarity against a negative pair that is generated via event replacement.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Reference	Event Stream Modelling		Training Strategy		Applications
Reference	Time	Type & Value	Supervised	Self-Sup.	Scenario	Task	Metric	Dataset
Reference	Event Stream Modelling		Training Strategy		Applications
Reference	Time	Type & Value	Supervised	Self-Sup.	Scenario	Task	Metric	Dataset
2016 Year:
( Cheng et al.,, 2016 )	Feature	Type Only	$\checkmark$	-	Hospital	Risk Prediction	AUC	IBM Warehouse
( Henriksson et al.,, 2016 )	Feature	Type Only	-	Skip-gram	Hospital	Risk Prediction	ACC, AUC	Stockholm EPR

Appendix

Table 2. Summary of EHR event stream modelling methods with supervised learning.

Electronic Health Records

Electronic Health Records (EHRs) are longitudinal digital patient records studied for clinical data integration, analysis, and decision support.

See all 111 papers

Structure-Aware Set Transformers: Temporal and Variable-Type Attention Biases for Asynchronous Clinical Time Series
Feb 18, 2026Joohyung Lee, Kwanhyung Lee, Changhun Kim +1Set Transformer Irregular Time-Series Modeling
AITRICS · Korea Advanced Institute of Science and Technology (KAIST)
EHR2Trace: Auditable EHR Data Infrastructure for Patient World Models and Clinical Agents
Sep 18, 2026Xinye Yang, Yuli Wang, Cheng Ting Lin +1Electronic Health Records Data Provenance
Department of Radiology University of Colorado Anschutz School of Medicine Aurora, CO, USA · Department of Radiology Johns Hopkins University School of Medicine Baltimore, MD, USA
FeatEHR-LLM: Leveraging Large Language Models for Feature Engineering in Electronic Health Records
Apr 24, 2026Hojjat Karami, David Atienza, Jean-Philippe Thiran +1Healthcare Electronic Health Records
1École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland

Feb 18, 2026J/K move · Enter open · S save

Structure-Aware Set Transformers: Temporal and Variable-Type Attention Biases for Asynchronous Clinical Time Series

Joohyung Lee, Kwanhyung Lee, Changhun Kim, Eunho Yang

AITRICS · Korea Advanced Institute of Science and Technology (KAIST)

Set Transformer Irregular Time-Series Modeling Time Series Transformers Longitudinal EHR Modeling Transformer Temporal Attention

Open paper AskPDF arXiv

Electronic health records (EHR) are irregular, asynchronous multivariate time series. As time-series foundation models increasingly tokenize events rather than discretizing time, the input layout becomes a key design choice. Grids expose time

\times

variable structure but require imputation or missingness masks, risking error or sampling-policy shortcuts. Point-set tokenization avoids discretization but loses within-variable trajectories and time-local cross-variable context (Fig.1). We restore these priors in STructure-AwaRe (STAR) Set Transformer by adding parameter-efficient soft attention biases: a temporal locality penalty

-|Δt|/τ

with learnable timescales and a variable-type affinity

B_{s_i,s_j}

from a learned feature-compatibility matrix. We benchmark 10 depth-wise fusion schedules (Fig.2). On three ICU prediction tasks, STAR-Set achieves AUC/APR of 0.7158/0.0026 (CPR), 0.9164/0.2033 (mortality), and 0.8373/0.1258 (vasopressor use), outperforming regular-grid, event-time grid, and prior set baselines. Learned

τ

and

B

provide interpretable summaries of temporal context and variable interactions, offering a practical plug-in for context-informed time-series models.

Abstract

Electronic Health Records

Explore similar work

Structure-Aware Set Transformers: Temporal and Variable-Type Attention Biases for Asynchronous Clinical Time Series

EHR2Trace: Auditable EHR Data Infrastructure for Patient World Models and Clinical Agents

FeatEHR-LLM: Leveraging Large Language Models for Feature Engineering in Electronic Health Records