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

Xinye Yang; Yuli Wang; Cheng Ting Lin; Harrison Bai

cs.LGSep 18, 2026

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

Authors: Xinye Yang, Yuli Wang, Cheng Ting Lin, Harrison Bai

Organizations: 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

Abstract

Patient world models and clinical agents aim to predict changes in patients' health and support clinical work. Developing these systems requires reliable histories of patient conditions, treatments, and the information available at each decision. Electronic health records (EHRs) contain these histories, but differences in how events are recorded make them difficult to use consistently. We present EHR2Trace, a system that converts EHRs from different sources into traceable patient events for model training and evaluation. It links events to source records, separates event time from information availability, and distinguishes medication orders, dispensing, and administration. A shared event representation supports both OMOP and MEDS exports, with automated validation and reproducible builds. Across three clinical datasets, EHR2Trace converted 846.4 million events, with every applicable check passing except one unit-consistency check on MIMIC-IV, and detected all 28 injected faults. A controlled prediction experiment showed that assigning later diagnoses to admission time substantially inflated measured performance, and that a model trained on such data lost accuracy when deployed on histories filtered by availability. EHR2Trace provides a reusable data foundation for patient world models and clinical agents, helping researchers inspect patient histories, check conversion decisions, and evaluate models with explicit data rules.

Figures & tables

Panel of Figure 1 — Figure 1: (a) EHR2Trace forms the data layer for patient world models and clinical agents. Dashed boxes show downstream stages. (b) The system converts source records into shared patient events, exports OMOP and MEDS, and stores audit records alongside the data. Automated checks validate the saved outputs. Counts show the number of checks in each group.

event_kind	hours from admission		source_name , and the detail an action carries	at $\tau$
	event	avail.
Subject 1, one encounter of 75 h; decision time $\tau$ at +44.08 h
drug_admin	4.37	4.38	CeftriaXONE 1 gm · Administered	✓
visit	7.98	·	EW EMER. AVAILABILITY_ASSUMED	✓
drug_order	12.32	·	Vitamin D 1,000 Unit Tablet 1000 · PO/NG · AVAILABILITY_ASSUMED	✓
drug_dispense	12.32	·	Docusate Sodium PO/NG · Discontinued via patient discharge · AVAILABILITY_BEFORE_EVENT	✓

Table 1: Two patient histories from the public MIMIC-IV demonstration subset [ 6 ] . Rows show canonical events in event-time order. A dot means event and availability times agree. Events marked “withheld” occurred before decision time

\tau

but became available afterward. Orders, dispensing, and administration remain separate records, with dose, route, status, and quality flags preserved. The NOT_ADMINISTERED flag records a treatment that was not given. Omitted events appear as counts.

Measure	JHU-CTPE	MIMIC-IV	CU-CTPA
Source rows	75,113,066	836,795,494	16,682,059
Identities encountered	22,982	364,673	127,955
Subjects with canonical events	22,980	364,673	127,955
Canonical events	33,396,779	799,153,396	13,880,372
Event-to-source links	101,522,252	801,049,552	16,385,410
Quarantine entries	750,121	1,074,435	232,123

Table 2: Conversion scale and validation results. Canonical and MEDS event counts match. Patient counts differ because some identities have no events or do not meet OMOP eligibility rules. A source row can generate multiple quarantine entries.

	JHU-CTPE		MIMIC-IV		CU-CTPA
Domain	Rows	Mapped	Rows	Mapped	Rows	Mapped
Condition	3,637,437	99.8%	6,451,372	98.9%	5,861,753	100.0%
Drug	11,713,070	86.7%	74,731,397	74.5%	4,206,761	99.1%
Measurement	15,911,634	97.1%	613,248,151	25.7%	982,655	100.0%
Procedure	572,119	100.0%	1,710,192	52.6%	719,423	100.0%
Observation	678,871	96.6%	11,531,859	9.8%	35	100.0%

Table 3: OMOP concept coverage by domain, weighted by row count. “Mapped” indicates the presence of a standard concept identifier.

Vital-status timestamps
Audit measure	JHU-CTPE	MIMIC-IV	CU-CTPA
Untimed VITAL_STATUS events after conversion	0	0	0
Medication action types and details
Dispensing records kept separate from administration	—	18,968,537	—
cabinet records among them	—	1,189,422	—
Non-medication requests stored as service orders	—	28,916,331	—

Table 4: Time and medication-action audit across all canonical events. Dispensing records indicate drug supply, while administration records describe whether and how a drug was given. Assumed availability is the share of timed events that default to event time because the source lacks an availability timestamp. These events carry AVAILABILITY_ASSUMED .

Property	EHR2Trace	mimic-iv-demo-omop	mimic-iv-demo-meds
Common data models produced	OMOP and MEDS	OMOP	MEDS
Source modules read on the subset	hospital, ICU	hospital, ICU	hospital, ICU
Output rows (OMOP) / events (MEDS)	897,487 / 971,545	423,057	916,166
Rows with a standard concept	18.9%	98.3%	–
Events linked to a concept	19.8%	–	13.5%
Source file and row recoverable	every event (972,567 links)	not carried	source keys on 95.2%, table not named

Table 5: Information retained by three conversions of the 100-patient MIMIC-IV demonstration subset. Results were measured from EHR2Trace outputs and the published baseline files. Input modules are listed for each conversion.

Panel of Figure 2 — Figure 2: Detection of injected faults, grouped by the affected data layer. DQD was run on nine faults that reach OMOP. The full suite covers all 28 faults.

Horizon	A $\rightarrow$ A	C $\rightarrow$ C	C $\rightarrow$ A	Evaluation inflation	Training deficit	AUPRC
	filtered	backdated	cross-rule	C $\rightarrow$ C $-$ C $\rightarrow$ A	A $\rightarrow$ A $-$ C $\rightarrow$ A	A $\rightarrow$ A / C $\rightarrow$ A
6 h	0.829	0.965	0.642	0.323 [0.295, 0.352]	0.188 [0.161, 0.215]	0.222 / 0.071
12 h	0.839	0.969	0.660	0.309 [0.280, 0.338]	0.179 [0.152, 0.206]	0.239 / 0.088
24 h	0.854	0.967	0.669	0.298 [0.269, 0.328]	0.185 [0.156, 0.214]	0.253 / 0.107
48 h	0.863	0.963	0.694	0.269 [0.241, 0.297]	0.169 [0.145, 0.194]	0.310 / 0.142

Table 6: Retrospective evaluation under matched and mismatched training–test rules. A denotes availability filtering and C diagnosis backdating. Arrows indicate training

\rightarrow

test rules. Performance columns report held-out AUROC, with paired percentile bootstrap intervals from 2,000 resamples for the two contrasts. The final column reports AUPRC for A

\rightarrow

A and C

\rightarrow

Panel of Figure 3 — Figure 3: Held-out performance under three time rules for the mortality-proxy task. Backdating diagnoses exposes early predictions to later information. Panel (c) shows AUROC increases relative to availability filtering on a logarithmic scale. Prediction horizons use a base-2 logarithmic axis. In panel (a), the AUROC axis begins at 0.80, and error bars show bootstrap intervals from 2,000 resamples of held-out patients.

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

Learning to Select Source-Traceable Evidence for Language-Model Prediction from Irregular Clinical Time Series
May 19, 2026Kwanhyung Lee, Juhwan Choi, Jongheon Kim +5Irregular Time-Series Modeling Healthcare
Kim Jaechul Graduate School of AI, KAIST · AITRICS
DT-Transformer: A Foundation Model for Disease Trajectory Prediction on a Real-world Health System
May 14, 2026Yunying Zhu, Andrew R Weckstein, Kueiyu Joshua Lin +1Clinical Prediction Time Series Transformers
Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA · Harvard T.H. Chan School of Public Health, Harvard University, Boston, MA, USA · Kempner Institute for the Study of Natural and Artificial Intelligence, Harvard University, MA, USA +2
Pretraining EHR Foundation Models with Patient-Aware Sampling
Jul 24, 2026Joshua Placidi, Yuxuan Liu, Jinpei Han +2EHR Foundation Models Longitudinal EHR Modeling
Dept. of Computing, Imperial College London, London, United Kingdom · Dept. of Bioengineering, Imperial College London, London, United Kingdom · UKRI Centres in AI for Health, United Kingdom +1

May 19, 2026J/K move · Enter open · S save

Learning to Select Source-Traceable Evidence for Language-Model Prediction from Irregular Clinical Time Series

Kwanhyung Lee, Juhwan Choi, Jongheon Kim, Joohyung Lee+4

Kim Jaechul Graduate School of AI, KAIST · AITRICS

Irregular Time-Series Modeling Healthcare Evidence Selection Clinical Outcome Prediction Longitudinal EHR Modeling

Open paper AskPDF arXiv

Numerical time-series models effectively process irregular electronic health record (EHR) trajectories, but do not expose which temporal patterns support each prediction as readable evidence. Existing text-based interfaces either serialize observations, preserving source traceability but offering limited clinical interpretation, or generate patient-level summaries that improve readability but can obscure links to source measurements. We introduce STEP-CTS (Source-Traceable Evidence for Prediction from Clinical Time-Series), which learns to select source-traceable text evidence for a language-model predictor. Multi-scale window statistics of each trajectory are verbalized as sets of deterministic threshold predicates, each set linked to its source window (e.g., "[5-7 h]: last temperature at least 38C"). An offline LLM, run once per unique predicate set without access to patient records, the prediction task, or outcome labels, attaches clinical concepts such as "fever" to their supporting predicates and abstains when none applies. Each predicate set, together with its concepts, forms an evidence unit. A learned Evidence Selector selects a fixed-size subset of the evidence units, which a pretrained clinical language-model encoder reads to make the prediction, with no patient-level text generation. Across three ICU benchmarks, STEP-CTS outperforms evaluated text-based baselines, improving AUPRC over the strongest by 5.1, 3.7, and 11.7 percentage points on P2012, MIMIC-III, and P2019, respectively, and is competitive with dedicated numerical time-series models. Ablations show that clinical concepts and learned evidence selection each contribute to predictive performance. In a blinded clinician study, the selected evidence is rated as traceable as deterministic serialization, near the ceiling of the scale, and above generated summaries on clinical interpretation.

Abstract

Electronic Health Records

Explore similar work

Learning to Select Source-Traceable Evidence for Language-Model Prediction from Irregular Clinical Time Series

DT-Transformer: A Foundation Model for Disease Trajectory Prediction on a Real-world Health System

Pretraining EHR Foundation Models with Patient-Aware Sampling