R-GEAN: Regimen-Guided Edit Action Network for Within-Admission Medication Change Prediction

Regan Mahat; Mansu Kim

cs.AISep 19, 2026

R-GEAN: Regimen-Guided Edit Action Network for Within-Admission Medication Change Prediction

Authors: Regan Mahat, Mansu Kim

Organizations: Gwangju Institute of Science and Technology (GIST), Gwangju, Republic of Korea

Abstract

The medications prescribed to a patient often change during a hospital admission as clinicians start, stop, or continue therapies. We study whether models can predict which medication classes are added or removed between 24 hours after admission and discharge. Metrics that compare the complete discharge regimen can reward models for copying medications that remain unchanged, even when they identify no actual changes. We therefore introduce a leakage-controlled benchmark that predicts net ATC3 additions and removals using only prior completed admissions and information available within the first 24 hours of the current admission. Addition candidates are classes not active at 24 hours, whereas removal candidates are classes active at that time. We also introduce R-GEAN, an asymmetric candidate-scoring network with independent addition and removal predictors. Across 240,480 admissions from 82,286 patients, R-GEAN achieves the highest predefined summary of addition, removal, changed-regimen, and action-pattern performance, termed the edit composite (0.464), compared with 0.435 for the strongest primary comparator. Reimplemented RETAIN, GAMENet, and MICRON baselines obtain 0.428, 0.420, and 0.288, respectively. R-GEAN's advantage is concentrated in correctly identifying medication classes no longer active at discharge, while rare additions and admissions with multiple medication changes remain difficult. Rankings based on micro-F1 over the reconstructed discharge regimen and the edit composite correlate weakly across the evaluated models (Spearman r = 0.20). The continuation baseline achieves the highest complete-regimen score despite predicting no additions or removals. These results show that complete-regimen and edit-level evaluation measure different aspects of medication prediction. The benchmark evaluates observed prescribing changes, not treatment appropriateness

Figures & tables

Figure 1: The model uses prior history and information available within the first 24 hours to predict medication classes added to or removed from the observed 24-hour regimen by discharge. The two predictions are combined to reconstruct the discharge regimen. No information recorded after the 24-hour boundary is available to the model.

Property	Value
Source	MIMIC-IV v3.1
Patients	82,286
Admission-level examples	240,480
Target vocabulary (ATC3)	78 classes
Train examples / patients	167,418 / 57,627
Validation examples / patients	24,030 / 8,215

Table 1: Benchmark and split statistics. The splits are patient-disjoint. Each test admission belongs to one of seven mutually exclusive medication-change patterns: continue 9,580, add 9,879, remove 7,818, switch 7,389, multi-edit 13,560, empty-to-nonempty 749, and nonempty-to-empty 57.

Figure 2: R-GEAN architecture. The addition and removal encoders receive 117- and 101-dimensional admission inputs, respectively, and produce 256-dimensional admission representations. Each representation is combined with a 32-dimensional medication-specific laboratory vector, a 16-dimensional exposure vector, and an 83-dimensional medication-class representation. The resulting 387-dimensional vector is scored by a

387\!\rightarrow\!128\!\rightarrow\!1

MLP. The addition branch evaluates classes absent from the observed 24-hour regimen (

\mathcal{C}^{+}=\mathcal{V}\setminus M_{A}

), whereas the removal branch evaluates classes present in it (

\mathcal{C}^{-}=M_{A}

). Their predictions reconstruct the discharge regimen as

\hat{M}_{T}=(M_{A}\setminus\hat{Y}^{-})\cup\hat{Y}^{+}

Contextual (non-learned) baselines
Model	Add F1 $\uparrow$	Remove F1 $\uparrow$	Changed Jacc. $\uparrow$	Action mF1 $\uparrow$	Composite $\uparrow$
continuation	0.0000	0.0000	0.6347	0.0467	0.1634
frequency	0.1942	0.5014	0.5968	0.3364	0.4012
Public models, reimplemented under this benchmark
RETAIN-FullSet	0.2687	0.4671	0.6209	0.3811	0.4275
RETAIN-Edit	0.0959	0.4780	0.5444	0.3608	0.3492

Table 2: Test results for the primary seed (49,032 admissions; 16,444 patients). Bold and underlined values indicate the best and second-best result in each column. The continuation baseline attains the highest changed-admission Jaccard despite predicting no additions or removals. The composite is the predefined summary of edit performance described in Equation 6 .

Variant	$\Delta\mathrm{F1}^{+}$	$\Delta\mathrm{F1}^{-}$	$\Delta S$
Add. w/o DDI	$-0.0786$	—	$-0.0409$
Add. w/o history	$-0.0299$	—	$-0.0224$
Add. w/o transition	$-0.0125$	—	$-0.0108$
Add. w/o exposure	$-0.0227$	—	$-0.0085$
Rem. w/o exposure	—	$-0.0772$	$-0.0293$
Shared predictor (rerun)	$-0.0570$	$-0.0077$	$-0.0298$

Table 3: Changes in test performance after removing individual components. Full R-GEAN obtains

\mathrm{F1}^{+}=0.2509

\mathrm{F1}^{-}=0.6013

, and

S=0.4643

. When one branch is modified, the other is held fixed. Complete results are provided in the supplement.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Result
Raw prescription rows	20,292,611
Records linked to eligible admissions	15,562,624
RxNorm entity coverage	64.6%
ATC coverage (by record)	88.6%
Resolved eligible records (after exclusions)	99.4%
Final ATC3 vocabulary	78 classes

Appendix

Table 4: Summary of medication mapping and benchmark construction. Coverage percentages use the denominators defined in the text.

Variable	Recurrent-care	Single-adm.	SMD
Admissions	240,480	109,899	—
Mean age (years)	62.9	56.6	0.33
Mean length of stay (d)	6.1	5.2	0.12

Appendix

Table 5: Cohort selection relative to otherwise-eligible single-admission adults (standardized mean differences).

Input block	Availability rule	Branch	Verdict
Prior-admission and admission features	Available by admission	Add, Rem	Pass
Anchor regimen (multi-hot)	Active at the 24-hour landmark	Add, Rem	Pass
Laboratory summaries and candidate-specific interactions	Recorded at or before the 24-hour cutoff	Add, Rem	Pass
Laboratory and vital-sign transition states	Recorded at or before the 24-hour cutoff	Add	Pass
Candidate-specific medication exposure	Order information available by the 24-hour cutoff	Add, Rem	Pass
Candidate metadata	Derived from the training split only	Add, Rem	Pass

Appendix

Table 6: Feature provenance, availability rules, branch usage, and automated leakage-check results.

#	Stratum	Condition
1	Empty-to-nonempty	anchor size $=0$ , target size $>0$
2	Nonempty-to-empty	anchor size $>0$ , target size $=0$
3	Continue	$\mathrm{add}=0$ and $\mathrm{rem}=0$
4	Add-only	$\mathrm{add}>0$ and $\mathrm{rem}=0$
5	Remove-only	$\mathrm{rem}>0$ and $\mathrm{add}=0$
6	Switch	$\mathrm{add},\mathrm{rem}>0$ and $\mathrm{add}+\mathrm{rem}\leq 3$

Appendix

Table 7: Action strata and their first-match precedence order.

Comparator	Metric	Diff.	95% CI
RETAIN-FullSet	composite	0.0368	[0.0341, 0.0395]
GAMENet-FullSet	composite	0.0448	[0.0424, 0.0472]
Tree	composite	0.0431	[0.0408, 0.0456]
Shared	composite	0.0293	[0.0276, 0.0310]
MICRON	composite	0.1759	[0.1731, 0.1788]
MICRON	addition F1	0.0229	[0.0193, 0.0264]

Appendix

Table 8: R-GEAN improvements over each comparator with 95% patient-paired bootstrap confidence intervals. Positive values favor R-GEAN.

Scheme	F1 +	F1 -	$J_{\text{chg}}$	F1 ${}_{\text{act}}$
Primary	0.35	0.30	0.25	0.10
Equal	0.25	0.25	0.25	0.25
Edit-heavy	0.40	0.40	0.10	0.10
Balanced	0.30	0.30	0.20	0.20
Addition-heavy	0.45	0.25	0.20	0.10
Removal-heavy	0.25	0.45	0.20	0.10

Appendix

Table 9: Prespecified composite-weight schemes over addition F1, removal F1, changed-admission Jaccard, and action macro-F1.

Property	MICRON	DualNN	R-GEAN
Prediction interval	within admission	between visits	within admission
Base regimen	observed 24-hour anchor	predicted	observed 24-hour anchor
Recursive prediction	no	no	no
Candidate spaces	one shared vector	two nets, shared space	two disjoint anchor spaces
Candidate-specific exposure	no	no	yes
Observation boundary	24-hour cutoff	visit-level input	explicit 24-hour cutoff

Appendix

Table 10: Comparison of the adapted MICRON baseline, DualNN, and R-GEAN under the within-admission prediction setting. MICRON columns describe the adapted implementation evaluated here; DualNN is listed as originally proposed and was not re-evaluated.

Model	2026	2027	2028	Mean
R-GEAN	$0.4672$	$0.4583$	$0.4655$	$0.4636$
Shared	$0.4350$	$0.4278$	$0.4375$	$0.4334$
RETAIN-FullSet	$0.4275$	$0.4292$	$0.4289$	$0.4285$
GAMENet-FullSet	$0.4195$	$0.4170$	$0.4182$	$0.4182$
MICRON-Adapted	$0.2884$	$0.2895$	$0.2894$	$0.2891$

Appendix

Table 11: Seed-specific edit composites and arithmetic means. Means are calculated from the full-precision seed-specific results.

Model	Parameters	Composite
Shared, original	363,394	$0.4333$ ${}\pm{}$	$0.0050$
Shared, param-matched	465,454	$0.4362$ ${}\pm{}$	$0.0070$
R-GEAN	465,924	$0.4636$ ${}\pm{}$	$0.0047$

Appendix

Table 12: Three-seed parameter-matched comparison. Models were retrained under a common capacity-control protocol, so the shared-model values differ slightly from the primary multiseed results.

Component	Parameters	Training time
R-GEAN addition branch	235,138	31.3 s
R-GEAN removal branch	230,786	4.4 s
R-GEAN combined	465,924	35.7 s

Appendix

Table 13: R-GEAN parameter counts and recorded branch training times for the primary seed-2026 run.

Model	Set $\mu$ F1	Set rank	Comp.	Edit rank
R-GEAN	$0.7769$	$4$	$0.4643$	$1$
Shared	$0.7334$	$7$	$0.4350$	$2$
RETAIN-FullSet	$0.7878$	$3$	$0.4275$	$3$
Tree	$0.7425$	$6$	$0.4212$	$4$
GAMENet-FullSet	$0.7938$	$2$	$0.4195$	$5$
Frequency	$0.7753$	$5$	$0.4012$	$6$

Appendix

Table 14: Set-level and edit-level ranks for the ten systems included in the rank-correlation analysis. Rank 1 is best.

Variant	Add F1	Remove F1	Changed Jacc.	Action macro-F1	Composite	$\Delta$ Comp.
Full R-GEAN	$0.2509$	$0.6013$	$0.6088$	$0.4392$	$0.4643$	$0.0000$
Addition, no transition states	$0.2384$	$0.6013$	$0.5888$	$0.4251$	$0.4535$	$-0.0108$
Addition, no candidate exposure	$0.2282$	$0.6013$	$0.6153$	$0.4173$	$0.4558$	$-0.0085$
Addition, no laboratory interaction	$0.2441$	$0.6013$	$0.5966$	$0.4322$	$0.4582$	$-0.0061$
Addition, no prior history	$0.2210$	$0.6013$	$0.5837$	$0.3825$	$0.4419$	$-0.0224$
Addition, no interaction regularizer	$0.1723$	$0.6013$	$0.5546$	$0.4410$	$0.4234$	$-0.0409$

Appendix

Table 15: Complete component ablations on the test set. The unmodified prediction direction is held fixed.

Variant	Comp.	Add F1	Add/adm
No regularizer	0.4223	0.1686	3.69
Interaction-derived regularizer	0.4618	0.2454	1.95
L1 expected-count control	0.4467	0.2084	2.77
Anchor-size count-matching	0.4615	0.2448	1.87

Appendix

Table 16: Addition-branch regularization (removal branch fixed): a count-matching control reproduces the predictive effect of the interaction-derived term. Pair-normalized interaction burden is 0.983 without the term and 0.994 with it.

Model	Rare-addition F1
Tree	0.0938
RETAIN-FullSet	0.0765
GAMENet-FullSet	0.0541
Shared	0.0231
R-GEAN	0.0004

Appendix

Table 17: Addition F1 across the 39 classes in the bottom half of training-set addition support.

Group	R@5	R@10	MRR	Macro-AUPRC
All classes	0.384	0.550	0.268	0.076
Rare classes	0.017	0.029	0.042	0.028
Common classes	0.440	0.630	0.302	0.123
Rarest quartile	0.000	0.000	0.019	0.007

Appendix

Table 18: Addition ranking quality by class-frequency group.

Direction	Method	ECE	Brier	F1
Addition	raw	0.0323	0.0235	0.2509
Addition	isotonic	0.0003	0.0213	0.2510
Removal	raw	0.1943	0.1370	0.6013
Removal	isotonic	0.0023	0.0904	0.6011

Appendix

Table 19: Validation-only calibration. Calibration improves probability quality with essentially unchanged thresholded F1.

Subgroup	Adm.	Add F1	Rem F1	Chg. Jacc.	Comp.
Age 18–44	8,254	0.2378	0.5982	0.6197	0.4613
Age 45–64	16,171	0.2361	0.6064	0.6133	0.4599
Age 65–74	10,865	0.2530	0.6065	0.5955	0.4624
Age 75–84	8,292	0.2509	0.5913	0.5875	0.4552
Age 85+	5,450	0.2414	0.5930	0.5698	0.4476
Sex F	25,473	0.2467	0.5974	0.6124	0.4628

Appendix

Table 20: R-GEAN performance by age band, sex, and admission type on the test split (internal stress tests).

Duration stratum	Classes	R-GEAN	RETAIN
Median order $\leq 48$ h	27	0.6925	0.5937
Median order $>48$ h	51	0.4769	0.2774

Appendix

Table 21: Removal F1 by class order-duration stratum. Classes are grouped by their median anchor-order duration on the test split for this post-hoc diagnostic; the grouping was not used for training, threshold selection, or model selection.

Event category	Count	%
Represented addition	121,836	17.2
Represented removal	74,174	10.5
Transient exposure	238,068	33.6
Stop–restart, membership unchanged	207,087	29.2
Restart of a removal class	26,601	3.8
Within-window stop of an addition class	40,383	5.7

Appendix

Table 22: Complete mutually exclusive taxonomy of medication start and stop events occurring between the 24-hour anchor and discharge. Net snapshot labels are directly represented by the first two rows; the final two rows are opposite-direction transitions of edit-labeled classes with multiple within-window episodes.

Explore similar work

RxEval: A Prescription-Level Benchmark for Evaluating LLM Medication Recommendation
May 14, 2026Shuhao Chen, Weisen Jiang, Changmiao Wang +4Healthcare Clinical Decision Support
The Hong Kong University of Science and Technology · Southern University of Science and Technology · The Chinese University of Hong Kong +2
GraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation
Mar 21, 2026Krati Saxena, Tomohiro ShibataClinical Decision Support Longitudinal EHR Modeling
Kyushu Institute of Technology, Kitakyushu, Fukuoka, Japan
SafeRx-Agent: A Knowledge-Grounded Multi-Agent Framework for Safe and Explainable Medication Recommendation
May 27, 2026Xinyu Wang, Hanwei Wu, Zhenghan Tai +7Multi-Agent LLM Systems Healthcare
McGill University · McMaster University · University of Toronto +1

May 14, 2026J/K move · Enter open · S save

RxEval: A Prescription-Level Benchmark for Evaluating LLM Medication Recommendation

Shuhao Chen, Weisen Jiang, Changmiao Wang, Xiaoqing Wu+3

The Hong Kong University of Science and Technology · Southern University of Science and Technology · The Chinese University of Hong Kong+2

Healthcare Clinical Decision Support Clinical Decision-Making Large Language Model-Based Recommendation Clinical Language Model Evaluation

Open paper AskPDF arXiv

Inpatient medication recommendation requires clinicians to repeatedly select specific medications, doses, and routes as a patient's condition evolves. Existing benchmarks formulate this task as admission-level prediction over coarse drug codes with multi-hot diagnostic and procedure code inputs, failing to capture the per-timepoint, information-rich nature of real prescribing. We propose RxEval, a prescription-level benchmark that evaluates LLM prescribing capability by multiple-choice questions: each question presents a detailed patient profile and time-ordered clinical trajectory, requiring selection of specific medication-dose-route triples from real prescriptions and patient-specific distractors generated via reasoning-chain perturbation. RxEval comprises 1,547 questions spanning 584 patients, 18 diagnostic categories, and 969 unique medications. Evaluation of 16 LLMs shows that RxEval is both challenging and discriminative: F1 ranges from 45.18 to 77.10 across models, and the best Exact Match is only 46.10%. Error analysis reveals that even frontier models may overlook stated patient information and fail to derive clinical conclusions.

Abstract

Explore similar work

RxEval: A Prescription-Level Benchmark for Evaluating LLM Medication Recommendation

GraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation

SafeRx-Agent: A Knowledge-Grounded Multi-Agent Framework for Safe and Explainable Medication Recommendation