HypeMed: Enhancing Medication Recommendations with Hypergraph-Based Patient Relationships

Xiangxu Zhang; Xiao Zhou; Hongteng Xu; Jianxun Lian

cs.IRMar 19, 2026

HypeMed: Enhancing Medication Recommendations with Hypergraph-Based Patient Relationships

Authors: Xiangxu Zhang, Xiao Zhou, Hongteng Xu, Jianxun Lian

Organizations: Gaoling School of Artificial Intelligence, Renmin University of China, China · Beijing Key Laboratory of Research on Large Models and Intelligent Governance · Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE · Microsoft Research Asia, China

Abstract

Medication recommendations aim to generate safe and effective medication sets from health records. However, accurately recommending medications hinges on inferring a patient's latent clinical condition from sparse and noisy observations, which requires both (i) preserving the visit-level combinatorial semantics of co-occurring entities and (ii) leveraging informative historical references through effective, visit-conditioned retrieval. Most existing methods fall short in one of both aspects: graph-based modeling often fragments higher-order intra-visit patterns into pairwise relations, while inter-visit augmentation methods commonly exhibit an imbalance between learning a globally stable representation space and performing dynamic retrieval within it. To address these limitations, this paper proposes HypeMed, a two-stage hypergraph-based framework unifying intra-visit coherence modeling and inter-visit augmentation. HypeMed consists of two core modules: MedRep for representation pre-training, and SimMR for similarity-enhanced recommendation. In the first stage, MedRep encodes clinical visits as hyperedges via knowledge-aware contrastive pre-training, creating a globally consistent, retrieval-friendly embedding space. In the second stage, SimMR performs dynamic retrieval within this space, fusing retrieved references with the patient's longitudinal data to refine medication prediction. Evaluation on real-world benchmarks shows that HypeMed outperforms state-of-the-art baselines in both recommendation precision and DDI reduction, simultaneously enhancing the effectiveness and safety of clinical decision support.

Figures & tables

Figure 1. Overview of Bob’s EHR spanning four medical visits. Each visit includes diagnosis, procedure, and medication codes, illustrating the progression of related respiratory conditions over time. Note that the medication section in Visit-4 is left unfilled, indicating it as the target for medication recommendation.

Notation	Definition
$\mathcal{R}$	EHR database containing all patients’ visit sequences
$\mathcal{S}^{(i)}_{t}$	The $t$ -th visit of patient $i$ (patient index $i$ is omitted when clear)
$\mathcal{D},\mathcal{P},\mathcal{M}$	Sets of all diagnosis, procedure, and medication codes
$\mathbf{d},\mathbf{p},\mathbf{m}$	Multi-hot vectors for diagnoses, procedures, and medications
$\mathbf{v}^{\mathcal{D}}$ , $\mathbf{v}^{\mathcal{P}}$ , and $\mathbf{v}^{\mathcal{M}}$	Visit representations derived from diagnoses, procedures, and medications
$\mathcal{H}_{\mathcal{X}}=(\mathcal{N}_{\mathcal{X}},\mathcal{E}_{\mathcal{X}})$	Hypergraph for domain $X\!\in\!\{\mathcal{D},\mathcal{P},\mathcal{M}\}$

Table 1. Summary of main notations used in this paper.

Figure 2. An example of a hypergraph with four nodes and two hyperedges:

e_{1}

connects nodes

n_{1}

n_{2}

, and

n_{3}

, while

e_{2}

connects nodes

n_{3}

and

n_{4}

Figure 3. Overall architecture of HypeMed . HypeMed comprises two stages: the Medical Entity Relevance Representation Stage (MedRep) and the Similar Visit Enhanced Medication Recommendation Stage (SimMR). MedRep focuses on encoding intra-visit set-level combinatorial semantics into a globally consistent, retrieval-friendly embedding space. SimMR integrates longitudinal history and visit-conditioned retrieved similar visits to refine latent condition estimation.

Figure 4. The detailed architecture of the Knowledge-aware Hypergraph Encoder (KHGE). The encoder consists of a Local Message Passing Network (LMPN) and a Knowledge-aware Global Attention Network (KGAN). The LMPN focuses on message propagation over the medical hypergraph, while the KGAN integrates medical knowledge into a global self-attention mechanism. The two outputs are combined through a feed-forward fusion layer with residual connections.

Item	MIMIC-III	MIMIC-IV	eICU
# of patients	6,350	9,036	7,855
# of visits	15,032	20,616	16,869
avg. # of visits	2.37	2.28	2.15
# of unique diag. codes	1,958	1,892	692
# of unique proc. codes	1,430	4,939	-
# of unique med. codes	131	131	46

Table 2. Statistics of processed MIMIC-III/MIMIC-IV and eICU.

Hyperparameter	Range
Learning rate	$[5\times 10^{-3},\,1\times 10^{-4}]$
Weight decay	$[1\times 10^{-4},\,1\times 10^{-5}]$
Dropout ratio	$[0,\,0.5]$
Epochs	$50$
Batch size	$16$
Pretraining epochs	$\{150,\,300,\,500,\,1500,\,3000\}$

Table 3. Hyperparameter search space.

Model	Jaccard $\uparrow$	F1 $\uparrow$	PRAUC $\uparrow$	DDI $\downarrow$	# Med. $\downarrow$
LR	$49.24_{\pm 0.27}$	$64.96_{\pm 0.25}$	$75.48_{\pm 0.31}$	$\color[rgb]{0,0.4,0.6}{8.30_{\pm 0.07}}$	$16.05_{\pm 0.14}$
ECC	$48.89_{\pm 0.26}$	$64.47_{\pm 0.23}$	$76.19_{\pm 0.20}$	$\color[rgb]{0,0.4,0.6}{8.57_{\pm 0.08}}$	$15.74_{\pm 0.07}$
LEAP	$45.50_{\pm 0.18}$	$61.68_{\pm 0.17}$	$65.46_{\pm 0.42}$	$\color[rgb]{0,0.4,0.6}{7.82_{\pm 0.08}}$	$18.57_{\pm 0.08}$
RETAIN	$48.60_{\pm 0.34}$	$64.66_{\pm 0.34}$	$75.98_{\pm 0.43}$	$\color[rgb]{1,0.25,0}{8.83_{\pm 0.10}}$	$18.92_{\pm 0.13}$
GAMENet	$50.18_{\pm 0.20}$	$65.84_{\pm 0.19}$	$76.48_{\pm 0.23}$	$\color[rgb]{1,0.25,0}{8.88_{\pm 0.09}}$	$27.39_{\pm 0.19}$
SafeDrug	$51.14_{\pm 0.25}$	$66.81_{\pm 0.23}$	$76.43_{\pm 0.23}$	$\color[rgb]{0,0.4,0.6}{5.87_{\pm 0.04}}$	$18.94_{\pm 0.13}$

Table 4. Performance comparison on MIMIC-III in terms of Jaccard (%), F1 (%), PRAUC (%), DDI (%) and # Med.. Numbers in bold and underlined indicate the best and the second-best performance, respectively, according to t-tests at the 95% confidence level. The ground-truth rate of DDI in the MIMIC-III dataset is 8.68%. In the DDI column, values in orange indicate that the mean+std exceeds the ground truth (inferior), while values in blue indicate that the mean-std falls below it (superior).

Model	Jaccard $\uparrow$	F1 $\uparrow$	PRAUC $\uparrow$	DDI $\downarrow$	# Med. $\downarrow$
LR	$47.65_{\pm 0.20}$	$63.31_{\pm 0.20}$	$74.08_{\pm 0.22}$	$\color[rgb]{0,0.4,0.6}{7.07_{\pm 0.09}}$	$13.85_{\pm 0.09}$
ECC	$45.82_{\pm 0.16}$	$61.39_{\pm 0.15}$	$73.58_{\pm 0.14}$	$\color[rgb]{0,0.4,0.6}{7.11_{\pm 0.14}}$	$13.08_{\pm 0.14}$
LEAP	$43.73_{\pm 0.26}$	$59.84_{\pm 0.25}$	$63.73_{\pm 0.26}$	$\color[rgb]{0,0.4,0.6}{6.57_{\pm 0.06}}$	$17.34_{\pm 0.07}$
RETAIN	$44.10_{\pm 0.49}$	$60.23_{\pm 0.50}$	$71.60_{\pm 0.40}$	$\color[rgb]{1,0.25,0}{7.29_{\pm 0.14}}$	$14.43_{\pm 0.17}$
GAMENet	$47.80_{\pm 0.26}$	$63.56_{\pm 0.28}$	$75.09_{\pm 0.20}$	$\color[rgb]{1,0.25,0}{7.85_{\pm 0.05}}$	$23.36_{\pm 0.17}$
SafeDrug	$48.47_{\pm 0.17}$	$64.24_{\pm 0.17}$	$74.27_{\pm 0.22}$	$\color[rgb]{0,0.4,0.6}{6.11_{\pm 0.11}}$	$18.15_{\pm 0.11}$

Table 5. Performance comparison on MIMIC-IV . The ground-truth DDI rate is 7.24%. Other notations are consistent with Tab. 4 .

Model	Jaccard $\uparrow$	F1 $\uparrow$	PRAUC $\uparrow$	# Med. $\downarrow$
LR	$39.96_{\pm 0.22}$	$55.33_{\pm 0.21}$	$\underline{73.24}_{\pm 0.36}$	$7.55_{\pm 0.07}$
ECC	$33.41_{\pm 0.26}$	$47.66_{\pm 0.29}$	${72.66}_{\pm 0.28}$	$5.62_{\pm 0.05}$
LEAP	$41.25_{\pm 0.18}$	$56.43_{\pm 0.21}$	$62.67_{\pm 0.29}$	$13.36_{\pm 0.04}$
RETAIN	$41.29_{\pm 0.35}$	$56.53_{\pm 0.34}$	$71.02_{\pm 0.35}$	$9.81_{\pm 0.06}$
GAMENet	${40.80}_{\pm 0.17}$	$56.00_{\pm 0.18}$	$68.86_{\pm 0.18}$	$12.45_{\pm 0.03}$
SafeDrug	$41.19_{\pm 0.23}$	$56.42_{\pm 0.21}$	$69.79_{\pm 0.30}$	$10.68_{\pm 0.06}$

Table 6. Performance comparison on eICU . There is no DDI information available in the eICU dataset. Other notations are consistent with Tab. 4 .

MIMIC-III
Model	Jaccard $\uparrow$	F1 $\uparrow$	PRAUC $\uparrow$	DDI $\downarrow$	#Med. $\downarrow$
HypeMed	$\mathbf{53.10_{\pm 0.43}}$	$\mathbf{68.49_{\pm 0.37}}$	$77.32_{\pm 0.37}$	$6.67_{\pm 0.06}$	$22.77_{\pm 0.13}$
– SimMR– w/o Sim.	$52.91_{\pm 0.42}$	$68.33_{\pm 0.37}$	$\mathbf{77.44_{\pm 0.34}}$	$6.40_{\pm 0.06}$	$22.50_{\pm 0.13}$
– SimMR– w/o Hist.	$51.90_{\pm 0.41}$	$67.40_{\pm 0.37}$	$76.40_{\pm 0.44}$	$6.12_{\pm 0.06}$	$23.16_{\pm 0.17}$
– MedRep–None	$51.70_{\pm 0.46}$	$67.23_{\pm 0.42}$	$76.28_{\pm 0.36}$	$6.52_{\pm 0.06}$	$23.51_{\pm 0.14}$
– MedRep–Fixed	$50.49_{\pm 0.23}$	$66.24_{\pm 0.21}$	$75.44_{\pm 0.33}$	$6.45_{\pm 0.04}$	$22.74_{\pm 0.13}$

Table 7. Ablation study on MIMIC-III and MIMIC-IV . Results are reported for both MedRep and SimMR ablations. Jaccard, F1, PRAUC, and DDI are expressed as percentages (%), while #Med. retains its original scale. Values denote mean

\pm

standard deviation over ten bootstrap sampling.

Dataset	Epoch	Jaccard $\uparrow$	F1 $\uparrow$	PRAUC $\uparrow$	DDI $\downarrow$	#Med. $\downarrow$
MIMIC-III	0	$52.27_{\pm 0.25}$	$67.75_{\pm 0.22}$	$\mathbf{77.08}_{\pm 0.32}$	$7.93_{\pm 0.10}$	$21.90_{\pm 0.17}$
	100	$\mathbf{52.48}_{\pm 0.23}$	$\mathbf{67.97}_{\pm 0.20}$	$76.10_{\pm 0.29}$	$6.00_{\pm 0.04}$	$21.93_{\pm 0.10}$
	300	$52.26_{\pm 0.25}$	$67.78_{\pm 0.22}$	$76.27_{\pm 0.39}$	$6.06_{\pm 0.04}$	$22.74_{\pm 0.11}$
	1000	$51.91_{\pm 0.22}$	$67.41_{\pm 0.19}$	$75.95_{\pm 0.37}$	$6.28_{\pm 0.04}$	$22.95_{\pm 0.14}$
MIMIC-IV	0	$49.87_{\pm 0.23}$	$65.48_{\pm 0.23}$	$75.60_{\pm 0.27}$	$6.30_{\pm 0.10}$	$16.50_{\pm 0.14}$
	100	$50.03_{\pm 0.23}$	$65.65_{\pm 0.22}$	$75.87_{\pm 0.27}$	$6.34_{\pm 0.07}$	$16.81_{\pm 0.17}$

Table 8. Pretraining-epoch sensitivity analysis on MIMIC-III and MIMIC-IV . Jaccard, F1, PRAUC, and DDI are reported in percentage (%), while #Med. retains its original scale. Values denote mean

\pm

standard deviation over ten bootstrap sampling.

Dataset	Top- $n$	Jaccard $\uparrow$	F1 $\uparrow$	PRAUC $\uparrow$	DDI $\downarrow$	#Med. $\downarrow$
MIMIC-III	0	$53.10_{\pm 0.43}$	$68.49_{\pm 0.37}$	$77.32_{\pm 0.37}$	$6.67_{\pm 0.06}$	$22.77_{\pm 0.13}$
	1	$\mathbf{53.23}_{\pm 0.40}$	$\mathbf{68.59}_{\pm 0.35}$	$\mathbf{77.50}_{\pm 0.35}$	$6.52_{\pm 0.04}$	$22.58_{\pm 0.13}$
	10	$52.93_{\pm 0.37}$	$68.33_{\pm 0.33}$	$77.43_{\pm 0.38}$	$6.25_{\pm 0.05}$	$22.63_{\pm 0.13}$
	100	$52.88_{\pm 0.40}$	$68.27_{\pm 0.35}$	$77.20_{\pm 0.36}$	$6.44_{\pm 0.06}$	$23.31_{\pm 0.13}$
	500	$52.80_{\pm 0.38}$	$68.22_{\pm 0.34}$	$77.13_{\pm 0.36}$	$6.39_{\pm 0.06}$	$23.20_{\pm 0.12}$
	1000	$52.86_{\pm 0.37}$	$68.26_{\pm 0.33}$	$77.16_{\pm 0.39}$	$6.40_{\pm 0.05}$	$22.97_{\pm 0.12}$

Table 9. Top-

n

sensitivity analysis on MIMIC-III and MIMIC-IV . Jaccard, F1, PRAUC, and DDI are expressed as percentages (%), while #Med. retains its original scale. Values denote mean

\pm

standard deviation over ten bootstrap sampling.

Dataset	#Views	Jaccard $\uparrow$	F1 $\uparrow$	PRAUC $\uparrow$	DDI $\downarrow$	#Med. $\downarrow$	GPU Mem.
MIMIC-III	$2$	$51.13_{\pm 0.17}$	$66.77_{\pm 0.16}$	$75.73_{\pm 0.34}$	$6.50_{\pm 0.05}$	$23.80_{\pm 0.14}$	$9.76$
	$3$	$51.26_{\pm 0.22}$	$66.85_{\pm 0.18}$	$75.86_{\pm 0.40}$	$6.23_{\pm 0.05}$	$24.60_{\pm 0.12}$	$14.09$
	$4$	$51.05_{\pm 0.20}$	$66.67_{\pm 0.19}$	$75.89_{\pm 0.43}$	$6.33_{\pm 0.04}$	$23.49_{\pm 0.12}$	$24.53$
MIMIC-IV	$2$	$48.67_{\pm 0.23}$	$64.40_{\pm 0.22}$	$73.57_{\pm 0.32}$	$5.90_{\pm 0.07}$	$20.93_{\pm 0.17}$	$14.46$
	$3$	$48.61_{\pm 0.21}$	$64.34_{\pm 0.20}$	$73.44_{\pm 0.27}$	$5.99_{\pm 0.06}$	$21.19_{\pm 0.18}$	$23.75$
	$4$	$48.66_{\pm 0.25}$	$64.39_{\pm 0.25}$	$73.31_{\pm 0.29}$	$6.03_{\pm 0.07}$	$20.87_{\pm 0.17}$	$38.61$

Table 10. Performance and GPU memory usage under different numbers of views on MIMIC-III and MIMIC-IV . Jaccard, F1, PRAUC, and DDI are reported in percentage (%); #Med. and GPU Memory (GiB) keep their original scales.

Panel of Figure 5 — Figure 5. Visualization of adaptive channel weighting and its association with model performance across different visit lengths. Blue and orange dots represent per-visit gate weights for the history and similarity channels, respectively. Black and purple lines indicate the Spearman correlations ( $\rho$ ) between Jaccard performance and each channel’s weight, computed within visit-length groups. To ensure statistical reliability, visit-length groups with fewer than 20 samples were excluded.

Dataset	Model	Jaccard $\uparrow$	F1 $\uparrow$	PRAUC $\uparrow$	DDI $\downarrow$	#Med. $\downarrow$
MIMIC-III	SafeDrug	$49.47_{\pm 1.24}$	$65.22_{\pm 1.03}$	$74.92_{\pm 0.91}$	$6.22_{\pm 0.20}$	$18.01_{\pm 0.26}$
	DAPSNet	$50.68_{\pm 0.95}$	$66.20_{\pm 0.79}$	$74.90_{\pm 1.00}$	$6.12_{\pm 0.11}$	$19.92_{\pm 0.33}$
	COGNet	$51.19_{\pm 1.20}$	$66.64_{\pm 1.11}$	$75.62_{\pm 0.69}$	$8.62_{\pm 0.15}$	$24.90_{\pm 0.27}$
	HypeMed	$\mathbf{51.60}_{\pm 0.86}$	$\mathbf{67.14}_{\pm 0.73}$	$\mathbf{75.83}_{\pm 0.98}$	$8.65_{\pm 0.29}$	$20.04_{\pm 0.31}$
	HypeMed -SimMR- w/o Hist.	$49.93_{\pm 1.00}$	$65.46_{\pm 0.92}$	$74.78_{\pm 1.04}$	$6.19_{\pm 0.25}$	$23.44_{\pm 0.32}$
	HypeMed -SimMR- w/o Sim.	$49.52_{\pm 1.08}$	$65.24_{\pm 0.98}$	$74.62_{\pm 0.96}$	$6.35_{\pm 0.25}$	$22.43_{\pm 0.28}$

Table 11. Performance comparison under the cold-start (first-visit) setting on MIMIC-III and MIMIC-IV . All metrics are reported as percentages (%), except for the average number of medications (#Med.). Values denote mean

\pm

standard deviation over ten bootstrapped runs. A total of 192 and 307 patients have only one visit in MIMIC-III and MIMIC-IV, respectively.

Figure 6. Retrieval-quality correlation analysis. We compare HypeMed with HypeMed -MedRep-None, a controlled variant that keeps the same prediction architecture but removes MedRep pre-training, resulting in retrieval within a non-pretrained embedding space. Test visits are grouped into bins by retrieval quality, measured as the F1 score between medications in the retrieved visits and the ground-truth medications. The x-axis denotes retrieval-quality bins (higher is better), and the y-axis reports the final prediction performance.

Figure 7. Comparative time complexity across models (FLOPs vs. Parameters)

Figure 8. Comparative Visualization of Diagnosis Entity Representations in MedRep using HGCN and KHGE Encoders.

	Diagnoses	Medications
Visit-1	V8741, V153, 79029, V1083, 5715, 1550, 07054	N02B, A01A, A02B , A06A , B05C , J05A , A12C , A07A , J01E , N01A , C03C , J01C , N02A , B01A , N05C , D01A , J01D , L04A , D11A , A04A , A07E , D07A , D04A , C03D
Visit-2	99682, 78552, 0389, 99592, 5848, 570, 28419, 4561, 5589, 78060, 07070, 5715, V1083, E8780, 4568, V168, V1582	N02B , A01A , A02B , A06A , B05C , A12A , A12C , C01C , A07A , J02A , J05A , V03A , J01E , J01C, N02A, A02A , J01M , B01A , N05C , J01D , B02B, C08C , D11A , A04A, A07E

Table 12. The example of Patient-908 with two visits for the case study. Medications recommended correctly without using similar visit information are highlighted in bold . Blue and orange are used respectively to denote additional correct medications introduced and erroneous medications eliminated after incorporating similar visit information.

Explore similar work

Mar 21, 2026

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

Recommending safe and effective medication combinations from electronic health records (EHRs) is a core clinical AI problem, yet it remains difficult because patient trajectories are long, noisy, and clinically heterogeneous. Existing methods typically excel at either temporal modeling across visits or pharmacological knowledge integration (e.g., drug-drug interactions, DDIs), but rarely achieve both while robustly suppressing noise. We present GraphDiffMed, a knowledge-constrained medication recommendation framework built on dual-scale Differential Attention v2. Differential attention is applied at both intra-visit and inter-visit levels to filter spurious signals within encounters and across longitudinal history, while pharmacological constraints are incorporated during learning. Experiments on MIMIC-III and ablation studies show that this design consistently improves recommendation quality and ranking over strong baselines while achieving a more favorable safety performance balance. We further find that the strongest-performing configuration uses only demographic auxiliary features under our experimental setting. Overall, GraphDiffMed demonstrates that combining noise-aware attention with pharmacological constraints yields more reliable and clinically meaningful medication recommendation. We open-source our code at https://github.com/saxenakrati09/GraphDiffMed.

Krati Saxena, Tomohiro Shibata

Kyushu Institute of Technology, Kitakyushu, Fukuoka, Japan

Clinical Decision Support Longitudinal EHR Modeling Electronic Health Records

Ask PDF

Oct 12, 2026

ARMED: Adaptive Retrieval-Augmented Medication Recommendation

Medication recommendation requires jointly considering diagnoses, comorbidities, longitudinal patient history, and treatment interactions. Although recent retrieval-augmented generation approaches using language models have shown promise for clinical recommendation, most existing systems rely on fixed retrieval pipelines that cannot adapt to diverse clinical retrieval scenarios and may introduce noisy or irrelevant evidence. We propose ARMED, an adaptive clinical RAG framework that dynamically expands retrieval according to retrieval sufficiency and current-admission relevance. ARMED first observes the initial retrieval outcome and selectively performs supplementary or fallback multi-query retrieval when additional evidence is needed. The framework also incorporates a relevance gate that activates only clinically relevant same-patient history and organizes retrieved evidence into structured query-level clinical evidence. Experiments on MIMIC-IV demonstrate that ARMED consistently outperforms previous static retrieval baselines and improves overall medication recommendation quality. By combining locally runnable small language models with adaptive retrieval, ARMED supports medication recommendation without relying on external language model APIs or task-specific fine-tuning.

Kyungkeon Chung, Se Un Park, Hyejin Lee +1

UX Factory, Inc., Seongnam-si, South Korea

Adaptive RAG Clinical Decision Support Large Language Model-Based Recommendation

Ask PDF

Oct 1, 2026

HADRec: A Hierarchy-Aware Drug Recommendation Framework by Fusing Molecular Knowledge and Electronic Health Record

Accurate medication recommendation is central to clinical decision-making, directly determining therapeutic efficacy and patient safety. However, existing methods suffer from two key limitations: drugs are often abstracted as discrete tokens, ignoring their molecular structures and pharmacological mechanisms, and the commonly used "flat" recommendation paradigm fails to leverage the hierarchical logic of the internationally standardized Anatomical Therapeutic Chemical (ATC) classification system. To address these issues, we propose HADRec, a Hierarchy-Aware Drug Recommendation framework that integrates molecular knowledge with electronic health records (EHRs). HADRec employs LLaMA-7B to encode clinical notes for rich patient representations and ChemBERTa to encode drug Simplified Molecular Input Line Entry System strings, building a global molecular knowledge base. A cross-attention mechanism then performs deep multimodal fusion between patient states and drug features. The framework further incorporates a hierarchical predictor and a novel consistency constraint loss to enforce strict adherence to ATC logical dependencies. Extensive experiments on MIMIC-III demonstrate that HADRec achieves state-of-the-art performance across Jaccard, F1, and PR-AUC. External validation on MIMIC-IV confirms strong generalization under distribution shifts, and calibration analysis shows well-calibrated predictive confidence on MIMIC-IV with ECE = 0.04, and Brier = 0.06. Counterfactual evaluation reveals clinically aligned reasoning, disentangling disease-specific treatments from general care. Together, these results establish HADRec as a high-performance, interpretable, and clinically grounded pathway toward safe and reliable AI-driven medication recommendation.

Junke Wang, Hongshun Ling, Li Zhang +4

School of Biomedical Engineering,South-Central Minzu University, 182 Minzu Avenue, Hongshan District, WuHan, 430074, HuBei, China

Clinical Decision Support

Ask PDF