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

Junke Wang; Hongshun Ling; Li Zhang; Jinjing Wu; Tong Shao; Fang Wang; Yuan Gao

cs.LGOct 1, 2026

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

Authors: Junke Wang, Hongshun Ling, Li Zhang, Jinjing Wu, Tong Shao, Fang Wang, Yuan Gao

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

Abstract

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.

Figures & tables

Figure 1: The workflow of the HADRec framework: (a) Multimodal representation learning employs LLaMA-7B and ChemBERTa to extract patient clinical profiles and drug molecular features, respectively. (b) Multimodal fusion performs deep interactions between patient states and drug molecular knowledge via a cross-attention mechanism. (c) The hierarchical prediction module constructs a four-level predictor to generate drug recommendations following the top-down ATC taxonomy. (d) The consistency constraint module incorporates a consistency loss function to enforce clinical logicality in the recommendations.

Figure 2: Illustration of the prompting template used for model input. The placeholders <…> are populated with patient EHR data.

Category	Item	Count
Patients	Total patients	6350
	Average visits patient	2.37
Clinical Events	Unique diagnosis codes	1958
	Unique procedure codes	1430
Medications	Unique medications (ATC-L4)	110
	Average medications per visit	19.57

Table 1: Summary statistics of the processed dataset derived from MIMIC-III.

Category	Model	Key-Technique	Drug-Knowledge Used	Use ATC?
Knowledge-Graph & EHR-based	GAMENet [ 27 ]	Memory-Network & GNN	Knowledge-Graph (KG)	$\times$
	SafeDrug [ 37 ]	DDI-controllable-Loss	Molecular-Graph	$\times$
	RASNet [ 45 ]	Reverse-Time Attention	–	$\times$
	EGNeT [ 25 ]	GNN & Explainable Module	External-Medical-KG	$\times$
Molecular-Structure-Aware	DEPOT [ 42 ]	Motif-Tree & Graph-Transformer	Molecular-Motifs	$\times$
	MoleRec [ 39 ]	Substructure Interaction & Annealed Loss	Molecular Substructures	$\times$

Table 2: Categorization and key characteristics of the baseline models.

Category	Notation	Description	Value
Model Architecture	$d_{dim}$	Hidden layer dimension	4096
	$d_{mol}$	Molecular embedding dimension	768
Attention Mechanism	$\lambda$	Residual connection scaling factor	0.7
	$\tau_{cross}$	Soft masking threshold for cross-attention	0.1
Loss Weighting	$(\lambda_{i})_{i=1}^{4}$	Weights for hierarchical BCE loss (L1-L4)	0.4, 0.4, 1.0, 1.5
	$\lambda_{cons}$	Global weight for consistency constraint loss	0.01

Table 3: Hyperparameter settings of the HADRec framework

Category	Model	Jaccard $\uparrow$	F1 $\uparrow$	PR-AUC $\uparrow$	DDI $\downarrow$	Avg.Med $\downarrow$
Knowledge-Graph & EHR-based	GAMENet [ 27 ]	0.5067±0.0025	0.6626±0.0025	0.7631±0.0030	0.0864±0.0006	27.21±0.11
	SafeDrug [ 37 ]	0.5213±0.0030	0.6768±0.0027	0.7647±0.0025	0.0589±0.0005	19.92±0.16
	RASNet [ 45 ]	0.5401±0.0021	0.6931±0.0019	0.7882±0.0025	0.0599±0.0009	21.04±0.15
	EGNeT [ 25 ]	0.5431±0.001	0.6955±0.0007	0.7842±0.0023	0.0693±0.0012	23.33±0.01
Molecular-Structure-Aware	DEPOT [ 42 ]	0.5352±0.0022	0.6890±0.003	0.7816±0.0035	0.0684±0.0007	20.62±0.19
	MoleRec [ 39 ]	0.5301±0.0025	0.6841±0.0022	0.7748±0.0022	0.0756±0.0006	22.22±0.17

Table 4: Performance comparison between HADRec and baseline models. Results are reported as mean ± standard deviation across multiple runs. The best performances are highlighted in bold.

MKR	HP	CL	Jaccard $\uparrow$	F1 $\uparrow$	PR-AUC $\uparrow$
$\times$	$\times$	$\times$	0.5397±0.0063	0.6904±0.0063	0.7946±0.0060
$\checkmark$	$\times$	$\times$	0.5405±0.0090	0.6915±0.0087	0.7944±0.0066
$\checkmark$	$\checkmark$	$\times$	0.5472±0.0051	0.6967±0.0050	0.7986±0.0052
$\times$	$\checkmark$	$\checkmark$	0.5450±0.0054	0.6950±0.0053	0.7971±0.0049
$\checkmark$	$\checkmark$	$\checkmark$	0.5517±0.0059	0.7010±0.0054	0.7994±0.0057

Table 5: Ablation study of key components of HADRec

Figure 3: Performance comparison demonstrates the superiority of full hierarchical supervision. The figure shows Jaccard (a-c), F1-score (d-f), and PR-AUC (g-i) across four ATC levels (L1-L4) under three supervision settings: L1+L4 (a, d, g), L1+L3+L4 (b, e, h), and full supervision L1+L2+L3+L4 (c, f, i).

Figure 4: Hierarchical attention distribution for a representative patient case. The heatmap visualizes the attention weights across Levels 1 to 4 for the ten drugs with the highest predicted probabilities at Level 4. Darker shading indicates stronger attention, clearly revealing the structured “parent-to-child” information flow constrained by the ATC hierarchy. Cells sharing the same background pattern correspond to drugs belonging to the same therapeutic class, highlighting consistent information transfer within hierarchical groups. Specifically, dotted (·), horizontal (—), and vertical (|) patterns denote A0-, A1-, and B0-level classes, respectively.

method	Threshold	Jaccard $\uparrow$	F1 $\uparrow$	Precision $\uparrow$	Recall $\uparrow$
Global Threshold	0.1	0.4578±0.0032	0.6186±0.0027	0.4789±0.0031	0.9184±0.0039
	0.2	0.5253±0.0057	0.6791±0.0050	0.5848±0.0052	0.8443±0.0065
	0.3	0.5491±0.0065	0.6992±0.0060	0.6589±0.0061	0.7741±0.0083
	0.35	0.5517±0.0059	0.7010±0.0054	0.6891±0.0064	0.7403±0.0072
	0.4	0.5470±0.0054	0.6969±0.0050	0.7126±0.0063	0.7069±0.0073
	0.5	0.5275±0.0051	0.6789±0.0050	0.7595±0.0075	0.6349±0.0062

Table 6: Performance comparison of global thresholding (0.1–0.9) and class-wise dynamic thresholding on HADRec. Results are reported as mean ± standard deviation across multiple runs. The best performances are highlighted in bold.

Setting	Disease Profile	Recommended Medication Combination (Top Categories)
Excluding diabetes (case 1)	Hypertension, Hyperlipidemia, Chronic kidney disease, Coronary artery disease, Obesity	B01A, C10A, C07A, A01A, A06A, N02B, C09A, B05C, A02B, A12B, A12C
Excluding hypertension (case 2)	Diabetes mellitus type II, Hyperlipidemia, Chronic kidney disease, Coronary artery disease, Obesity	B01A, C10A, A06A, B05C, A01A, C07A, A12C, A10A, H04A, N02B, A02B, A12B, A04A
Full condition set (case 3)	Diabetes mellitus type II, Hypertension, Hyperlipidemia, Chronic kidney disease, Coronary artery disease, Obesity	B01A, C10A, A06A, C07A, A01A, B05C, H04A, A10A, A12B, A02B, N02B, C09A

Table 7: Counterfactual medication recommendations under selective disease removal.

Model	Jaccard $\uparrow$	F1 $\uparrow$	PR-AUC $\uparrow$	DDI $\downarrow$	Avg.Med $\downarrow$
KNN [ 27 ]	$0.3978\pm 0.0023$	$0.5473\pm 0.0022$	$0.3917\pm 0.0023$	$0.0815\pm 0.0013$	$12.78\pm 0.2203$
LR [ 24 ]	$0.4499\pm 0.0033$	$0.5985\pm 0.0027$	$0.7259\pm 0.0024$	$0.0769\pm 0.0008$	$10.36\pm 0.1111$
ECC [ 26 ]	$0.4350\pm 0.0017$	$0.5809\pm 0.0020$	$0.7193\pm 0.0021$	$0.0681\pm 0.0003$	$8.76\pm 0.2115$
RETAIN [ 7 ]	$0.4182\pm 0.0020$	$0.5729\pm 0.0016$	$0.6775\pm 0.0012$	$0.0835\pm 0.0004$	$11.29\pm 0.2007$
LEAP [ 41 ]	$0.4123\pm 0.0022$	$0.5661\pm 0.0019$	$0.6026\pm 0.0025$	$0.0667\pm 0.0005$	$12.13\pm 0.1530$
GAMENet [ 27 ]	$0.4594\pm 0.0015$	$0.6126\pm 0.0011$	$0.7229\pm 0.0020$	$0.0803\pm 0.0006$	$18.86\pm 0.0002$

Table 8: External validation and cross-dataset transferability evaluation of HADRec on MIMIC-IV.

Dataset	ECE	Brier Score
MIMIC-III	0.09	0.12
MIMIC-IV	0.04	0.06
Transfer (III→IV)	0.04	0.07

Table 9: Calibration and uncertainty evaluation results.

Model	Backbone	GPU Device	Memory Usage	Avg. Inference Time
HADRec	LLaMA-7B	NVIDIA RTX3090	12.75GB	3.14 s/sample

Table 10: Inference efficiency and deployment cost analysis.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Longitudinal EHR context analysis across different visit window configurations on the MIMIC-III cohort (N=6,350). (A) Distribution of patient counts by number of visits (log scale). (B) Cumulative patient coverage as a function of window size; window size 3 covers 89.0% of patients. (C) Marginal coverage gain diminishes rapidly beyond window size 3. (D) Information retention rate for truncated patients (n=701) with mean 63.1%. (E) Temporal distance of visits to last admission; truncated visits are significantly older (median 2.4 years). (F) Key summary statistics.

Explore similar work

SafeRx-Agent: A Knowledge-Grounded Multi-Agent Framework for Safe and Explainable Medication Recommendation
May 27, 2026Xinyu Wang, Hanwei Wu, Zhenghan Tai +7Medication Leaflet Mimic-Iv Datasets
McGill University · McMaster University · University of Toronto +1
GRAIN: Molecules Are Not the Right Granularity -- Active-Ingredient Modeling for Safe Medication Recommendation
Jul 30, 2026Juao Fan, Jinhan Li, Shengxin ZhuMedication Leaflet Drug-Drug Interactions
Guangdong Provincial Key Laboratory of Interdisciplinary Research and Application for Data Science, Beijing Normal–Hong Kong Baptist University, Zhuhai, China
ChatHealthAI: Aligning Electronic Health Record Representations with Large Language Models for Grounded Clinical Reasoning
Jun 1, 2026Bo-Hong Wang, Baicheng Peng, Ruilin Wang +3Medical World Model Clinical Reasoning Training
School of Computer Science, McGill University, Montreal, QC, Canada · Mila - Quebec AI Institute, Montreal, QC, Canada

May 27, 2026, cs.CLJ/K move · Enter open · S save

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

Xinyu Wang, Hanwei Wu, Zhenghan Tai, Sicheng Lyu+6

McGill University · McMaster University · University of Toronto+1

Medication Leaflet Mimic-Iv Datasets International Classification Of Disease Coding

Open paper AskPDF arXiv

Medication recommendation predicts medications for patient visits, but existing methods still face two key challenges. At the model level, traditional drug recommendation methods only predict structured drug codes with limited evidence grounding, while LLM agents can use richer clinical context but may lack safety verification and traceability. At the task level, existing benchmarks often use broad medication categories, which ignore subgroup-level safety differences and can lead to risk overestimation. We introduce the first fine-grained medication recommendation setting based on fourth-level ATC code generation. We propose Safe Prescription Agent (SafeRx-Agent), a knowledge-grounded multi-agent framework that uses patient context, external clinical knowledge, and safety verification to recommend traceable medication sets. Experimental results on MIMIC-III and MIMIC-IV datasets show that SafeRx-Agent improves fine-grained medication prediction accuracy while controlling drug interactions, contraindications, and medication set size.

Abstract

Explore similar work

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

GRAIN: Molecules Are Not the Right Granularity -- Active-Ingredient Modeling for Safe Medication Recommendation

ChatHealthAI: Aligning Electronic Health Record Representations with Large Language Models for Grounded Clinical Reasoning