基于Transformer与特征映射的可解释性药物推荐方法

打开文本图片集

中图分类号：TP391 文献标识码：A

文章编号：1006-8228（2025）12-37-07

Explainable Medication Recommendation Method Based on Transformer and Feature Mapping

Li Yeting，Xu Chun

（SchoolofInformationManagementXinjiangUniversityofinanceandEconomics，UrumqiXinjiang3olhina）

Abstract：MedicationrecommendationisakeyaplicationofAIinhealthcare.Toadressthelimitationsofexistingmethods— suboptimalaccuracyandsafetyaswellaslimitedinterpretabilityofrecommendationresults—thispaperproposesExpMR，an explainablemedicationrecommendationmethod.First，aTransformerencodermodelslongitudinalEHRstolearnrobustpatient representations.Second，afeaturemappingmechanismprojectsthe"black-box"embedingsofdiagnoses，procedures，and medicationsontoanexplainabledimensionalfeaturespace，enhancingtransparencyandtraceabilityThird，acontrolabletheshold strategydyamicallyreweightsthedrug-druginteraction（DDI）losstoensuresafetyExperimentsonMMIC-IIandMIMIC-IV demonstratethatExpMRsurpasesstrongbaselinesonJaccardcoeffcient，F1，andPRAUC，whileachievinglowerDDIates， confirmingitscombined advantages in accuracy，safety，and interpretability.

Keywords：Medication Recommendation；Transformer;Feature Mapping；Interpretability

0引言

药物推荐旨在依据患者就诊时产生的电子健康记录（ElectronicHealthRecords，EHR），为医生提供精准、安全且有效的药物组合建议，从而提升处方效率和医疗服务质量。（剩余9528字）