Hybrid Symbolic–Neural Framework for Explainable Knowledge Discovery: An Industry-Oriented Approach for EHR Analytics

カテゴリ：研究会(論文単位)

論文No：PI25072

グループ名：【C】電子・情報・システム部門 知覚情報研究会

発行日：2025/12/23

タイトル(英語)：Hybrid Symbolic–Neural Framework for Explainable Knowledge Discovery: An Industry-Oriented Approach for EHR Analytics

著者名：Bhoite Harshraj(LTIMindtree)

著者名(英語)： Harshraj Bhoite(LTIMindtree)

キーワード：Explainable AI,Neuro-Symbolic AI,Knowledge Discovery,EHR,Clinical Decision Support,Ontologies,Explainable AI,Neuro-Symbolic AI,Knowledge Discovery,EHR,Clinical Decision Support,Ontologies

要約(日本語)：Healthcare organizations increasingly rely on machine learning for risk stratification, early warning, and operational optimization. Yet, the lack of transparency of black-box models limits clinical adoption and erodes trust. This paper presents a Hybrid Symbolic–Neural Framework (HSNF) that unifies neural representation learning with rule-driven symbolic reasoning to deliver high predictive performance and human centered explanations. HSNF comprises (i) a Neural Representation Module (NRM) that encodes multimodal EHR signals (structured labs/vitals and unstructured notes), (ii) a Symbolic Reasoning Engine (SRE) that grounds learned concepts into clinical predicates aligned with ontologies (e.g., SNOMED CT, ICD) and enforces differentiable logic constraints, and (iii) an Explanation Engine (EE) that surfaces faithful, stepwise justifications in clinician friendly terms. Using MIMIC-III, we benchmark HSNF for infection risk identification and 48-hour readmission prediction against pure DNNs, Logic Tensor Networks, and tree ensembles. HSNF matches strong baselines on AUROC while substantially improving interpretability metrics and error localization. Extensive ablations quantify the contribution of ontology constraints, predicate calibration, and rule sparsity; we additionally report latency, memory, and MLOps considerations for production deployment on common hospital data stacks. Our results suggest that neuro-symbolic systems can close the gap between model accuracy and actionable explainability required in regulated healthcare environments.

要約(英語)：Healthcare organizations increasingly rely on machine learning for risk stratification, early warning, and operational optimization. Yet, the lack of transparency of black-box models limits clinical adoption and erodes trust. This paper presents a Hybrid Symbolic–Neural Framework (HSNF) that unifies neural representation learning with rule-driven symbolic reasoning to deliver high predictive performance and human centered explanations. HSNF comprises (i) a Neural Representation Module (NRM) that encodes multimodal EHR signals (structured labs/vitals and unstructured notes), (ii) a Symbolic Reasoning Engine (SRE) that grounds learned concepts into clinical predicates aligned with ontologies (e.g., SNOMED CT, ICD) and enforces differentiable logic constraints, and (iii) an Explanation Engine (EE) that surfaces faithful, stepwise justifications in clinician friendly terms. Using MIMIC-III, we benchmark HSNF for infection risk identification and 48-hour readmission prediction against pure DNNs, Logic Tensor Networks, and tree ensembles. HSNF matches strong baselines on AUROC while substantially improving interpretability metrics and error localization. Extensive ablations quantify the contribution of ontology constraints, predicate calibration, and rule sparsity; we additionally report latency, memory, and MLOps considerations for production deployment on common hospital data stacks. Our results suggest that neuro-symbolic systems can close the gap between model accuracy and actionable explainability required in regulated healthcare environments.

本誌：2025年12月26日知覚情報研究会

本誌掲載ページ：7-10p

原稿種別：英語

PDFファイルサイズ：778Kバイト