Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SHIELD: Semantic Heterogeneity Integrated Embedding for Latent Discovery in Clinical Trial Safety Signals

Francois Vandenhende, Anna Georgiou, Theodoros Psaras, Ellie Karekla

TL;DR

SHIELD presents an integrated pipeline that fuses information-theoretic disproportionality analysis with semantic clustering of MedDRA PT embeddings to uncover latent safety syndromes in clinical trials. It adapts an IC-like statistic for multi-arm data, applies hierarchical Bayesian shrinkage, and builds a semantic-augmented utility graph $U=ZS Z$ to drive spectral clustering with Ward linkage, yielding syndrome-level PT clusters annotated by LLMs. The method, demonstrated on a Duchenne trial, recovers known signals (e.g., liver-related AEs) and provides an interpretable network representation of treatment-specific safety patterns, enhancing causal interpretation and safety assessment. By bridging traditional signal detection with NLP-driven semantic structure, SHIELD offers a scalable, interpretable framework for comprehensive safety evaluation in multi-arm clinical trials.

Abstract

We present SHIELD, a novel methodology for automated and integrated safety signal detection in clinical trials. SHIELD combines disproportionality analysis with semantic clustering of adverse event (AE) terms applied to MedDRA term embeddings. For each AE, the pipeline computes an information-theoretic disproportionality measure (Information Component) with effect size derived via empirical Bayesian shrinkage. A utility matrix is constructed by weighting semantic term-term similarities by signal magnitude, followed by spectral embedding and clustering to identify groups of related AEs. Resulting clusters are annotated with syndrome-level summary labels using large language models, yielding a coherent, data-driven representation of treatment-associated safety profiles in the form of a network graph and hierarchical tree. We implement the SHIELD framework in the context of a single-arm incidence summary, to compare two treatment arms or for the detection of any treatment effect in a multi-arm trial. We illustrate its ability to recover known safety signals and generate interpretable, cluster-based summaries in a real clinical trial example. This work bridges statistical signal detection with modern natural language processing to enhance safety assessment and causal interpretation in clinical trials.

SHIELD: Semantic Heterogeneity Integrated Embedding for Latent Discovery in Clinical Trial Safety Signals

TL;DR

SHIELD presents an integrated pipeline that fuses information-theoretic disproportionality analysis with semantic clustering of MedDRA PT embeddings to uncover latent safety syndromes in clinical trials. It adapts an IC-like statistic for multi-arm data, applies hierarchical Bayesian shrinkage, and builds a semantic-augmented utility graph to drive spectral clustering with Ward linkage, yielding syndrome-level PT clusters annotated by LLMs. The method, demonstrated on a Duchenne trial, recovers known signals (e.g., liver-related AEs) and provides an interpretable network representation of treatment-specific safety patterns, enhancing causal interpretation and safety assessment. By bridging traditional signal detection with NLP-driven semantic structure, SHIELD offers a scalable, interpretable framework for comprehensive safety evaluation in multi-arm clinical trials.

Abstract

We present SHIELD, a novel methodology for automated and integrated safety signal detection in clinical trials. SHIELD combines disproportionality analysis with semantic clustering of adverse event (AE) terms applied to MedDRA term embeddings. For each AE, the pipeline computes an information-theoretic disproportionality measure (Information Component) with effect size derived via empirical Bayesian shrinkage. A utility matrix is constructed by weighting semantic term-term similarities by signal magnitude, followed by spectral embedding and clustering to identify groups of related AEs. Resulting clusters are annotated with syndrome-level summary labels using large language models, yielding a coherent, data-driven representation of treatment-associated safety profiles in the form of a network graph and hierarchical tree. We implement the SHIELD framework in the context of a single-arm incidence summary, to compare two treatment arms or for the detection of any treatment effect in a multi-arm trial. We illustrate its ability to recover known safety signals and generate interpretable, cluster-based summaries in a real clinical trial example. This work bridges statistical signal detection with modern natural language processing to enhance safety assessment and causal interpretation in clinical trials.
Paper Structure (24 sections, 31 equations, 3 figures)

This paper contains 24 sections, 31 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Incidence Table Construction
  4. Semantic Embeddings and Similarity Matrix
  5. Information Component Calculation
  6. Expected counts under the null.
  7. Raw Information Component (IC).
  8. Frequentist likelihood-ratio test (G-test).
  9. Directional assignment for two-arm trials.
  10. Hierarchical Bayesian shrinkage via Dirichlet prior.
  11. Hyperprior estimation.
  12. Posterior sampling.
  13. Posterior summaries.
  14. Multi-arm ($k>2$) and two-sided interpretation.
  15. Semantic Utility Matrix Construction
  16. ...and 9 more sections

Figures (3)

  • Figure 1: SHIELD network graph of AEs for the 4-arm disproportionality analysis of the DMD trial.
  • Figure 2: Dendogram plot of AE clusters for the 2-arm disproportionality analysis of the DMD trial during Part 1.
  • Figure 3: SHIELD network graph of AEs for the single-arm incidence analysis of the DMD trial.