TRACE: reconstructing fragmented microbial landscapes for high-resolution antimicrobial resistance surveillance

TL;DR

TRACE is a new computational framework that improves the detection of antimicrobial resistance by reconstructing fragmented clinical data into detailed microbial profiles.

Contribution

TRACE introduces a novel triple-based reconstruction framework that combines clinical data aggregation with in silico simulation to recover lost microbial signals.

Findings

TRACE achieves a Macro-F1 score of 15.4% and Precision@5 of 70.3% on the MIMIC-IV dataset.

The framework successfully disentangles clinically similar but biologically distinct resistance phenotypes.

TRACE reduces misidentification of rare pathogen-drug combinations by capturing subtle resistance patterns.

Abstract

Comprehensive mapping of Antimicrobial Resistance (AMR) dynamics is a cornerstone of systems microbiology and global health security. However, the precise characterization of microbial interactions within clinical ecosystems is severely confined by the fragmentation of phenotypic evidence. Critical data points, including pathogen identity, anatomical niche, and susceptibility profiles, are often buried within unstructured clinical narratives. This fragmentation creates a blind spot in surveillance, where the ecological signals of rare but critical resistance phenotypes are lost, preventing a systems-level understanding of microbial evolution.

To bridge this gap, we propose the Triple-based Reconstruction for Antimicrobial Clinical Evidence (TRACE) framework, a systems-computational approach that treats clinical records as unstructured microbiological sensors. Unlike traditional classification models, our approach focuses on the biological reconstruction of evidence. It first aggregates scattered findings into coherent Organism-Specimen-Susceptibility biological clusters. To resolve the scarcity of data for underrepresented pathogens, which represent the long-tail of biodiversity, we employ a constraint-based in silico simulation strategy. This component utilizes a large language model to synthesize biologically plausible phenotypic profiles that are rigorously verified for factual consistency. Finally, the framework employs parameter-efficient adaptation to capture the subtle semantic nuances of complex resistance patterns.

Validated on the MIMIC-III and MIMIC-IV datasets, TRACE demonstrates a superior ability to recover lost microbial signals. On the challenging MIMIC-IV benchmark, it achieves a Macro-F1 score of 15.4% and a Precision@5 of 70.3%. More importantly, the system successfully disentangles clinically similar but biologically distinct resistance phenotypes, significantly reducing the misidentification of rare pathogen-drug combinations.

By resolving evidence fragmentation and effectively expanding the feature space for underrepresented microorganisms, TRACE provides a robust solution for fine-grained infectious disease characterization. This framework establishes a foundation for high-precision AMR phenotype identification, paving the way for more reliable monitoring of resistance trends in clinical ecosystems.

Figures