Agentic discovery of blood biomarker from distilled private health records
TLDR
Authors distill private health records into a released scoring tool, enabling an agentic propose-score-refine loop to discover blood biomarkers with external validation.
Reasoning
The paper's strengths include a privacy-preserving distillation approach using a large real-world patient panel, external validation across multiple cohorts, and demonstrated improvements over literature-seeded candidates. Weaknesses are limited visible details on the agent design, candidate search space, and cohort-specific variability in gains.
Read-first score
Read-first score 49.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 60.
Field roles
Rank sensitivity
Stability: volatile; rank range: 120.
Keyword Scores
Deep Analysis
Innovations
- Distillation of private health records into a released scoring tool by training graph attention networks inside the data boundary and releasing only trained weights
- Agentic propose-score-refine loop grounded in real-world data without exposing patient data
- Privacy-preserving biomarker hypothesis generation from routine complete blood counts across 13 immune-mediated diseases
Methodology
For each of 13 immune-mediated diseases, a graph attention network was trained inside the Clalit Health Services data boundary to predict the case-control AUC of candidate CBC expressions, and only the trained weights were released as a scoring tool. An agent uses this tool in a propose-score-refine loop to evaluate candidate CBC expressions. External validation was performed across three independent cohorts, comparing agent-discovered expressions to literature-seeded starting points and reranking candidates from three frontier research tools.
Key Results
Agent-discovered expressions improved on literature-seeded starting points by a median of 4.18 AUC percentage points. Across three independent cohorts, reranking candidates from three frontier research tools improved on their first choices in most comparisons, with gains varying by cohort.