Research map / Probabilistic and scientific modeling / Causal discovery

DoAtlas-1: A Causal Compilation Paradigm for Clinical AI

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Counterfactual identification and bounds · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes causal compilation, compiling heterogeneous medical studies into executable interventional estimand objects (DoAtlas-1) with conflict-aware graphs and validation against the Human Phenotype Project.

Problem

Medical foundation models produce narrative evidence summaries but cannot quantify intervention effects, detect evidence conflicts, or validate literature claims against real-world data, limiting clinical auditability.

Approach

Standardize research evidence into structured estimand objects (population, contrast, outcome, time horizon, measure) with canonicalization, bucket-based comparability, decidable conflict predicates, and a typed query layer (six causal query operators) executed over a multi-edge estimand atlas, plus external validation against HPP real-world data.

Key results

• Canonicalization semantic preservation 98.5% on 200 sampled kernels, reviewed by three clinical epidemiologists (Cohen's kappa 0.94).
• Query executability 80.5% across six query types (e.g., 200/250 for Q_do); conflict detection recall and precision 100% on 42 flagged buckets.
• External validation: direction agreement 1.000 with 0.000 false positives at coverage 0.625, versus 0.712 agreement and 0.630 false positives for text-only baseline.

Surprising findings

• Ten representative literature-to-HPP mismatch cases include point-estimate sign flips, showing that many literature claims do not reproduce in real-world data.
• Removing canonicalization drops bucket purity from 96.8% to 68.3%.

Limitations

• Coverage limited to cardiovascular, diabetes, and metabolic domains.
• Executability is 80.5%, so roughly one fifth of queries cannot be answered.
• Validation mappings to HPP are approximate; many mismatch cases attributed to population, contrast, and outcome non-comparability.

Experimental setup

Datasets: Human Phenotype Project (HPP) 754 clinical studies compiled into 1,445 effect kernels

Benchmarks: Human Phenotype Project external validation canonicalization sample of 200 kernels six query suite

Metrics: canonicalization accuracy F1 conflict detection recall/precision executability rate direction agreement false positive rate purity

Baselines: Baseline-1 (no canonicalization) Baseline-2 (text-only + coarse cohort) Med-PaLM GPT-4-Med

Builds on: do-calculus structural causal models random-effects meta-analysis heterogeneity (I^2)

Closest papers by meaning

• Simulating Viva Voce Examinations to Evaluate Clinical Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.87)
• Improving the Generation and Evaluation of Synthetic Data for Downstream Medical Causal Inference (NeurIPS 2025 · similarity 0.85)
• End-To-End Causal Effect Estimation from Unstructured Natural Language Data (NeurIPS 2024 · similarity 0.85)
• ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence (NeurIPS 2026 · similarity 0.85)
• CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models (NeurIPS 2026 · similarity 0.84)
• MedEvoEval: Evaluating Continual Evolution of Doctor Agents through Simulated Clinical Episodes (NeurIPS 2026 · similarity 0.84)