DoAtlas-1: A Causal Compilation Paradigm for Clinical AI

Yulong Li; Jianxu Chen; Xiwei Liu; Chuanyue Suo; Rong Xia; Zhixiang Lu; Yichen Li; Xinlin Zhuang

cs.AIFeb 22, 2026

DoAtlas-1: A Causal Compilation Paradigm for Clinical AI

Authors: Yulong Li, Jianxu Chen, Xiwei Liu, Chuanyue Suo, Rong Xia, Zhixiang Lu, Yichen Li, Xinlin Zhuang,

Organizations: Mohamed bin Zayed University of Artificial Intelligence · Xi’an Jiaotong-Liverpool University.

Abstract

Medical foundation models generate narrative explanations but cannot quantify intervention effects, detect evidence conflicts, or validate literature claims, limiting clinical auditability. We propose causal compilation, a paradigm that transforms medical evidence from narrative text into executable code. The paradigm standardizes heterogeneous research evidence into structured estimand objects, each explicitly specifying intervention contrast, effect scale, time horizon, and target population, supporting six executable causal queries: do-calculus, counterfactual reasoning, temporal trajectories, heterogeneous effects, mechanistic decomposition, and joint interventions. We instantiate this paradigm in DoAtlas-1, compiling 1,445 effect kernels from 754 studies through effect standardization, conflict-aware graph construction, and real-world validation (Human Phenotype Project, 10,000 participants). The system achieves 98.5% canonicalization accuracy and 80.5% query executability. This paradigm shifts medical AI from text generation to executable, auditable, and verifiable causal reasoning.

Figures & tables

	Narrative Generation	Causal Compilation (DoAtlas)
Representation	Text with citations	Executable estimand objects
Query Interface	Natural language questions	Typed queries ( $Q_{do}$ , $Q_{med}$ , $Q_{CATE}$ , …)
Output Format	“Studies suggest…” + citations	$\hat{\theta}=0.73$ [0.60–0.88] ✓ validated
Effect Quantification	✗ Qualitative descriptions only	✓ Quantitative estimates with CI
Conflict Detection	✗ Implicit (“results vary”)	✓ Explicit (directional, interval, $I^{2}$ )
External Validation	✗ Cannot verify claims	✓ Statistical test vs. real-world data

Table 1: Paradigm shift from narrative generation to causal compilation for clinical decision support.

Figure 1: Query Gallery. DoAtlas supports six executable query types: Qdo (interventional effect), Qcf (counterfactual effect under an alternative exposure), Qtraj (time-indexed effect trajectory), QCATE (heterogeneous effects across populations/strata), Qmed (direct/indirect effects via a mediator), and Qjoint (joint intervention effect). Each answer is computed from a witness subset of evidence objects (EV-*); when a query is not executable, the system returns diagnostic flags instead of a numeric answer.

Figure 2: Illustration of the HPP study data. (a) Multi-modal and multi-omic data architecture of HPP. Molecular tests include genetics, microbiome metagenomics, metabolomics, proteomics, and single-cell RNA sequencing. Clinical tests capture deep physiological states via retinal imaging, liver and carotid ultrasound, DXA body composition, and continuous monitoring through CGM and multi-night sleep monitoring. Biobanking involves the systematic preservation of blood and stool samples for longitudinal discovery. (b) Timing of sleep monitoring with respect to all other phenotypes. The body system characteristics were measured within a period of ±6 months from the visit, with the sleep monitoring performed in three nights within a 2-week time period after that visit. ABI, ankle–brachial index; CGM, continuous glucose monitoring; DXA, dual-energy X-ray absorptiometry; ECG, electrocardiogram; HMO, health maintenance organization; IMT, intimamedia thickness; MS, mass spectrometry; PWV, pulse wave velocity; RDS, recent depressive symptoms.

Figure 3: Overview of DoAtlas. DoAtlas compiles heterogeneous clinical evidence into standardized interventional estimand objects with explicit contrasts, effect scales, time horizons, and target populations. It organizes comparable claims into conflict-aware causal graphs, supports executable causal queries, including do-calculus estimation, mechanistic pathway decomposition, counterfactual reasoning, individualized treatment effect estimation, combination intervention modeling, and dynamic prognostic trajectory modeling. External validation signals from HPP are used to assess reliability, adjudicate evidential conflicts, and support auditable clinical decision support.

Query	Signature	Required evidence pattern	Output type
$Q_{\mathrm{do}}$	$(X,Y,\varepsilon)\mapsto$	$\exists\,s\ \text{s.t.}\ \mathcal{B}_{s}\neq\varnothing\ \wedge\ s\ \text{matches}\ (X,Y,\varepsilon)$	$\mathsf{Ans}$
$Q_{\mathrm{med}}$	$(X,M,Y,\varepsilon)\mapsto$	$\exists\,s_{1},s_{2}\ \text{s.t.}\ \mathcal{B}_{s_{1}}\neq\varnothing\ \wedge\ \mathcal{B}_{s_{2}}\neq\varnothing\ \wedge\ s_{1}\ \text{matches}\ (X,M,\varepsilon)\ \wedge\ s_{2}\ \text{matches}\ (M,Y,\varepsilon)$	$(\mathsf{Ans}_{\mathrm{TE}},\mathsf{Ans}_{\mathrm{NDE}},\mathsf{Ans}_{\mathrm{NIE}})$
$Q_{\mathrm{joint}}$	$(X_{1},X_{2},Y,\varepsilon)\mapsto$	$\exists\,s_{1},s_{2}\ \text{s.t.}\ \mathcal{B}_{s_{1}}\neq\varnothing\ \wedge\ \mathcal{B}_{s_{2}}\neq\varnothing\ \wedge\ s_{1}\ \text{matches}\ (X_{1},Y,\varepsilon)\ \wedge\ s_{2}\ \text{matches}\ (X_{2},Y,\varepsilon)$	$\mathsf{Ans}$
$Q_{\mathrm{cf}}$	$(X,Y,z,\varepsilon)\mapsto$	$\exists\,s\ \text{s.t.}\ \mathcal{B}_{s}\neq\varnothing\ \wedge\ s\ \text{matches}\ (X,Y,\varepsilon)\ \wedge\ s\ \text{matches individual }z\ \text{(counterfactual)}$	$\mathsf{Ans}$
$Q_{\mathrm{CATE}}$	$(X,Y,z,\varepsilon)\mapsto$	$\exists\,s\ \text{s.t.}\ \mathcal{B}_{s}\neq\varnothing\ \wedge\ s\ \text{matches}\ (X,Y,\varepsilon)\ \wedge\ s\ \text{matches subgroup }z$	$\mathsf{Ans}$
$Q_{\mathrm{traj}}$	$(X,Y,\mathcal{T},\varepsilon)\mapsto$	$\forall\,\hat{\tau}\in\mathcal{T},\ \exists\,s(\hat{\tau})\ \text{s.t.}\ \mathcal{B}_{s(\hat{\tau})}\neq\varnothing\ \wedge\ s(\hat{\tau})\ \text{matches}\ (X,Y,\varepsilon,\hat{\tau})$	$\{\mathsf{Ans}(\hat{\tau})\}_{\hat{\tau}\in\mathcal{T}}$

Table 2: Six query operators in DoAtlas and their executability conditions. Here, “matches” means that the bucket key

s

agrees with the queried endpoint pair

(X,Y)

and the estimand semantic dimensions encoded in the key, including the population category, contrast type, aligned time class, and measure family. “Matches subgroup

z

” means that the population category encoded by

s

equals the subgroup identifier

z

. “Matches individual

z

(counterfactual)” refers to an individual-level context for counterfactual querying, whereas “matches subgroup

z

” refers to a population-category constraint for subgroup effects. More detailed explanations are in Appendix D.4 – D.5 .

Case	Patient Profile	Clinical Question	Query	Effect Estimate	Interpretation
1	68yo male, T2DM, HFrEF (EF 35%), eGFR 45	SGLT2i HF benefit	do(Dapagliflozin vs Placebo) $\to$ HF worsening or CV death	HR=0.73 (0.60-0.88), 18mo	27% RRR
		SGLT2i renal protection	do(Dapagliflozin 10mg) $\to$ Composite kidney endpoint	HR=0.61 (0.51-0.72), 2.3yr	39% RRR
2	72yo female, HFpEF (EF 55%), dyspnea	Symptom improvement	do(Dapagliflozin vs Placebo) $\to$ KCCQ score	+5.8 pts (2.3-9.2), 12wk	Exceeds MCID
3	64yo male, CKD stage 3, eGFR 38	Slow CKD progression	do(Empagliflozin vs Placebo) $\to$ Kidney progression or CV death	HR=0.72 (0.64-0.82), 2yr	28% RRR
4	70yo male, HFpEF, recent hospitalization	Prevent rehospitalization	do(Empagliflozin vs Placebo) $\to$ HF hospitalization	HR=0.71 (0.60-0.83), 26mo	29% RRR

Table 3: Clinical decision support cases. Demonstration of DoAtlas-1 in real-world clinical scenarios. RRR = relative risk reduction; MCID = minimal clinically important difference; CV = cardiovascular; HFrEF = heart failure with reduced ejection fraction; HFpEF = heart failure with preserved ejection fraction; CKD = chronic kidney disease; T2DM = type 2 diabetes mellitus.

Intervention (out-degree)	Degree	Outcome (in-degree)	Degree
TRE	22	All-cause mortality	41
Finerenone	17	Stroke	26
Dapagliflozin 10mg + CR	15	Cardiovascular disease	20
Dapagliflozin	15	Myocardial infarction	16
Early ECLS	14	Heart failure	13

Table 4: Top-5 hub nodes. Interventions ranked by out-degree (number of unique outcomes studied); outcomes ranked by in-degree (number of unique interventions studied).

Query Type	Total	Executable	Exec Rate
$Q_{do}$ (do-calculus)	250	200	80.0%
$Q_{cf}$ (counterfactual)	250	200	80.0%
$Q_{traj}$ (trajectory)	45	36	80.0%
$Q_{CATE}$ (subgroup)	200	163	81.5%
$Q_{med}$ (mediation)	36	29	80.6%
$Q_{joint}$ (joint intervention)	134	107	79.9%

Table 5: Query suite performance. Negative samples: 25%. Executability rate = queries with evidence; execution correctness = correct answers;

Configuration	Buckets	Purity	Query Acc	Key Loss
full	1,110	96.8%	100%	-
no_canonical	$\sim$ 1,450	68.3%	87.3%	HR/logHR spurious splits
no_align_tau	$\sim$ 1,320	83.7%	93.6%	“12 months”/“1 year” over-splitting
weak_key	$\sim$ 890	78.9%	94.8%	per_unit/binary over-merging

Table 6: Marginal contribution of canonicalization components. Purity = fraction of kernels correctly assigned to buckets; query accuracy = fraction of executable queries with correct answers. no_canonical : skip scale normalization (HR remains raw, not converted to logHR); no_align_tau : skip time horizon alignment; weak_key : weaken bucket key (ignore contrast_type).

Method	Coverage	Direction agreement	False positive
Ours (causal compilation)	0.625	1.000	0.000
Baseline-1 (no canonicalization)	0.643	0.713	0.047
Baseline-2 (text-only + coarse cohort)	0.861	0.712	0.630

Table 7: External validation comparison.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Module	Definition and Standards
Paper	The paper’s title and abstract, journal, and year are used as basic bibliographic information, and PMID and DOI are extracted to establish digital anchors; the registration number of clinical research records is used to associate the trial protocol.
Provenance	Establish a foundation for data auditing by creating a physical mapping between structured data and unstructured raw text. This includes a list of specific chart numbers and page ranges for core conclusions, as well as relevant information for confirming supplementary materials.
Design	Documenting the methodological context in which the evidence was generated provides context for assessing the strength of the evidence. This includes the type of research structure (e.g., ”RCT”, ”Cohort”, ”Cross-sectional”, ”Case-control”), statistical analysis methods (e.g., ”Cox Proportional Hazards”, ”Linear Mixed Model”, ”Logistic Regression”, ”Difference-in-Differences”), total sample size, total number of study groups, blinding implementation (e.g., ”Double-blind”, ”Single-blind”, ”None”, ”Open-label”), randomization scheme (e.g., ”Block”, ”Stratified”, ”Cluster”, ”None”), and estimation objectives (e.g., ”ITT” for intention-to-treat analysis, ”Per-protocol” for per-protocol analysis, and ”ATE” for average treatment effect).
Population	Structured fields are used to characterize the biological boundaries of research subjects, serving as a direct basis for screening the HPP local validation cohort. These include descriptions of the data population’s age, sex composition, specific disease or health conditions (such as ”Type 2 Diabetes patients”, ”Healthy volunteers”, ”Obstructive Sleep Apnea”), and descriptions of key inclusion and exclusion criteria.
Transport Signature	Record the research center, data collection period, geographical location, and medical setting to assess the external validity of the evidence and its transferability from the source population to the target population. Data includes the name of the research center where the data was collected, the specific start and end dates or year range of data collection, the geographical region, the medical setting (e.g., ”Community”, ”Hospital”, ”Clinic”, ”Laboratory”), and the nature of the data source (e.g., ”Clinical Trial”, ”Registry”, ”EHR”, ”Cohort Study”).
Variables	A ”node_role” separation design is adopted to define the physical properties of variables and their functions in the causal network. It mainly consists of variable nodes and their roles. Variable nodes are defined by an identifier, their name in the paper, data type, unit of measurement, and the UCUM standard code for the unit. Role assignments consist of the variable’s role in the current evidence pathway, generally including exposure variables (X), outcome variables (Y), mediating variables (M), causal variables (C), covariates (Z), and environmental or instrumental variables (E).

Appendix

Table 8: The four categories share a common subcategory overview.

Figure 4: Examples of evidence card files for each category.

Card type	Component	Parameter	Value / Definition
Mechanism	Glucose tolerance equation	$c^{\prime}$	$6.0\%$ increase
	Glucose tolerance equation	$b$	$-1.0$ (directional)
	Insulin sensitivity equation	$a$	$-1.0$ (directional)
Causal	MR-weighted median	$\tau$	$0.113$
Causal	MR-Egger	$\tau$	$0.095$
Interventional	Risk difference	RD	$P(Y{=}1\mid I{=}1)-P(Y{=}1\mid I{=}0)$

Appendix

Table 9: Registry of key parameters and hyperparameters across evidence card categories. Structural equations are summarized in Figure 4 ; this table reports numerical parameters, operational thresholds, and implementation-specific settings used in each evidence card.

Causal Reasoning

Causal reasoning studies how causes generate outcomes and how interventions and counterfactuals support explanation, prediction, and decision-making.

See all 79 papers

PROMETHEUS: Automating Deep Causal Research Integrating Text, Data and Models
May 13, 2026Sridhar MahadevanCausal World Models Causal Attribution
Adobe Research · University of Massachusetts, Amherst
EviDAG: Auditable Causal DAG Authoring with Biomedical Literature
Jul 23, 2026Yi-han Sheu, Michael R. Steigman, Yu Zhou +3Data Provenance Evidence-Grounded Reasoning
Massachusetts General Hospital, USA · Harvard University, USA · Northeastern University, USA
MedAgent-R1: Faithfulness-Aware Reinforcement Learning for Evidence-Grounded Medical Reasoning
Aug 31, 2026Jiangwang Chen, Chenghao Zhang, Hengxing CaiLLM Hallucination Mitigation Evidence-Grounded Reasoning
Tsinghua University · DP Technology

May 13, 2026J/K move · Enter open · S save

PROMETHEUS: Automating Deep Causal Research Integrating Text, Data and Models

Sridhar Mahadevan

Adobe Research · University of Massachusetts, Amherst

Causal World Models Causal Attribution Causal Reasoning in Language Models Causal Reasoning Causal Inference Large Language Model-Guided Scientific Discovery

Open paper AskPDF arXiv

Large language models can extract local causal claims from text, but those claims become more useful when organized as persistent, navigable world models rather than as flat summaries. We introduce PROMETHEUS, a framework that turns retrieved literature, filings, reviews, reports, agent traces, source data, code, simulations, and scientific models into causal atlases: sheaf-like families of local causal predictive-state models over an explicit cover of a research substrate. Each local region contains causal episodes, structured claim tables, predictive tests, support statistics, and provenance; restriction maps compare overlapping regions; gluing diagnostics expose agreement, drift, contradiction, and underdetermination. The resulting Topos World Model is not a single universal graph. It is a research instrument for navigating what a corpus says, where it says it, how strongly it is supported, and where local claims fail to assemble into a coherent global view. Three literature-atlas case studies -- ocean-temperature impacts on marine populations, GLP-1 weight-loss evidence, and resveratrol/red-wine health-benefit claims -- illustrate deep causal research from text with explicit locality, evidence, persistent state, and gluing tension. Four grounded-counterfactual case studies -- a Nature Climate Change microplastics forcing paper, an Indus Valley hydrology paper with VIC-derived figure data and model code, the canonical Sachs protein-signaling study with single-cell perturbation data, and a Nature singing-mouse study with MAPseq projection matrices -- show a stronger mode: when a paper ships source data, simulation outputs, or code, PROMETHEUS can evaluate a counterfactual against that scientific substrate and then rebuild the sheaf world model around the

Abstract

Causal Reasoning

Explore similar work

PROMETHEUS: Automating Deep Causal Research Integrating Text, Data and Models

EviDAG: Auditable Causal DAG Authoring with Biomedical Literature

MedAgent-R1: Faithfulness-Aware Reinforcement Learning for Evidence-Grounded Medical Reasoning