PaperlayerView exact revision on arXiv

Source-linked AI summary

Causal Modeling of Adverse Pregnancy Outcomes via Adaptive LLM Proposals

Kavimayil P. Komarasamy, Saurabh Mathur, Ameet Soni, David M. Haas, Kristian Kersting, Sriraam Natarajan

arXiv:2608.21079v1cs.LG

TL;DR

Causal discovery for adverse pregnancy outcomes is difficult because clinical knowledge is incomplete, obstetric data are scarce, and LLM proposals are inconsistent. CLARA iteratively samples and evaluates LLM-generated graphs, then updates the LLM context with high-scoring structures. On clinical and synthetic evaluations, it matched expert-validated edges and achieved stronger structure recovery than the reported baselines.

  • Problem

    Causal discovery for adverse pregnancy outcomes is limited by incomplete domain knowledge, sparse obstetric data, and unreliable LLM-generated causal graphs.

  • Method

    CLARA treats an LLM as an adaptive proposal distribution, scoring generated causal graphs against data and using high-scoring graphs to update subsequent generations.

  • Results

    CLARA recovers all 31 expert-validated causal edges, proposes 30 additional edges, and achieves the lowest SID across methods on ALARM and nuMoM2b.

  • Takeaways & Limitations

    Expert validation rated 26 of CLARA’s 30 additional edges as established and four as plausible, indicating alignment with obstetric domain knowledge.

Abstract

from arXiv · show

Adverse Pregnancy Outcomes (APOs) such as preterm birth and gestational diabetes can have long-term consequences for both the mother and child, yet an understanding of their causes remains elusive. Causal discovery in this domain is especially challenging due to a paucity of data and incomplete domain knowledge. As a result, pure data-driven methods fail, and Large Language Model (LLM) outputs remain inconsistent or contradictory. We introduce a neurosymbolic framework for generating plausible causal hypotheses that iteratively combines the broad prior knowledge of LLMs with empirical scoring on data. Our method treats the LLM as an adaptive proposal distribution, generating hypotheses that are scored against empirical data; the resulting high-scoring graphs are then used to update the LLM's context, steering subsequent generations toward more promising regions of the hypothesis space. We evaluate our approach on a real-world clinical dataset for modeling APOs and their risk factors, comparing our results against an expert-constructed causal graph. Our method recovers all expert-validated edges and identifies additional plausible causal relations not previously listed by experts, potentially providing new insights for targeted interventions.

I. INTRODUCTION

Causal discovery for adverse pregnancy outcomes is hindered by incomplete clinical knowledge, sparse obstetric data, and unreliable one-shot LLM proposals. CLARA combines LLM-generated hypotheses with empirical evaluation and iterative updating to improve causal graph generation.

  • APOs arise from diverse interacting risk factors, while existing medical knowledge remains incomplete about their overall causal dynamics.
  • Sparse high-quality obstetric data and the difficulty of interventional studies limit data-driven causal discovery.
  • LLM outputs can confuse association with causation and produce contradictory, empirically ungrounded graphs.
  • CLARA treats the LLM as an adaptive proposal distribution, scoring sampled causal structures against empirical data and updating subsequent prompts with high-scoring graphs.
  • On a real-world APO dataset, CLARA recovers expert-constructed edges and identifies additional plausible causal relations.

A. Causal Bayesian Networks

Causal Bayesian networks represent causal relations in directed acyclic graphs and support interventional reasoning. Existing approaches use data, expert knowledge, or LLM-generated hypotheses, while CLARA adapts distribution-based optimization into an iterative LLM-guided search.

  • Causal Bayesian networks encode direct causal relations in a directed acyclic graph and support interventional reasoning.
  • Data-driven methods learn causal graphs from observational data using structural assumptions such as faithfulness and causal sufficiency.
  • Theory refinement starts from an expert-informed graph and applies local edge additions, deletions, or reversals to improve an empirical structure score.
  • LLM-generated initial hypotheses are broad but stochastic and inconsistent, so local refinement can miss clinical nuances when starting from one hypothesis.
  • Estimation of Distribution Algorithms maintain and update a distribution over the search space, reducing the risk of becoming trapped in local optima.
  • MIMIC samples candidates, evaluates them, and learns common structural features of top-scoring samples to update the sampling distribution.
  • CLARA replaces explicit distribution learning with an LLM sampling distribution updated by changing its in-context examples to top-scoring samples.

III. CAUSAL LEARNING VIA ADAPTIVE RESAMPLING AND AGGREGATION

CLARA frames causal model construction as reconciling LLM knowledge, incomplete expert constraints, and empirical data. Its iterative procedure addresses the incompleteness and unreliability of these information sources.

  • The target is a causal graph over discrete variables that reconciles a dataset, forbidden edges from incomplete expert knowledge, and a pretrained LLM.
  • Clinical expertise is precise but siloed, observational data are scarce, and LLM outputs are stochastic and unreliable for causal graph generation.
  • CLARA uses a Sample-Evaluate-Update loop to shape a pretrained LLM’s broad prior toward data-validated plausible causal graphs.

A. Sampling Causal Graphs

CLARA samples causal hypotheses from an LLM and evaluates them against observational data using a structure score. Tree BIC accounts for context-specific structure while penalizing graph complexity.

  • Sampling Causal Graphs: CLARA samples candidate causal graphs from the LLM using variables, descriptions, and forbidden-edge constraints.Candidates violating forbidden-edge or acyclicity constraints are discarded and resampled.
  • Sampling Causal Graphs: Bayesian Information Criterion (BIC) scores each proposed graph by balancing data fit against model complexity.The score uses the data log-likelihood and penalizes graph structures with more parameters.
  • Sampling Causal Graphs: Tree BIC replaces full conditional-probability-table penalties with decision-tree leaf counts to represent context-specific independencies.This can reduce the complexity penalty when parent configurations share conditional structure.
  • Sampling Causal Graphs: Graphs are ranked by Tree BIC, and only the highest-scoring candidates contribute to the next generation’s context.This ranking connects empirical evaluation to subsequent LLM sampling.

C. Updating the Sampling Distribution

CLARA updates LLM generation iteratively by conditioning new proposals on high-scoring causal graphs. After the iterations, it aggregates the best candidates into a final graph.

  • Updating the Sampling Distribution: CLARA uses adjacency-list representations so in-context examples share the directed-edge semantics of the LLM’s graph outputs.A compressed common-edges representation is evaluated as an efficiency alternative.
  • Updating the Sampling Distribution: CLARA initializes a history of candidate graphs, scores them, and repeatedly selects the top K graphs for representation.The algorithm uses a Generate-Evaluate-Update loop rather than relying on one LLM output.
  • Updating the Sampling Distribution: INDUCEREPRESENTATION constructs context from the current top-K graphs, while SAMPLEFROMLLM generates new graphs conditioned on that context.The context can encode full graphs or common edges.
  • Updating the Sampling Distribution: After T iterations, CLARA combines the top-K high-scoring graphs with a union-based strategy and breaks resulting cycles to construct the final graph.The aggregation and cycle-breaking procedures are described separately in the paper.

IV. EMPIRICAL EVALUATION

The empirical evaluation tests whether CLARA improves causal hypotheses across synthetic and clinical settings. It examines comparisons with LLM-based, search-based, data-driven, and representation variants.

  • IV. EMPIRICAL EVALUATION: The evaluation asks whether combining LLM priors with search outperforms purely LLM-based generation, pure structure search, and LLM-initialized local search.
  • IV. EMPIRICAL EVALUATION: It tests whether CLARA performs better than data-driven causal discovery in data-scarce and noisy domains.
  • IV. EMPIRICAL EVALUATION: The evaluation also examines whether structural summaries reduce prompt size while maintaining performance.
  • IV. EMPIRICAL EVALUATION: A final question assesses whether CLARA proposes plausible causal hypotheses for real obstetrics data that closely match domain knowledge.The supplied passage truncates the comparison’s final wording.

A. Dataset Description

The study evaluates CLARA on synthetic ALARM data and clinical nuMoM2b data. The clinical analysis focuses on early-pregnancy risk factors and adverse pregnancy outcomes.

  • Dataset Description: The synthetic dataset is based on the 37-node, 46-edge ALARM Bayesian network and contains 3,000 sampled observations with 20% entry-wise corruption.
  • Dataset Description: The clinical dataset comes from nuMoM2b, a longitudinal study of first-time mothers with singleton pregnancies.
  • Dataset Description: The clinical analysis uses first-prenatal-visit variables to model early-pregnancy risk factors’ causal impact on adverse pregnancy outcomes.The risk factors span demographics, family history, pre-existing conditions, and lifestyle domains.
  • Dataset Description: 3,856 examples remained after excluding data points with missing values for the structural search.

B. Method and Baselines

CLARA is compared with data-driven, one-shot LLM, theory-refinement, and non-LLM iterative-search baselines, alongside component variants using common-edge representations.

  • Baselines: CLARA is compared with PC and FCI data-driven discovery, one-shot LLM generation, LLM-augmented theory refinement, and MIMIC.MIMIC isolates the iterative search loop from CLARA’s LLM prior.
  • Graph handling: Cycle-breaking removes the lowest-frequency edge among five responses, with lexicographic tie-breaking.For nuMoM2b, the forbidden edge set encodes temporal-order constraints; ALARM has no blacklist edges.
  • Variants: The evaluation includes fully implemented CLARA and a common-edges variant that represents only adjacencies shared across all top-K scoring graphs.The common-edge representation enforces direct adjacencies and reflects full agreement across the top-K set.

C. Evaluation Metrics

The study evaluates learned graphs against synthetic ground truth and an expert-constructed obstetrics graph, prioritizing causal-intervention agreement over structural similarity alone.

  • Reference graphs: Structural correctness is assessed against a synthetic reference graph and an expert-constructed graph based on current obstetric medical consensus.The expert graph encodes relationships known to the clinicians at the time of evaluation.
  • Metrics: The reported metrics are Structural Hamming Distance, Precision, Recall, and Structural Intervention Distance.SHD, Precision, and Recall measure topological alignment, while SID measures differences in causal conclusions.
  • Metric priority: SID is the primary metric because a graph with higher SHD can still support correct interventional reasoning.SHD is treated as secondary when making comparative claims.

D. Results and Discussion

Across synthetic and real-world evaluations, CLARA improves causal agreement over competing approaches and remains stable under data scarcity and noise. It also recovers all expert-validated obstetric edges while proposing additional clinically supported relations.

  • Results: 32% better SID than one-shot LLM generation, 55% better than data-driven discovery, and 28% better than hybrid theory refinement on ALARM.CLARA achieves the lowest SID across all methods, with the improvement observed for both evaluated LLMs.
  • Results: CLARA achieves the lowest SID across all methods on nuMoM2b and outperforms one-shot, data-driven, and hybrid theory-refinement baselines.MIMIC marginally outperforms purely data-driven baselines on nuMoM2b but has the worst SID on ALARM.
  • Robustness: CLARA maintains low, stable SID across dataset sizes and noise levels, whereas PC and FCI remain elevated or degrade under higher noise.PC and FCI fail to match CLARA’s SID even when the dataset size is tripled.
  • Efficiency: Up to three times fewer prompt tokens preserves SID close to full CLARA on three of four dataset–LLM combinations.On ALARM with GPT-5.2, the common-edges variant is 10% worse than CLARA but 21% better than other baselines.
  • Clinical discoveries: CLARA recovers all 31 expert-validated causal edges and proposes 30 additional edges, of which 26 are rated established and four plausible.The four plausible relations are PCOS→PTB, HiBP→GDM, DiabHist→PTB, and Race→PhyAc.
  • Clinical discoveries: Additional hypotheses include risk-factor relationships involving family history, race, and PCOS that are supported by cited clinical or population evidence.Examples include effects on BMI, high blood pressure, physical activity, smoking, and preeclampsia.
  • Clinical discoveries: CLARA also recovers direct effects that may represent mediated relationships when intermediate variables are absent from the dataset.Examples include maternal age and pre-pregnancy BMI affecting preterm birth, and PCOS affecting preeclampsia and preterm birth.
  • Clinical discoveries: A proposed APO cascade links gestational diabetes to new hypertension and new hypertension to preterm birth.The relations are supported by evidence involving shared metabolic mechanisms and a prospective cohort.

V. CONCLUSION

CLARA combines pretrained LLM priors with empirical structure scoring to generate plausible causal hypotheses for data-scarce APO modeling. The framework is presented as effective, while extensions remain open for richer clinical constraints, scalable search, and more specialized language models.

  • CLARA reconciles pretrained LLM knowledge with empirical structure scoring to generate plausible causal hypotheses for Adverse Pregnancy Outcomes.
  • The framework uses a Sample-Evaluate-Update loop to mitigate stochastic one-shot proposals and address data sparsity in complex clinical domains.
  • Figure 3 evaluates SID across dataset sizes and noise levels for PC, FCI, and CLARA; lower SID indicates better causal-structure recovery.
  • Hyperparameter sensitivity is assessed through mean and standard deviation of SID across combinations of top-K graphs and iteration counts on ALARM.
  • Future work includes specialized language-model fine-tuning to improve the reliability of causal-hypothesis generation.
Loading 2608.21079v1…