🤖 machine learning

Causal Modeling of Adverse Pregnancy Outcomes via Adaptive LLM Proposals

This paper introduces a neurosymbolic framework that iteratively combines Large Language Model-generated causal hypotheses with empirical data scoring to effectively discover and validate causal relationships for Adverse Pregnancy Outcomes, successfully recovering expert-validated edges and identifying novel plausible risk factors.

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

Published 2026-08-24

📖 4 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Understanding why a pregnancy might go wrong is one of the most difficult puzzles in modern medicine. Conditions like preterm birth or gestational diabetes do not happen in isolation; they are the result of a complex web of factors ranging from a mother's family history and lifestyle to her age and pre-existing health. To prevent these outcomes, doctors need to know which factors actually cause the others, rather than just appearing alongside them. This is the difference between seeing two things happen together and knowing that one pushes the other to happen. Finding these true causes is notoriously hard because medical data is often incomplete, noisy, and difficult to gather, while the number of possible connections between variables is so vast that computers struggle to sort through them all.

For years, researchers have tried to solve this by using two different tools. One approach relies entirely on data, looking for statistical patterns in patient records. The other uses the vast knowledge stored in large language models, which are computer systems trained on millions of medical texts and can suggest likely connections based on what they have read. However, neither method works perfectly on its own. Data-driven methods often fail when there is not enough information, while language models can be inconsistent, sometimes suggesting connections that sound plausible but are not supported by evidence. They can also get stuck on a single idea and miss the bigger picture.

A new study introduces a method called CLARA, which acts as a bridge between these two approaches. Instead of asking a computer to guess the answer once or to crunch numbers alone, the researchers created a system that treats the language model as a flexible guide. The system asks the model to generate a batch of possible cause-and-effect maps. It then checks each map against real clinical data to see how well it fits. The maps that fit the data best are not discarded; instead, they are used to teach the language model what a good answer looks like. The system then asks the model to try again, but this time with the instruction to focus on the patterns that worked well before. This cycle of guessing, checking, and refining repeats itself, slowly steering the computer away from wild guesses and toward a map that is both logically sound and supported by the evidence.

The researchers tested this method on a real-world dataset involving nearly 4,000 first-time mothers, tracking variables such as race, age, body mass index, family history of disease, and lifestyle habits like smoking or exercise. They compared the results of their new system against the work of human experts and standard computer algorithms. The findings were striking. The new method successfully recovered every single causal link that the human experts had identified, a feat that the standard computer algorithms failed to achieve. Furthermore, the system uncovered thirty additional connections that the experts had not listed. When a clinical specialist reviewed these new links, they confirmed that twenty-six were already established medical facts, while the remaining four were plausible hypotheses that fit with current medical understanding. These included connections suggesting that a history of polycystic ovary syndrome could contribute to high blood pressure, and that gestational diabetes might be a direct cause of new hypertension.

The study also showed that this approach is remarkably robust. In tests where the data was intentionally made messy or incomplete, the new method continued to find the correct connections, while traditional data-only methods became confused and produced errors. The researchers found that they could make the process even more efficient by summarizing the best results into a short list of common edges rather than feeding the entire complex maps back into the system, without losing any accuracy. This suggests that the method is not only effective but also practical for real-world use. By combining the broad knowledge of artificial intelligence with the rigorous testing of real data, the researchers have created a tool that can navigate the uncertainty of medical science to reveal the hidden causes of adverse pregnancy outcomes, offering a clearer path toward targeted interventions and better care for mothers and children.

Technical Summary: Causal Modeling of Adverse Pregnancy Outcomes via Adaptive LLM Proposals

Problem Statement

Reliable clinical reasoning requires robust causal models, yet these are rarely available in complex domains like obstetrics. Modeling Adverse Pregnancy Outcomes (APOs)—such as preterm birth, gestational diabetes, and preeclampsia—is particularly challenging due to the interplay of diverse risk factors (demographics, family history, lifestyle) and the incompleteness of current medical literature regarding overall causal dynamics.

Existing approaches face significant limitations:

Data-driven methods (e.g., PC, FCI, GES) struggle with the paucity of high-quality obstetric data, the presence of noise, and the super-exponential growth of possible graph structures. They often fail to recover plausible structures in data-scarce environments.
Large Language Models (LLMs) possess broad prior knowledge from vast training corpora but lack genuine causal reasoning capabilities. They often confuse association with causation, produce inconsistent outputs across runs, and are highly sensitive to prompting, leading to unstable one-shot graph generation.
Hybrid Theory Refinement methods, which refine an initial hypothesis via local edits, are sensitive to their starting points and prone to getting trapped in local optima if the initial hypothesis is poor.

Methodology: The CLARA Framework

The authors propose CLARA (Causal Learning via Adaptive Resampling and Aggregation), a neurosymbolic framework that treats an LLM as an adaptive proposal distribution. Inspired by the MIMIC (Mutual-Information-Maximizing Input Clustering) stochastic optimization algorithm, CLARA operates via a Generate-Evaluate-Update loop to iteratively shape the LLM's sampling distribution toward data-validated causal graphs.

The CLARA Loop

Sampling (Generation):
- The framework queries a pretrained LLM to generate a batch of $K$ candidate causal hypotheses.
- The LLM is prompted with the list of variables, their descriptions, and expert knowledge in the form of "forbidden edges" (constraints).
- Graphs violating acyclicity or forbidden edge constraints are discarded and resampled.
Evaluation:
- Each candidate graph is scored against the empirical dataset using the Tree BIC (Bayesian Information Criterion) score.
- Unlike standard BIC, which penalizes complexity based on the number of parent configurations (often leading to over-sparse graphs), Tree BIC represents local conditionals as decision trees. This accounts for Context-Specific Independencies (CSI), merging parent configurations that do not significantly reduce description length, thereby offering a more nuanced penalty for structure complexity.
- Graphs are ranked by their Tree BIC scores and stored in a history set $H$ .
Update (Adaptive Resampling):
- The top- $K$ highest-scoring graphs from the history set are used to construct an in-context representation $C$ .
- This representation can take two forms:
  - Full Graphs: The adjacency lists of the top-scoring graphs.
  - Common Edges: A compressed representation containing only edges present in all top- $K$ graphs.
- The LLM is prompted again with this updated context $C$ , effectively narrowing its sampling distribution toward high-confidence structures supported by both the LLM's prior and the empirical data.
Aggregation:
- After $T$ iterations, the top candidates in the history set are aggregated (via union) to form the final causal graph. Cycles introduced during aggregation are broken by removing the least frequent edge.

Key Contributions

Neurosymbolic Framework: Introduction of CLARA, which integrates approximate knowledge from LLMs, incomplete expert constraints, and small empirical datasets to generate plausible causal hypotheses.
Empirical Efficacy: Demonstration of the method's success on a real-world clinical dataset (nuMoM2b) for APOs. The framework recovers all expert-validated edges and identifies additional plausible causal relations not previously listed by experts.
Efficiency Analysis: Investigation into replacing full graph representations with "common edges" summaries, showing that prompt size can be significantly reduced without sacrificing performance.

Experimental Results

The authors evaluated CLARA on two datasets: a synthetic ALARM benchmark (37 nodes, 3,000 samples with 20% noise) and the real-world nuMoM2b dataset (9 risk factors, 4 outcomes, 3,856 examples).

Performance Metrics

The primary metric was Structural Intervention Distance (SID), which measures the error in predicted interventional distributions (critical for clinical reasoning), alongside Structural Hamming Distance (SHD), Recall, and Precision.

Comparison with Baselines: CLARA outperformed purely data-driven methods (PC, FCI), one-shot LLM generation, and LLM-initialized theory refinement.
- On the ALARM dataset, CLARA achieved the lowest SID, improving by 32% over one-shot LLM generation and 55% over data-driven discovery.
- On the nuMoM2b dataset, CLARA achieved a SID of 0.0, perfectly matching the expert-constructed graph's interventional predictions, while other methods (including theory refinement) showed significantly higher errors.
Robustness: CLARA maintained low and stable SID across varying dataset sizes and noise levels (up to 60% noise), whereas data-driven methods (PC, FCI) degraded significantly as noise increased or data decreased.
Efficiency: The "Common edges" variant reduced prompt token counts by approximately 3x while maintaining SID performance within 10% of the full-graph variant.

Clinical Validation

In the obstetrics domain, CLARA recovered all 31 expert-validated causal edges. It proposed 30 additional edges, which were reviewed by a clinical expert:

26 edges (87%) were rated as established medical knowledge.
4 edges (13%) were rated as plausible, including relationships such as PCOS $\to$ Preterm Birth and Gestational Diabetes $\to$ New Hypertension.
The framework successfully identified causal relationships among risk factors (e.g., Race $\to$ BMI) and sequential complication cascades among APOs.

Significance and Claims

The paper claims that CLARA effectively bridges the gap between the broad, stochastic priors of LLMs and the empirical rigor required in clinical settings. By framing causal discovery as an iterative optimization loop rather than a one-shot generation or a greedy local search, CLARA mitigates the unreliability of LLMs and the data-scarcity limitations of traditional algorithms.

The authors assert that their method provides a viable path for causal discovery in complex, data-sparse medical domains where pure data-driven approaches fail and pure LLM approaches are too unreliable. The framework offers a mechanism to "steer" LLM generations toward high-confidence regions of the hypothesis space, potentially uncovering new insights for targeted interventions while adhering to known medical constraints.

Future work suggested by the authors includes incorporating more expressive constraints (e.g., qualitative influence statements), developing hierarchical versions of the algorithm for high-dimensional domains, and exploring fine-tuning strategies for the underlying language models to specialize them further in causal hypothesis generation.