Causally Constrained and Uncertainty-Aware Cardiopulmonary Risk Prediction

Yuxiang Lai, Jiashuo Zhang, Liang Zhao, Xiaofeng YangarXiv 2026 · arXiv

This paper introduces the Causal Envelope Agent (CE-Agent), a framework designed to predict both cardiovascular disease (CVD) and lung cancer risk from a single routine chest CT scan. By combining established epidemiological knowledge with modern machine learning, the authors create a model that is more robust, interpretable, and data-efficient than standard "black-box" approaches.

TL;DR

CE-Agent uses "Causal Envelopes"—mathematical constraints derived from medical literature—to prevent machine learning models from learning biologically impossible associations. This approach allows for accurate, uncertainty-aware prediction of both lung cancer and heart disease from a single scan, outperforming standard clinical calculators even when patient records are incomplete.

What Problem Does This Paper Address?

Predicting cardiopulmonary risk is currently hampered by three major issues:

  1. Siloed Tools: Clinical calculators usually evaluate one disease at a time (e.g., just heart disease) and often ignore rich data available in chest CT scans.
  2. The "Statin Paradox" (Confounding): Unconstrained ML models often learn "wrong" directions for risk factors. For example, because high-risk patients are prescribed statins to lower their cholesterol, a naive model might incorrectly learn that low cholesterol is a risk factor for death.
  3. Missing Data: Real-world patient records are often incomplete. Standard methods for handling missing data (like Monte Carlo sampling) are too slow for real-time clinical use or fail to provide calibrated uncertainty intervals.

Methods

The CE-Agent operates in a two-stage framework:

  • Stage I: Causal Envelope Compilation: Before looking at any patient outcomes, the system uses a Directed Acyclic Graph (DAG) to screen variables and compiles peer-reviewed epidemiological evidence (hazard and odds ratios from PubMed) into a "Causal Envelope." This envelope defines the allowed range and direction for every risk factor's weight.
  • Stage II: Constrained Estimation: The model then fits weights to local cohort data, but it is strictly "constrained" to stay within the Causal Envelope. This prevents the model from overfitting or learning confounded associations.
Overview of the CE-Agent framework, showing the two-stage process from prior compilation to uncertainty-aware triage.
  • Analytic Uncertainty Propagation: Instead of slow simulations, the authors derived a closed-form formula to propagate uncertainty from missing clinical or CT inputs. This allows the model to return a "90% risk interval" in milliseconds.

Datasets

The authors evaluated CE-Agent on two major US cohorts:

  • National Lung Screening Trial (NLST): 53,452 participants from 33 hospitals, focusing on heavy smokers.
  • Multi-Ethnic Study of Atherosclerosis (MESA): 6,814 adults from four racial/ethnic groups, providing a diverse validation set for cardiovascular outcomes.

Key Insights and Ideas

  • Knowledge as a Regularizer: By using medical literature to bound the model, the authors effectively "pre-train" the model's logic. This lowered the required sample size for reliable model development by up to fivefold.
  • Confounding Protection: The Causal Envelope successfully "zeroed out" confounded weights. In cases where unconstrained models learned that a larger pulmonary artery (a sign of lung disease) was "protective" against heart disease due to complex correlations, CE-Agent maintained biologically sound directions.
  • Triage Efficiency: The model's uncertainty estimates can guide clinical workflow. If the 90% risk interval crosses a treatment threshold, the system can specifically flag that patient for a manual chart review or a specialized CT read.
Comparison of CE-Agent weights vs. unconstrained fits, illustrating how the Causal Envelope keeps weights within biologically plausible ranges.

Key Results

  • Superior Accuracy: CE-Agent outperformed standard cardiovascular calculators (like the Pooled Cohort Equations) and matched or exceeded unconstrained ML models.
  • Calibration Correction: It corrected the well-known 1.7-fold risk overestimation of the Pooled Cohort Equations, achieving an expected-to-observed event ratio of 1.04.
  • Speed: The analytic uncertainty propagation ran over 200 times faster than Monte Carlo sampling, making it viable for high-volume screening environments.
  • Generalization: A model trained at a single hospital generalized to 32 other hospitals better than unconstrained models, thanks to the stability provided by the causal constraints.

Limitations and Future Work

While CE-Agent is highly effective, it relies on the availability of high-quality epidemiological priors. If the literature for a specific sub-population is missing or biased, the "envelope" may be suboptimal.

Future work includes extending the framework to incorporate longitudinal health records (serial CT scans) and integrating it with automated 3D segmentation tools to further reduce the manual burden on radiologists. The authors also suggest that this "agentic" approach—where a model follows a fixed, auditable sequence of modules—is a promising path for clinical AI that requires high levels of trust and transparency.

Summary updated

Abstract

Routine chest computed tomography (CT) captures pulmonary, cardiac, and vascular phenotypes that jointly stratify lung cancer and cardiovascular disease, the two leading causes of death in lung-screening populations. Yet clinical risk calculators evaluate each disease in isolation and omit CT findings, whereas unconstrained machine learning overfits small hospital cohorts, learns treatment- and illness-confounded risk directions, and conceals uncertainty when patient records are incomplete. Here we present the Causal Envelope Agent (CE-Agent), a two-stage framework that compiles peer-reviewed epidemiological evidence into explicit constraints on risk-factor weights before any outcome is seen, while estimating local baseline risk and newly added CT phenotypes from cohort data. At deployment, closed-form uncertainty propagation accounts for missing clinical or CT inputs within milliseconds per patient, running over two orders of magnitude faster than Monte Carlo sampling and returning calibrated risk intervals that direct chart review and selective imaging to borderline patients. Across 53,452 lung-screening participants from 33 US hospitals and 6,814 multi-ethnic adults, CE-Agent outperforms cardiovascular calculators in lung screening, exceeds the Pooled Cohort Equations on hard atherosclerotic events while correcting their 1.7-fold risk overestimation, and generalizes from a single hospital or CT scanner to external sites while lowering required sample sizes up to fivefold. Applied unchanged to lung cancer, CE-Agent matches established screening models with accurate calibration even when restricted to risk-factor directions alone, and pairing both predictions improves joint cardiopulmonary triage from a single screening visit.
BibTeX
@misc{lai2026causally,
  author = {Yuxiang Lai and Jiashuo Zhang and Liang Zhao and Xiaofeng Yang},
  title = {{Causally Constrained and Uncertainty-Aware Cardiopulmonary Risk Prediction}},
  year = {2026},
  eprint = {2611.17819},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17819},
}
Figure