acceptodds
Withdrawn from ICLR 2027

Time-aware Neuro-symbolic Abduction with Structured Conformal Prediction for Biomedical Causal Discovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unraveling deep mechanistic pathways underlying disease progression and drug action is a core pursuit of modern pharmacology and targeted therapeutics. However, transforming sparse and asynchronous clinical observations into statistically verifiable causal structures remains a long-standing challenge.Existing paradigms relying heavily on static knowledge graphs or large language models (LLMs) suffer from critical limitations: structural hallucinations due to insufficient causal constraints, search explosion in open-domain literature mining, rapid knowledge obsolescence for dynamic clinical scenarios, and the absence of distribution-free statistical safety bounds for causal predictions.In this paper, we introduce TRACE, a unified time-aware neuro-symbolic abductive framework integrated with structured conformal prediction, which transforms the passive candidate path validation to active open-domain causal discovery. Concretely, a time-aware hierarchical topological representation module that aligns sparse and asynchronous clinical observations into explicit causal subgraphs is first designed to resolve the mismatch between static knowledge representations and dynamic clinical data. To mitigate search explosion and semantic inconsistency in open-domain evidence retrieval, we propose an active state-graph backward search algorithm to guide multi-hop causal chain assembly across heterogeneous literature streams.Then, we incorporate macro-skill discovery and temporal update mechanism with continual learning objectives, which reduces online computational overhead and mitigates knowledge aging by continuously distilling verified reasoning trajectories.Furthermore, we establish a structured conformal prediction module tailored for multi-hop directed acyclic graphs, which defines path-level joint non-conformity scores via atomic triple decoupling and constructs distribution-free confidence bounds with finite-sample coverage guarantees, addressing the lack of statistical reliability in existing causal path predictions.We validate TRACE across four benchmark datasets: MIMIC, ADE-Time, CoMAGC, and FB15k-Subset. Experimental results demonstrate that TRACE achieves an absolute MRR gain of up to 10.4% over state-of-the-art baselines, while strictly maintaining the target 95% statistical coverage rate guaranteed by the conformal prediction module, reducing average prediction set sizes by over 50.0% via tighter confidence bounds, and accelerating inference speed by 54% through macro-skill reuse.

Primary area:
causal reasoning
Keywords:
biomedical causal discovery, neuro-symbolic abduction, conformal prediction, time-aware

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ZYcfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This article studies the problem of discovering multi hop biomedical causal mechanisms from sparse and asynchronous clinical observations, proposes the TRACE framework, and combines time aware graph construction, literature verification, MCTS reverse causality search, macro skill reuse, and structured conformal prediction. The core viewpoint of the author is to narrow down the candidate mechanism space through time and knowledge constraints, then use literature evidence to support active search, and use conformal predictive control to control the uncertainty of output. The paper reported good candidate ranking performance on MIMIC, ADE Time, and CoMAGC, and reported near target 95% coverage on MIMIC and FB15k Subset, while demonstrating the inference acceleration brought by macro skill reuse. The research question has practical value, and combining mechanism search with uncertainty control is also attractive. However, the relationship between "candidate mechanism credibility", "true causal structure", and "statistical coverage guarantee" has not been fully established in current papers. There are still some key issues regarding method definition, theoretical basis, and experimental verifiability.

Critical strengths

The paper reported good candidate ranking performance on MIMIC, ADE Time, and CoMAGC, and reported near target 95% coverage on MIMIC and FB15k Subset, while demonstrating the inference acceleration brought by macro skill reuse. The research problem has practical value, and combining mechanism search with uncertainty control also has certain appeal

Critical weaknesses
  1. The coverage objects and prediction sets of structured conformal prediction are not fully defined, and existing algorithms are insufficient to support the core guarantee of the paper. Equations (10) to (12) and the pseudocode did not use yi, only calculating the validation score of the graph itself. The triple probability product of equation (10) can be used as a heuristic inconsistency score, but without a dependency structure explanation, it cannot be directly explained as a joint probability that the entire path holds.

  2. The construction of a time perception graph has not yet established an effective identification process from clinical events to causal structures. The time similarity in equation (3) is symmetrical and can express the degree of event proximity, but it cannot determine the causal direction itself. Although formula (4) introduces Bayesian structural scoring and acyclic constraints, it does not define how clinical events are composed of random variables and repeated observation samples, nor does it explain the specific form of likelihood, missing value handling, and causal identification assumptions.

  3. Insufficient information on dataset construction and baseline adaptation makes it difficult to determine the reliability and fairness of the main results. Section 4.1 only briefly lists four data sources, without providing the task sample size, input-output instances, annotation process of real root causes or paths, candidate space size, and complete partitioning of training, validation, calibration, and testing sets.

The implementation of active open domain search and continuous updates is not yet sufficient to support its method proposition. Section 3.3 allows the language model to propose new upstream reasons, but equation (7) relies on the Rotate embedding of candidate entities. The paper does not explain how new entities outside the knowledge graph are represented, nor does it distinguish between "discovering new relationships between known entities" and "discovering unincorporated entities and mechanisms".

    1. The existing experiments can provide preliminary signals of performance improvement, but are not sufficient to validate the core mechanisms and statistical conclusions. In Table 2, the absolute MRR improvement of TRACE relative to the best baseline of each dataset is 0.090, 0.080, and 0.069, respectively, which cannot directly correspond to the highest 10.4% state-of-the-art baseline improvement mentioned in the abstract. The author should clarify and unify the comparison criteria. The coverage rates of 95.2% and 95.4% in Table 3 do not have matching test sample sizes, confidence intervals, or duplicate partitioning results, so it cannot be claimed to "strictly maintain" target coverage based solely on point estimates; The results of 94.8% and 94.9% in Table 4 cannot be explained as module removal leading to guarantee failure in the absence of statistical analysis. If the MRR decreases simultaneously after removing the conformal module in Table 4, it is necessary to explain whether conformal gating has changed the sorting, evaluation samples, or training process. The 54.3% acceleration brought by macro skills has potential value, but should be included in the cost of building and maintaining the skill library, reporting hit rate, cold start performance, and accuracy under the same time budget, and compared with regular path caching.
Questions for authors
  1. Can provide clear prediction objectives, set construction, coverage events, and theorem proofs, and report the overall coverage rate, rejection rate, and error rate in the accepted samples separately?

  2. Clarify the semantics of nodes and edges, implement global constraints on the merged graph, and report the accuracy of edge direction, structural distance, and performance under mixed conditions on simulated data or expert annotated data with known structures?

  3. Can compare greedy, bundle search, MCTS, and caching strategies under a fixed validator, compare different conformal scores under a fixed candidate set, and report coverage performance divided by path length, disease category, and time?

1 comment

My native language is Chinese. After reading this manuscript, I provided my opinion in Chinese and had it translated into English by AI. The editor seems to believe that this is an AI generated opinion, and they are requesting me to make further comments. But I think their viewpoint is obviously inappropriate.

Reviewer_ouk4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TRACE, a pipeline for finding causal pathways behind clinical observations. Timestamped events are clustered using temporal and PubMedBERT similarity, and a local DAG is learned for each cluster with a knowledge graph prior. An MCTS backward search then queries an LLM for upstream causes, filters them with UMLS normalization and a transfer probability, and scores paths with a trained verifier. Verified sub-paths are stored as reusable macro-skills with time-decayed evidence weights. A structured conformal module scores each graph by its weakest path and either accepts the graph or abstains. Experiments on MIMIC, ADE-Time, CoMAGC, and an FB15k subset report MRR gains over KG, RAG, and abduction baselines, 95% coverage with smaller sets than CSP, and lower latency from skill reuse.

Critical strengths
  • The problem matters. Moving from checking given KG paths to searching for unknown upstream causes, with an option to abstain, is a useful goal for pharmacovigilance.
  • The ablation table covers every module, and the paper reports latency, which is rare for LLM-based pipelines.
Critical weaknesses
  • The conformal guarantee is not stated correctly. The calibration score Φ(G_i) does not use the label y_i, so it is unclear what event the coverage refers to. A single accept or abstain decision on one graph is not a prediction set, yet set sizes are reported. Coverage of what, over which candidates, is never defined. Clinical events over time are also not exchangeable, which the authors only mention as a limitation.
  • The evaluation cannot be checked. The paper does not explain how MIMIC is turned into an abduction task with ground-truth root causes, or how ADE-Time is built. CoMAGC is cited as an anonymous 2025 PRCV paper. FB15k is not biomedical, although the paper calls all four datasets biomedical benchmarks. No variance, seeds, or significance tests are given. The baseline settings and the LLM backbone are not reported.
  • The method is not causal discovery in any formal sense. Edges come from LLM proposals, literature similarity, and a learned verifier. The paper gives no identification argument or link to interventional meaning. The local Bayesian DAG step does not say what data the likelihood is computed on. Eq. 7 multiplies a scalar by a document vector, so P_trans is not a probability as written.
  • The showcased case study contains a wrong edge. It ranks the path “hyperkalemia to acute kidney injury” with confidence 0.92. In this setting, kidney injury causes hyperkalemia, not the reverse, and losartan affects the kidney directly through renal hemodynamics. A confidently accepted reversed edge undercuts the claim that the system reduces structural hallucinations.
Questions for authors
  • What exactly is covered with probability 1 − ε? Please define the label, the candidate set, and the prediction set, and explain why the calibration score does not depend on y_i.
  • How are ground-truth root causes and pathways defined for MIMIC and ADE-Time? How many test cases are there?
  • Which LLM is used? How were the baselines configured and tuned?
Reviewer_4J8eOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes TRACE, a time-aware neuro-symbolic framework for biomedical causal discovery that integrates temporal graph construction, MCTS-based abductive reasoning, macro-skill reuse, and structured conformal prediction. The framework aims to identify literature-supported multi-hop causal pathways while improving inference efficiency and uncertainty estimation. Experiments on MIMIC, ADE-Time, CoMAGC, and FB15k-Subset report improvements in MRR, Hit@K, prediction-set efficiency, and inference latency compared with existing baselines.

Critical strengths

The paper presents a comprehensive framework integrating time-aware graph modeling, MCTS-based abductive reasoning, and structured conformal prediction for biomedical causal discovery. Its main strengths are the evidence-guided exploration of multi-hop causal pathways and the consistent improvements in MRR and Hit@K across multiple benchmarks. The macro-skill mechanism also demonstrates substantial inference efficiency gains, supporting the framework's potential for complex biomedical reasoning tasks.

Critical weaknesses
  1. Several mathematical formulations are insufficiently defined and appear to contain substantial methodological flaws. In particular, the proposed framework conflates temporal association with causality. In Eq. (5), causal edges are inferred from temporal precedence ((t_u < t_v)) and knowledge graph connectivity. However, temporal ordering does not establish causality, as (P(Y\mid X)\neq P(Y\mid do(X))). The authors should clarify the underlying causal assumptions, explain how confounding factors are addressed, and provide stronger validation of the inferred causal relationships.

  2. The causal pathway presented in the case study is inconsistent with the temporal observations. The inferred pathway suggests "Hyperkalemia → AKI," whereas Figure 5 indicates that AKI occurred on Day 10 and hyperkalemia on Day 12. This directly contradicts the framework's assumption that causes must precede their effects.

  3. Furthermore, losartan-associated AKI is commonly related to reduced intraglomerular pressure, while hyperkalemia may result from impaired renal potassium excretion associated with AKI or aldosterone suppression. Therefore, the proposed causal direction requires stronger clinical justification.

  4. The manuscript contains several inconsistencies and reproducibility concerns. Section 3 inconsistently reports the number of components and steps, while Appendix B.1 is empty and duplicates B.2's title. Several hyperparameters (k=60k=60, M=5M=5, θagent=0.7\theta_{\mathrm{agent}}=0.7) are insufficiently explained, and mathematical notation is inconsistent or reused ambiguously (e.g., α\alpha, cpuctc_{\mathrm{puct}} vs. cexplorec_{\mathrm{explore}}). Furthermore, FB15k is incorrectly characterized as a biomedical benchmark, MIMIC-IV requires credentialed access despite claims of open accessibility, and some references contain errors (e.g., the GraphRAG first author's name). These issues undermine methodological clarity and reproducibility.

Questions for authors

Could the authors clarify how ground-truth labels are incorporated into conformal calibration and how the claimed 95% causal-path coverage is guaranteed? How does the framework distinguish temporal associations from genuine causal relationships while accounting for confounding factors? Additionally, how do the authors explain the inconsistency between the inferred causal pathway and the clinical timeline in Figure 5? Finally, could they provide more details on ground-truth construction, dataset splits, and hyperparameter settings to ensure reproducibility?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.