acceptodds
Withdrawn from ICLR 2027

From Measurements to Concepts: LLM-Guided Latent Concept Causal Discovery for Medical Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Causal discovery provides a structural basis for intervention and counterfactual analysis, making it valuable for understanding complex disease processes from observational medical data. However, medical datasets often contain multiple measurements reflecting shared but unobserved clinical states. Existing latent-variable causal discovery methods primarily search for hidden structure within predefined statistical model classes, but do not explicitly determine the appropriate level of causal abstraction. We formulate this problem as causal representation selection, coupling causal representation selection with causal graph learning. We propose LLCD, a framework explicitly addresses latent-aware causal discovery from this representation-selection perspective. LLCD uses LLM-derived semantics to propose candidate latent concepts, learns them with a missing-aware variational model, and selects among alternative latent–observed representations using held-out structural adequacy. We evaluate LLCD on semi-synthetic benchmarks with known latent causal structure and real-world ADNI and PPMI cohorts, assessing latent concept recovery, causal structure recovery, and causal interventional analysis. On the semi-synthetic benchmarks, LLCD improves directed F1 by \(66.0%\) and \(63.5%\) over the strongest baselines on ADNI and PPMI, respectively. LLCD also reduces causal-effect NMAE by approximately \(20%\) on both datasets. The code has been made public.

Primary area:
causal reasoning
Keywords:
Causal Discovery, LLM, Healthcare

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GNUSOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper tackles representation selection in medical causal modeling: deciding whether observed indicators should be retained as distinct variables or abstracted into latent clinical concepts. The proposed framework, LLCD, uses LLMs to derive candidate concepts, variable groupings, and causal tiers from metadata, learns latent representations via a missingness-aware variational model, and executes BIC-based graph search under tier constraints. Final representations and graphs are selected based on validation residual correlation and structural complexity. Experiments on ADNI, PPMI, and semi-synthetic benchmarks evaluate latent and graph recovery, SCM fitting, and interventional estimation. Although LLCD performs favorably, the distinct contributions of individual components and the sufficiency of the causal evaluation require further scrutiny.

Critical strengths
  1. The investigated problem carries clear clinical relevance, and the proposed framework provides a coherent, end-to-end pipeline tailored to representation selection. In medical domains, clinical conditions are frequently manifested across multiple surrogate indicators; thus, optimizing the causal node space itself is both well-motivated and practically meaningful. Crucially, the authors recognize that raw BIC scores across non-identical variable spaces are not directly comparable, appropriately using validation structural goodness-of-fit to govern representation selection: a perspective that offers genuine methodological insight.

  2. The empirical evaluation covers multiple complementary dimensions with encouraging initial results. Beyond indicator grouping, the paper assesses causal graph recovery and interventional effect estimation. Table 2 benchmarks against diverse latent causal discovery baselines, while Table 3 reports strong projected directed F1 scores of 0.873 and 0.981 on two semi-synthetic benchmarks. Methodological details such as subject-level data splits on real cohorts and reporting variance across five random seeds lend credible support to the framework's feasibility.

Critical weaknesses
  1. Insufficient ablation of core contributions and baselines: Table 5 lacks a key ablation isolating the latent module from the tier prior (retaining latents without hierarchy). Crucially, the paper lacks two essential controls under identical candidate groupings: (1) a baseline that naively retains all candidate concepts, and (2) classical dimensionality reduction (PCA/Factor Analysis) with identical groupings and tier priors. Given that the latent-free baseline on ADNI already reaches a directed F1 of 0.835 (vs. 0.873 for the full model), controlled ablations are necessary to prove that gains stem from the selection criterion rather than grouping or directional priors. Additionally, the latent baselines in Table 2 should be carried over to Tables 3 and 4.

  2. Potential intra-instance information leakage and constrained causal verification: In Eq. (14) and Appendix A.2.2, latent parents are inferred from indicator sets and then used to predict those same indicators. This risks target information leakage where high R2R^2 and low NLL merely reflect autoencoding/reconstruction artifacts rather than structural discovery: an issue that subject-level splits cannot resolve. Furthermore, quantitative causal evaluation is limited to two linear-Gaussian semi-synthetic benchmarks, with only one intervened variable per dataset and no reported variance, leaving the model's robustness to noisy tier priors or varied causal queries unverified.

Questions for authors

Q1. Evaluation protocol & target leakage (Appendix A.2.2 / Table 3): Does inferring the latent parent of a test indicator XiX_i utilize XiX_i itself from the same instance? If so, please provide results under a leave-one-out / target-masked inference protocol to eliminate circular reconstruction gains and ensure fair structural evaluation.

Q2. Controlled baselines for core contributions: Under identical groupings, representations, and search constraints, what is the performance gap between the proposed selection strategy and (1) naively retaining all candidate latents, or (2) classical PCA / Factor Analysis baselines? This is necessary to isolate the true value of representation selection and variational learning.

Q3. Sensitivity to noisy or absent priors: Please report the missing ablation in which latent modeling is kept, but tier priors are removed. Additionally, how robust is the framework to misspecified groupings or incorrect hierarchical constraints: can the data-driven stage reject invalid hypotheses?

Q4. Breadth and variance of interventional evaluation (Table 4): Please evaluate a broader spectrum of pre-specified intervention targets and strengths on the semi-synthetic systems, including error intervals across multiple seeds, to confirm that the reported interventional gains are consistent and not query-dependent.

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

LLCD presented latent-aware causal discovery on clinical data as representation selection. From variable metadata, an LLM proposes KK constructs (indicator sets Ik\mathcal I_k, context sets Uk\mathcal U_k) and a causal layering ℓ\ell. Each r∈{0,1}Kr \in \lbrace 0,1 \rbrace^K either keeps Ik\mathcal I_k observed or replaces it with ZkZ_k, learned by a mask-aware, context-conditioned VAE whose posterior mean z^k\hat z_k (d=2d=2 for ADNI, d=4d=4 for PPMI) becomes a vector node. LLCD then runs layer-constrained linear-Gaussian BIC search and selects the sparsest rr within ϵ\epsilon of the best held-out residual-dependence score vCIv_{\mathrm{CI}}.

Critical strengths
  1. In clinical data, the question is which semantically meaningful grouping should be a causal node. The paper correctly notes that BIC cannot compare different node spaces (Eq.2Eq. 2) and formalizes the choice as selection over r∈{0,1}Kr \in \lbrace 0,1 \rbrace^K. This framing is useful.

  2. Prompts, ground-truth graphs, projection rules, and the generator, which is decoupled from the VAE (Eqs.23–30)(Eqs. 23–30), are specified. This makes the work reproducible and its flaws diagnosable.

Critical weaknesses
  1. z^k\hat z_k is a re-encoding. dk=∣Ik∣d_k = |\mathcal I_k| for 6 of 8 ADNI latent nodes, and dk>∣Ik∣d_k > |\mathcal I_k| for 3 of 4 PPMI construct types. With β=10−3\beta = 10^{-3}, the encoder can losslessly encode its indicators (plus context). LLCD is grouped-variable structure learning over an LLM-fixed partition; the one-dimensional ground-truth factor (Eq. 24) is not estimated.

  2. Replacing a group of indicators Ik\mathcal I_k by one latent ZkZ_k is valid only if (i) the indicators are independent of each other once ZkZ_k is known, and (ii) no other variable affects or is affected by a single indicator directly, bypassing ZkZ_k. But vCIv_{\mathrm{CI}} is computed after the indicators have been replaced (Eqs. 18–19), so these are not tested on data. The complexity term C(r)=∣E∣+λZ∣Zr∣C(r) = |E| + \lambda_Z |Z_r| systematically rewards using latents: collapsing ∣Ik∣|\mathcal I_k| nodes into one removes all their edges, while the added cost is only λZ=1\lambda_Z = 1.

  3. ADNI's answer is "select all", and the partition cannot change, so C-Jac and Meas-F1 equal PureLLM's (Table 2). The only real decision, PPMI's spurious DAT construct, is retained in 100% of seeds at d=2d = 2 and 20% at d=4d = 4 (Fig. 3). This contradicts Table 2's ΔK=0.0±0.0\Delta K = 0.0 \pm 0.0.

  4. In the downstream SCM, LLCD predicts each indicator XkjX_{kj} from its latent z^k\hat z_k through the measurement edge Zk→XkjZ_k \to X_{kj} (A.2.2). But z^k\hat z_k is the encoder's output computed from those same test indicators (Eq. 14). LLCD is scored on reconstructing its own inputs. This explains the size of the reported gains: real PPMI R2R^2 rises when latents are added, and synthetic NLL drops (ADNI and PPMI) (Table 5). It also explains why NLL keeps falling as dd grows while F1\text{F1} does not (Fig. 2): a larger code reconstructs the indicators better without improving the graph. Since every real-data result in the paper rests on these metrics, none of them currently supports the method.

  5. Most of the structure-recovery gain comes from the LLM's ordering prior and the evaluation protocol, not from latent modelling. The synthetic ground-truth graph follows amyloid→tau→neurodegeneration→cognition\text{amyloid} \to \text{tau} \to \text{neurodegeneration} \to \text{cognition}, which is exactly the ordering the LLM layer prior ℓ\ell supplies. Because the model is linear-Gaussian, the data alone cannot determine most edge directions, so directed F1\text{F1} largely measures whether ℓ\ell matches the authors' own ground truth. The ablation (Table 5) confirms this: plain BIC hill-climbing without latents or ℓ\ell already scores 0.7100.710 (ADNI) and 0.6990.699 (PPMI), higher than every baseline in Table 3. The latent-variable baselines listed in Sec. 5.1 are not included in Tables 3–4, so the claimed gains are measured only against methods that ignore latents.

Questions for authors
  1. Could you recompute Tables 3 and 5 so that each test indicator is predicted only from its latent's structural parents, i.e. from E[Zk∣Pa(Zk)]\mathbb{E}[Z_k \mid \mathrm{Pa}(Z_k)], never from z^k\hat z_k computed on that same indicator. Does LLCD still outperform "w/o latent" and the baselines?

  2. Could you add a semi-synthetic case where an LLM-plausible group is not a common-cause block, e.g. an indicator with its own parent or an indicator that is not caused by the construct (ICV). Does LLCD reject that latent?

  3. Table 2 reports ΔK=0\Delta K = 0 on PPMI at d=4d = 4, but Fig. 3 shows the spurious DAT construct selected in 11 of 55 seeds at d=4d = 4. Which data does Fig. 3 use? Was dd chosen using ground-truth F1\text{F1} (Fig. 2)?

  4. Please list all Uk\mathcal U_k, and state whether the LLM's suggested_edges\verb|suggested_edges| are used anywhere. I think there is a need to ablate an encoder that does not receive the context: how many of the recovered edges between ZkZ_k and Uk\mathcal U_k remain?

  5. In the PPMI ground truth, NP3TOT_ON_OFF→NP3TOT\verb|NP3TOT_ON_OFF| \to \verb|NP3TOT| is a direct edge into one indicator. LLCD cannot represent such an edge, yet it reports SHD=0\text{SHD} = 0 on PPMI. How was this edge scored?

Reviewer_Y4ENOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

LLCD is a framework for causal model selection with latent variables, in which an LLM proposes latent concepts and causal ordering constraints. Each candidate representation is constructed using variational modelling and then associated with a causal graph learned through constrained BIC-based search. The final representation is selected by minimizing model complexity among those exhibiting low residual correlation on validation data.

Critical strengths

The paper addresses a relevant and interesting problem in latent-variable causal discovery and presents several strengths, both in its methodological perspective and empirical evaluation.

First, the paper introduces an interesting perspective on causal representation selection by considering alternative representations involving observed and latent variables, rather than assuming a fixed causal variable space. The integration of LLM-derived semantic knowledge with data-driven causal discovery is particularly relevant in medical applications, where domain knowledge can help identify clinically meaningful latent constructs. The proposed LLCD framework combines latent concept proposal, missing-aware variational representation learning, causal graph discovery, and latent variables selection within a structured methodological pipeline.

Second, the experimental evaluation is conducted on two relevant medical datasets, ADNI and PPMI, using both semi-synthetic benchmarks with known latent causal structures and real-world cohorts. The experiments cover several aspects of the framework, including latent concept recovery, causal structure recovery, and intervention-effect estimation. The reported improvements over existing baselines are promising, particularly in terms of directed F1 and intervention-effect estimation errors. The inclusion of comparisons with established causal discovery methods and some ablation studies further strengthens the empirical evaluation.

Critical weaknesses

Despite its interesting perspective and promising empirical results, the paper raises several methodological concerns, particularly regarding the causal interpretation of the learned representations and the experimental validation of its main contribution.

Major weaknesses:

First, the causal interpretation of the learned latent representations remains insufficiently justified. The proposed framework learns latent constructs through separate context-conditioned VAEs, while their causal relationships are estimated only afterward. It is therefore unclear to what extent the learned representations correspond to meaningful causal factors rather than statistically informative embeddings. Moreover, the role of the auxiliary context variables within the underlying SCM is not clearly specified. These variables may capture population heterogeneity or distribution shifts, but their effects could operate directly on the latent constructs, on their observed measurement indicators, or on both. These alternatives correspond to different causal assumptions and may have important consequences for the interpretation of the learned graph and subsequent interventions. The authors should clarify the causal positioning of context variables and better justify the separation between latent representation learning and causal structure discovery.

Second, the experimental evaluation does not sufficiently isolate the contribution of the proposed causal representation selection mechanism, which appears to be the central novelty of the paper. Although the authors provide ablations on latent modeling and causal ordering constraints, they do not directly compare the proposed selection strategy with fixed representations, such as retaining all candidate latent constructs, or with alternative selection criteria. The contribution of context conditioning is also not evaluated independently, and the ablations of latent modeling and causal ordering constraints are not fully disentangled. Consequently, the reported improvements demonstrate the effectiveness of the overall pipeline but do not establish which components are responsible for the gains. Furthermore, the semi-synthetic benchmarks rely on predefined clinically meaningful latent structures, potentially favoring the LLM-generated proposals. Experiments involving incorrect or redundant latent candidates, alternative representations, and imperfect semantic priors would be particularly valuable for assessing the robustness and actual benefit of the selection mechanism.

Minor weaknesses:

Several additional methodological choices would benefit from further justification. In particular, the use of VAEs for relatively low-dimensional datasets, with some latent constructs inferred from only two observed indicators, raises questions about their added value compared with simpler factor analysis or latent-variable modeling approaches. The authors also impose the same latent dimension on all constructs within each dataset, despite their different clinical meanings and numbers of indicators. The interpretation of these multiple latent dimensions, and their treatment as a single causal node, remains unclear. Although the datasets contain longitudinal measurements, the proposed approach does not explicitly model temporal dynamics beyond causal ordering constraints. Finally, the intervention experiments would benefit from a clearer discussion of the causal interpretation of the interventions, particularly the intervention on the follow-up interval in PPMI.

Questions for authors
  1. Contribution of the causal representation selection mechanism:

The main contribution of LLCD appears to be the selection of an appropriate causal representation among alternative combinations of observed and latent variables. However, the current ablation studies do not directly isolate the contribution of this selection mechanism. How much of the reported improvement can be attributed specifically to representation selection, rather than to latent modeling, LLM-derived semantic knowledge, or causal ordering constraints? In particular, how would LLCD perform if all admissible candidate latent constructs were retained without applying the proposed selection criterion? Such a comparison would be important for assessing the paper's central contribution.

  1. Causal interpretation of the learned latent representations:

The latent constructs are learned through separate context-conditioned VAEs, while their causal relationships are estimated afterward. What justifies interpreting these learned representations as causal variables rather than statistically informative embeddings? In particular, what assumptions ensure that the learned representations preserve the causal relationships and intervention semantics of the underlying system? Could the authors clarify why LLM-derived causal knowledge is not incorporated directly into the representation-learning objective, as in causally structured latent representation approaches?

  1. Role and causal positioning of context variables:

The auxiliary context variables may capture population heterogeneity or distribution shifts, but their causal role within the underlying SCM remains unclear. Are these variables assumed to influence the latent constructs, their observed measurement indicators, or both? These alternatives correspond to different measurement and causal models. How is this distinction handled in LLCD, and what prevents context conditioning from introducing dependencies that may compromise the causal interpretation of the learned latents? Additionally, what evidence supports the benefit of context conditioning, independently of the other components?

  1. Validity of the representation selection criterion:

The proposed criterion compares candidate representations using average absolute residual correlations on held-out data. However, different representations involve different variable spaces, latent dimensions, and sets of residual independence constraints. How do the authors justify the comparability of these scores across representations? Moreover, since zero correlation does not generally imply independence outside specific distributional assumptions, under what conditions can this criterion reliably identify an appropriate causal representation? Could different or overly dense graphs obtain favorable scores simply because they induce fewer testable constraints?

  1. Robustness to imperfect latent concept proposals:

The semi-synthetic benchmarks rely on predefined clinically meaningful latent structures, which appear relatively aligned with the concepts proposed by the LLMs. How robust is the representation selection mechanism when the LLM proposes incorrect, redundant, overlapping, or missing latent constructs? Does the method reliably reject inappropriate candidates, and how does this affect causal graph recovery? Clarifying this point, ideally through controlled perturbations of the candidate set, would help establish whether the proposed selection mechanism is effective beyond favorable semantic priors.

  1. Justification of the latent modeling choices:

The datasets contain relatively few observed variables, and some clinical constructs are associated with only two measurement indicators. What motivates the use of VAEs rather than simpler factor models or latent-variable SEM approaches in this setting? Furthermore, why is the same latent dimension assigned to all constructs within each dataset, despite differences in their clinical meaning and measurement structure? How should the multiple dimensions of a given construct be interpreted when the construct is subsequently treated as a single causal node?

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

Related papers

Discussion (0)

Sign in to comment.