{"id":"9af5f1cf-3bac-4d69-928a-92e757e42075","arxiv_id":"2501.15973","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"PCF integrates causal Bayesian networks and probability-tree ensembles to deliver predictions plus intervention and counterfactual analysis on three healthcare datasets.","lead":"This paper proposes a framework, PCF, that combines causal Bayesian networks with probability trees to predict patient outcomes and to simulate what-if interventions. It applies the framework to ICU length of stay, heart disease, and diabetes data, with predictive accuracy comparable to standard machine learning models plus causal explanations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central causal-intervention claim is not established: forcing a branch in a single-parent empirical probability tree does not implement a Pearl do-intervention unless the learned CBN is causally correct and the tree encodes the joint distribution, neither of which is stated or tested.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the validity of treating a forced PTree branch as a causal intervention. I agree, and the concern is central because the paper's stated novelty is not predictive parity alone but 'causal reasoning capabilities' through interventions and counterfactuals. The predictive results in Tables 1-3 are credible but do not validate the causal claims, and no causal benchmark or ground-truth comparison is offered. The internal ambiguity about root vs. variable order further weakens the claim that the tree faithfully encodes the CBN's joint distribution. However, the gap is testable and potentially repairable: a synthetic-data check with a known structural causal model would settle whether the method's do-estimates correspond to true interventional effects. Since the paper could be made defensible by adding such validation and clearly stating identification assumptions, the conditional verdict is appropriate rather than outright rejection.","tokens_in":25312,"tokens_out":5000,"duration_ms":52521,"concrete_test":"Generate synthetic data from a known structural causal model with confounding, e.g., U -> B, U -> A, B -> A, with A as the outcome and B as the intervened variable; discretize all variables. Run the full PCF pipeline (Bayesys structure learning, model averaging, topological ordering, ensemble PTrees) and compute the Section 5.3 intervention estimate by setting the B-branch transition to 1. Compare this estimate to the true interventional distribution obtained by simulating the SCM with B fixed, and also report the naive conditional P(A|B) for reference. If PCF's do-estimate tracks the naive conditional rather than the true interventional distribution, the causal-interpretation claim is not supported; if it matches the true effect, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central added value over predictive ML is the claim in Section 5.3 that 'intervention involves the strategic modification of transition probabilities to ensure a specific event occurs with certainty' and that this yields P(A|do(B)). This inference is only valid if (i) the DAG learned from observational data is the true causal graph, with no unmeasured confounders and correct edge orientations; and (ii) the PTree's factorization exactly represents the CBN's joint distribution. Neither condition is stated, derived, or checked. On (ii), Section 3.2.1 computes transition probabilities as P(X_current | X_parent) using a single parent from a total topological order, but a node in the learned CBN can have several parents; a chain of single-parent conditionals does not in general equal the CBN joint. The description is also internally inconsistent: root partitioning is based on the target variable Y, while the CBN order is said to lead to Y, and Algorithm 1 uses variable order[0] as the current variable. Consequently the 'interventional' numbers in Figures 4-6 and the counterfactual statements in Section 5.4.2 are descriptive conditionals under a modified chain factorization, not causal effects. No causal baseline or ground-truth comparison is provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Probabilistic Causal Fusion (PCF) framework, which combines Causal Bayesian Network structure learning (via Bayesys model averaging) with an ensemble of probability trees whose variable ordering is taken from a topological sort of the learned graph. PCF is evaluated on three healthcare datasets (MIMIC-IV ICU length of stay, Framingham CHD risk, and BRFSS diabetes risk) and compared against conventional ML classifiers. The authors report comparable predictive accuracy/AUC and use the learned structure for sensitivity analysis, SHAP explanations, 'interventions' implemented by forcing PTree transition probabilities to 1, and 'counterfactual' statements following the probability-tree algorithms of Genewein et al. [8]. The central claim is that PCF provides a single tool that quantifies factor impacts and simulates hypothetical interventions from observational data while maintaining predictive performance comparable to standard ML.","tokens_in":25614,"tokens_out":5366,"duration_ms":50536,"significance":"If the causal claims were established, the framework would be a useful contribution: the predictive component is compared with several baselines on three real datasets and the reported metrics (Tables 1-3) support the 'comparable predictive performance' part of the claim. The method description is sufficiently detailed to permit re-implementation, and the use of external ML baselines avoids circularity. However, the distinctive added value—causal intervention and counterfactual reasoning from observational data—is not supported by the current derivation or validation. The intervention operation is defined purely as a change in an empirical tree's transition probabilities, with no causal identifiability conditions, no proof that the tree's factorization equals the CBN's joint distribution, and no causal baseline or ground-truth comparison. Thus the paper's significance currently rests on its predictive results plus a set of descriptive conditional analyses; the causal reasoning capability advertised in the abstract remains unsubstantiated.","major_comments":[{"comment":"","section":"Section 5.3 and Section 3.2.1"},{"comment":"","section":"Algorithm 1 and Section 3.2.1(b)"},{"comment":"","section":"Section 5.4.1 and 5.4.2"},{"comment":"","section":"Section 5.1"}],"minor_comments":[{"comment":"","section":"Section 5.3"},{"comment":"","section":"Figures 5 and 6"},{"comment":"","section":"Algorithm 2"},{"comment":"","section":"Section 3.1.2"},{"comment":"","section":"Section 5.4.2"},{"comment":"","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The causal inference contribution is the main selling point and is currently not established; the paper would be better positioned as a predictive modeling paper with descriptive conditional analyses unless the authors can supply identifiability conditions and validation for the intervention and counterfactual claims. I would encourage the editor to ask for a revision that either proves the correspondence between PTree interventions and causal effects under testable assumptions, or clearly reframes the claims as non-causal conditional modeling."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the predictive half of this paper is credible and the pipeline is a reasonable incremental combination of known pieces. The causal half—interventions and counterfactuals—is where it falls apart. The paper treats forcing a branch in an empirical probability tree as a Pearl-style do-operation, but never states or verifies the conditions under which that would be valid.\n\nWhat's actually new: the model-averaged CBN topological order is used to drive an ensemble of PTrees, with SHAP and sensitivity wrapping the whole thing. That specific combination doesn't appear in the cited prior art. The empirical work is real: three datasets, comparison against eight ML baselines, and PCF lands in the same performance band. Tables 1–3 support the 'comparable to ML' claim, though there are no error bars or significance tests, and hyperparameters like ensemble size, pruning threshold, and resampling choice are not reported in a reproducible way. No code is released either, which further limits reproducibility.\n\nThe causal claims, unfortunately, are not supported. Section 5.3 defines intervention as setting transition probabilities to 1 and writes the result as P(A|do(B)). That only works if the learned CBN is the true causal graph, all confounders are observed, and the tree's single-parent chain of conditionals equals the joint distribution. None of those is stated, derived, or tested. The internal description is also inconsistent: root partitioning uses the target variable Y, while the CBN order is said to lead to Y, and Algorithm 1 uses variable order[0] as current variable. The counterfactual sections (5.4) are descriptive conditionals under a modified chain, not true counterfactuals. The paper even acknowledges the features were chosen because the literature says they matter—that's soft circularity in the validation.\n\nThat said, the authors are not being deceptive. They cite the probability-tree causal reasoning work [8] and the fuzzy decision-tree hybrid [7], and the limitations section does mention some gaps, just not the ones that matter here. The paper reads like a serious attempt to combine tools, but the causal language overstates what the method delivers.\n\nWho this is for: a reader interested in interpretable clinical prediction might get value; a reader looking for causal inference will be disappointed. I'd send it to peer review—the predictive results and the pipeline are worth a serious look—but the causal claims need major revision, either by adding proper identification machinery or by reframing what the intervention numbers mean.\n\nRecommendation: engage with it as a conditional, requiring a thorough causal-identification check.","headline":"PCF is a plausible incremental pipeline with credible predictive results, but its causal intervention and counterfactual claims do not follow from the method as described.","tokens_in":26138,"tokens_out":2837,"would_cite":false,"duration_ms":23422,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PCF fuses causal Bayesian networks with probability trees to predict outcomes, quantify factor impacts, and simulate hypothetical interventions, with accuracy comparable to standard machine-learning models.","keywords":["causal Bayesian networks","probability trees","causal inference","interventional analysis","counterfactual analysis","healthcare machine learning","SHAP","ensemble learning"],"falsifier":"Compare the model's $P(\\mathrm{CHD}=1 \\mid \\mathrm{do}(\\mathrm{cigsPerDay}=0))$ estimate for heavy smokers against the observed 10-year coronary heart disease incidence among smokers who quit during the Framingham follow-up; if the model's counterfactual probability does not fall within the cohort's confidence interval, the tree-clamping operation is not reproducing a real intervention effect.","tokens_in":25069,"feed_emoji":"🩺","tokens_out":12083,"duration_ms":102614,"temperature":0.7,"pith_summary":"PCF is a proposed framework that joins Causal Bayesian Networks with ensembles of Probability Trees so that a single model can predict a clinical outcome, quantify how strongly each factor influences it, and simulate what would happen under a hypothetical intervention. The CBN supplies a causal ordering of variables; the trees express the resulting conditional probabilities as branches; interventions are represented by forcing a branch to probability 1 and reading the outcome as $P(A \\mid \\mathrm{do}(B))$. On the MIMIC-IV, Framingham Heart Study, and BRFSS-2015 diabetes datasets, the framework reports predictive accuracy and AUC comparable to standard classifiers such as gradient boosting, random forests, and logistic regression, while also producing sensitivity analyses and SHAP explanations of factor importance. The paper's point is that prediction and causal reasoning need not be separate toolkits: one tree-based model can give clinicians both a risk score and an answer to 'what if we changed this factor?'","feed_headline":"Match ML accuracy and answer what-if questions with fused causal trees","feed_subtitle":"From observational data, it matches ML predictions and simulates what-if interventions.","key_machinery":"The load-bearing object is the Probability Tree, a branching structure in which nodes are events and edges carry conditional probabilities, whose branching order is fixed by the topological order of a model-averaged Causal Bayesian Network. The CBN contributes the causal parent–child relationships that decide which variables branch before which, while empirical data determine the transition probabilities at each branch. Interventions are implemented by setting a chosen transition probability to 1, giving the paper's $P(A \\mid \\mathrm{do}(B))$ semantics; counterfactuals are implemented by altering the variable order or resetting factual states in a modified tree. The ensemble averages predictions over $k$ trees built on disjoint data subsets, and the averaged prediction is thresholded to produce a class label.","core_discovery":"The authors claim that the causal structure of a Causal Bayesian Network can fix the variable order of a Probability Tree, turning the tree into a reasoning tool rather than just a predictor. They learn several network structures from the data, average them into a single DAG (adding or reversing edges to avoid cycles), take its topological order, and build an ensemble of trees on random subsets of the data using that order. Transition probabilities are empirical counts of a variable given its CBN parent, and the root split reflects the marginal distribution of the target. An intervention is performed by setting a branch's transition probability to 1, which the paper writes as $P(A \\mid \\mathrm{do}(B))$, and counterfactuals are obtained by reordering variables or replacing observed states with hypothetical ones and recomputing path probabilities. On MIMIC-IV, Framingham, and BRFSS-2015, the method achieves accuracy and AUC-ROC within the range of the compared machine-learning baselines while adding sensitivity analysis, SHAP feature attribution, and the ability to simulate interventions.","pith_inferences":["Because each node's transition probability is conditioned on a single parent, the tree factorization may not capture the full joint distribution implied by the CBN, so an intervention in one branch could propagate differently than the network's multi-parent structure would dictate.","A direct way to test the causal layer would be to benchmark PCF's intervention curves against a known randomized effect, such as the Framingham cohort's observed outcomes following smoking cessation, and check whether the model's counterfactual reductions match the longitudinal data.","The same tree-building recipe could be applied to continuous outcomes, time-to-event data, or multi-category outcomes, and the intervention operation could be adapted to shift transition probabilities fractionally instead of forcing them to 1."],"forward_implications":["A clinician could query the model for the effect of lowering blood pressure, quitting smoking, or reducing BMI on the outcome and receive a probability change computed from the same model that gives the risk prediction.","The CBN-derived variable order removes the need for an expert to hand-specify the probability tree's branching sequence, replacing a subjective step with a data-driven one.","On the three datasets, the framework's accuracy and AUC-ROC are comparable to gradient boosting, random forests, support vector machines, and logistic regression, with a more balanced sensitivity–specificity profile than several baselines.","Sensitivity analysis and SHAP values can be produced from the same model, letting clinicians see both parameter-level and feature-level explanations.","Pre-trained PCF models with their learned causal graphs could be shared across hospitals, so smaller institutions can start from collective causal knowledge and refine it with local data."],"supporting_citations":[{"why":"Supplies the Bayesian network structure-learning and model-averaging machinery that produces the averaged DAG and its topological order.","marker":"[24]"},{"why":"Provides the probability-tree algorithms for causal reasoning and counterfactual computation that the PTree side of PCF builds on.","marker":"[8]"},{"why":"The earlier hybrid of probabilistic fuzzy decision trees with causal reasoning that PCF extends by making the tree construction data-driven.","marker":"[7]"},{"why":"Foundational treatment of causality and the do-operator that underwrites the paper's $P(A \\mid \\mathrm{do}(B))$ interpretation of interventions.","marker":"[1]"},{"why":"Supplies the MIMIC-IV preprocessing steps and the length-of-stay categorization into short and long stays used in the ICU case study.","marker":"[40]"},{"why":"Defines SHAP, the game-theoretic feature attribution method the framework wraps around the ensemble for interpretability.","marker":"[34]"},{"why":"SMOTE oversampling that the authors used to handle class imbalance in the MIMIC-IV and Framingham experiments.","marker":"[56]"},{"why":"ADASYN oversampling that gave the best imbalance handling for the diabetes dataset.","marker":"[57]"}],"fun_headline_variants":["Causal trees answer what-if clinical questions with ML-grade accuracy","New framework fuses causal networks and trees for what-if health analytics","Ask what-if: causal trees match ML and explain why","Probabilistic causal fusion: ML accuracy plus causal insight","Beyond prediction: causal trees for clinical what-if scenarios"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The causal claims stand on the assumption that the network learned from observational data is the true causal graph with all confounders observed, and that forcing a branch in the fitted tree to probability 1 is the same as intervening in the world.","fun_headline_variants_meta":{"raw":{"variants":["Causal trees answer what-if clinical questions with ML-grade accuracy","New framework fuses causal networks and trees for what-if health analytics","Ask what-if: causal trees match ML and explain why","Probabilistic causal fusion: ML accuracy plus causal insight","Beyond prediction: causal trees for clinical what-if scenarios"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000695,"raw_usage":{"total_tokens":3180,"prompt_tokens":1022,"completion_tokens":2158,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":2075}},"tokens_in":638,"tokens_out":2158,"duration_ms":13773,"temperature":1.0,"reasoning_tokens":2075,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:50:31.523833+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the model's $P(\\mathrm{CHD}=1 \\mid \\mathrm{do}(\\mathrm{cigsPerDay}=0))$ estimate for heavy smokers against the observed 10-year coronary heart disease incidence among smokers who quit during the Framingham follow-up; if the model's counterfactual probability does not fall within the cohort's confidence interval, the tree-clamping operation is not reproducing a real intervention effect.","supporting_citations":[{"cited_title":"Constantinou, The Bayesys user manual, http://bayesianai.eecs.qmul.ac.uk/bayesys/, [Online] Accessed 21 April 2023 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian network structure-learning and model-averaging machinery that produces the averaged DAG and its topological order."},{"cited_title":"Algorithms for Causal Reasoning in Probability Trees","cited_arxiv_id":"2010.12237","evidence_quote":"Provides the probability-tree algorithms for causal reasoning and counterfactual computation that the PTree side of PCF builds on."},{"cited_title":"Assisting clinical practice with fuzzy probabilistic decision trees","cited_arxiv_id":"2304.07788","evidence_quote":"The earlier hybrid of probabilistic fuzzy decision trees with causal reasoning that PCF extends by making the tree construction data-driven."},{"cited_title":"Pearl, Causality: Models, Reasoning and Inference, Cambridge University Press, (2009)","cited_arxiv_id":null,"evidence_quote":"Foundational treatment of causality and the do-operator that underwrites the paper's $P(A \\mid \\mathrm{do}(B))$ interpretation of interventions."},{"cited_title":"Hempel, S","cited_arxiv_id":null,"evidence_quote":"Supplies the MIMIC-IV preprocessing steps and the length-of-stay categorization into short and long stays used in the ICU case study."}],"review_version":1}