Auditing Sybil: Explaining Deep Lung Cancer Risk Prediction Through Generative Interventional Attributions

Bartlomiej Sobieski; Jakub Grzywaczewski; Karol Dobiczek; Mateusz W\'ojcik; Matthew Tivnan; Patryk Szatkowski; Przemyslaw Biecek; Przemys{\l}aw Bombi\'nski; Tomasz Bartczak

arxiv: 2602.02560 · v2 · pith:GBDOXIASnew · submitted 2026-01-30 · 💻 cs.LG · cs.AI· cs.CV

Auditing Sybil: Explaining Deep Lung Cancer Risk Prediction Through Generative Interventional Attributions

Bartlomiej Sobieski , Jakub Grzywaczewski , Karol Dobiczek , Mateusz W\'ojcik , Tomasz Bartczak , Patryk Szatkowski , Przemys{\l}aw Bombi\'nski , Matthew Tivnan

show 1 more author

Przemyslaw Biecek

This is my paper

Pith reviewed 2026-05-16 09:21 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CV

keywords Sybillung cancer risk predictioninterventional attributionsgenerative auditingdeep learningCT imagingmodel interpretabilitycausal verification

0 comments

The pith

Sybil lung cancer risk model differentiates malignant nodules like radiologists but shows sensitivity to artifacts and radial bias.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces S(H)NAP, a model-agnostic framework that applies generative interventions via 3D diffusion bridge modeling to CT scans in order to measure each feature's causal effect on Sybil's future risk score. This moves beyond observational correlations to test what the model actually uses for its decisions. A sympathetic reader would care because current clinical validations of such tools rely on correlations that can hide failure modes capable of producing incorrect risk assessments. The audit finds that Sybil often separates malignant from benign nodules in ways that align with expert judgment, yet it also exhibits dangerous sensitivity to clinically irrelevant artifacts and a consistent radial bias.

Core claim

By constructing realistic 3D diffusion bridge modifications that isolate object-specific changes and validating the resulting attributions with expert radiologists, the authors deliver the first interventional audit of Sybil and show that the model frequently behaves like an expert in distinguishing malignant pulmonary nodules from benign ones while also displaying critical failure modes of sensitivity to unjustified artifacts and a distinct radial bias.

What carries the argument

S(H)NAP, a generative interventional attribution framework that uses 3D diffusion bridge modeling to systematically alter anatomical features in CT scans and thereby isolate their causal contributions to the risk prediction.

If this is right

Sybil requires targeted corrections for its artifact sensitivity and radial bias before safe clinical deployment.
Interventional auditing can replace or supplement purely observational metrics when assessing deep learning tools for medical risk prediction.
Expert radiologist validation of generative attributions provides a practical way to confirm or refute the model's reasoning on specific cases.
Similar failure modes may exist in other deep models trained on CT data, making systematic causal audits necessary across the domain.
The shift to causal verification improves the reliability of automated screening decisions that affect patient management.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The radial bias could arise from systematic patterns in how training CT volumes are centered or reconstructed, suggesting a data-preprocessing fix.
Generative auditing frameworks like S(H)NAP could be adapted to audit other high-stakes medical AI systems for hidden confounders.
If the artifact sensitivity persists across different model architectures, it may indicate a broader limitation of current 3D convolutional approaches on CT data.
Routine use of such audits before deployment could reduce the risk that subtle, clinically irrelevant features drive screening recommendations.

Load-bearing premise

The modifications produced by the 3D diffusion bridge isolate genuine causal contributions without creating new confounding artifacts that the model could exploit.

What would settle it

A controlled set of diffusion-modified CT scans in which an added artifact or altered nodule changes the Sybil risk score in a manner that contradicts the expert-validated attribution map for that scan.

read the original abstract

Lung cancer remains the leading cause of cancer mortality, driving the development of automated screening tools to alleviate radiologist workload. Standing at the frontier of this effort is Sybil, a deep learning model capable of predicting future risk solely from computed tomography (CT) with high precision. However, despite extensive clinical validation, current assessments rely purely on observational metrics. This correlation-based approach overlooks the model's actual reasoning mechanism, necessitating a shift to causal verification to ensure robust decision-making before clinical deployment. We propose S(H)NAP, a model-agnostic auditing framework that constructs generative interventional attributions validated by expert radiologists. By leveraging realistic 3D diffusion bridge modeling to systematically modify anatomical features, our approach isolates object-specific causal contributions to the risk score. Providing the first interventional audit of Sybil, we demonstrate that while the model often exhibits behavior akin to an expert radiologist, differentiating malignant pulmonary nodules from benign ones, it suffers from critical failure modes, including dangerous sensitivity to clinically unjustified artifacts and a distinct radial bias.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

S(H)NAP gives a first interventional audit of Sybil via 3D diffusion edits but the causal claims rest on thin qualitative evidence without checks on edit fidelity.

read the letter

The paper's main move is to audit Sybil by using a 3D diffusion bridge to make targeted changes to CT volumes and measure shifts in the model's risk score. This produces attributions that are interventional rather than purely observational or gradient-based, and the authors pair it with expert radiologist review to flag both sensible behavior on malignant nodules and problems like artifact sensitivity and radial bias. That combination is new for this specific model and task. The approach is model-agnostic and directly targets safety questions that matter for clinical deployment. The execution is straightforward and the failure modes it surfaces are worth knowing. The soft spots are clear from the abstract and stress-test note. No quantitative results appear on how many cases were tested, what the expert agreement rates were, or whether the diffusion edits preserve distributional properties of real scans. Without perceptual metrics, realism tests, or controls for new high-frequency artifacts, a change in Sybil output could reflect sensitivity to synthetic noise rather than the intended anatomical feature. The weakest assumption—that the bridge isolates causal contributions cleanly—therefore stays untested in the reported work. This paper is for researchers who build or evaluate deep models for medical imaging and want practical auditing tools. It is not yet ready for strong claims about causal isolation, but the direction is useful and the method can be tightened. It deserves peer review so referees can ask for the missing quantitative controls and distributional checks.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces the S(H)NAP auditing framework, which uses 3D diffusion bridge modeling to generate interventional attributions for auditing the Sybil deep learning model for lung cancer risk prediction from CT scans. It claims to provide the first such interventional audit, showing that Sybil behaves similarly to expert radiologists in distinguishing malignant from benign nodules but has failure modes including sensitivity to clinically unjustified artifacts and a radial bias, validated by expert radiologists.

Significance. If the generative interventions faithfully isolate causal features without introducing artifacts, this approach could be highly significant for developing trustworthy AI systems in medical imaging by enabling causal explanations and identification of biases that observational methods miss. It applies existing diffusion models to a new auditing task in a high-stakes domain.

major comments (2)

[Abstract] Abstract: The central claim that the 3D diffusion bridge isolates object-specific causal contributions to Sybil's risk score is load-bearing but rests on the unverified assumption that edits leave the remainder of the CT volume distributionally unchanged; no quantitative distributional fidelity metrics (e.g., perceptual metrics or blinded realism tests) are reported to rule out new synthetic confounders that the model could latch onto.
[Expert validation description] Expert validation: The abstract and framework description provide no quantitative results, error bars, number of audited cases, or details on how expert agreement was measured; this leaves the demonstration of failure modes (artifact sensitivity and radial bias) as purely qualitative without visible controls.

minor comments (2)

The acronym S(H)NAP is introduced without expansion or explanation of its components, which may confuse readers unfamiliar with the framework.
[Abstract] Consider adding a brief statement on the specific number of CT volumes or nodules used in the audit to provide context for the scope of the qualitative validation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful comments, which help clarify how to better substantiate the causal claims and expert validation in our S(H)NAP auditing framework. We respond to each major comment below and indicate the revisions we will make.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that the 3D diffusion bridge isolates object-specific causal contributions to Sybil's risk score is load-bearing but rests on the unverified assumption that edits leave the remainder of the CT volume distributionally unchanged; no quantitative distributional fidelity metrics (e.g., perceptual metrics or blinded realism tests) are reported to rule out new synthetic confounders that the model could latch onto.

Authors: We agree that explicit quantitative evidence of distributional fidelity is necessary to support the interventional attributions. Although the 3D diffusion bridge is conditioned to preserve the original volume's distribution outside the edited region, we did not report supporting metrics in the initial submission. In the revised manuscript we will add Fréchet Inception Distance (FID) scores computed on lung patches and full volumes, together with results from a blinded expert realism study, to quantify fidelity and address the possibility of introduced confounders. revision: yes
Referee: [Expert validation description] Expert validation: The abstract and framework description provide no quantitative results, error bars, number of audited cases, or details on how expert agreement was measured; this leaves the demonstration of failure modes (artifact sensitivity and radial bias) as purely qualitative without visible controls.

Authors: We acknowledge that the current presentation of the expert validation is primarily qualitative and lacks the requested quantitative details. The manuscript describes confirmation of the identified failure modes by expert radiologists, but we agree that reporting the number of cases, inter-rater agreement statistics, and error bars would strengthen the section. We will revise the abstract, methods, and results to include these specifics (number of audited nodules, agreement metrics, and error bars on relevant summaries) so that the controls for the qualitative findings are explicit. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the auditing framework

full rationale

The paper introduces S(H)NAP as an auditing method that applies 3D diffusion bridge modeling to generate feature interventions on CT volumes and then measures resulting changes in Sybil's risk score. This chain does not reduce any claimed result to its own inputs by construction: the diffusion edits are produced by a separate generative process, the attribution is defined as the observed delta in model output, and expert validation is external to the equations. No self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations appear in the described derivation. The approach is an application of existing generative techniques to a new auditing task, remaining self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the assumption that 3D diffusion bridge modeling can generate realistic counterfactual CT scans that isolate specific anatomical features without introducing new artifacts; no free parameters or invented entities are explicitly quantified in the abstract.

axioms (1)

domain assumption Generative models can produce anatomically plausible modifications that preserve all non-targeted features.
Invoked to justify that changes in risk score reflect causal contribution of the modified feature.

invented entities (1)

S(H)NAP auditing framework no independent evidence
purpose: Model-agnostic method for constructing generative interventional attributions
New named framework introduced to perform the audit; no independent evidence provided beyond the paper's own application.

pith-pipeline@v0.9.0 · 5529 in / 1278 out tokens · 19322 ms · 2026-05-16T09:21:52.223195+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

By leveraging realistic 3D diffusion bridge modeling to systematically modify anatomical features, our approach isolates object-specific causal contributions to the risk score.
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Sybil employs a 3D ResNet18 encoder... attention mechanism... pairwise interactions over pulmonary nodules.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.