arxiv: 2605.07785 · v2 · submitted 2026-05-08 · 💻 cs.CV

Recognition: no theorem link

Radiologist-Guided Causal Concept Bottleneck Models for Chest X-Ray Interpretation

Amy Rafferty , Rishi Ramaesh , Ajitha Rajan

Authors on Pith no claims yet

Pith reviewed 2026-05-15 06:22 UTC · model grok-4.3

classification 💻 cs.CV

keywords concept bottleneck modelscausal modelschest X-raymedical imaging interpretabilitynoisy-ORradiologist guidanceMIMIC-CXRpathology classification

0 comments

The pith

XpertCausal uses radiologist-curated causal structure and a noisy-OR model to improve accuracy, calibration, and clinical alignment in chest X-ray concept bottleneck models.

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

The paper introduces XpertCausal, a concept bottleneck model that explicitly represents how pathologies generate observable radiographic concepts through a probabilistic noisy-OR process. Radiologist-defined associations constrain the model to clinically plausible pathways, after which Bayesian inference recovers pathology probabilities from the predicted concepts. Evaluated on MIMIC-CXR, the approach yields higher AUROC, better calibration, stronger explanation quality, and concept-pathology relationships closer to expert knowledge than both non-causal baselines and unconstrained causal variants. A sympathetic reader would care because the method embeds domain knowledge directly into the generative structure rather than learning associations from data alone, offering a route to more trustworthy and interpretable medical imaging models.

Core claim

XpertCausal models pathology-to-concept relationships with a radiologist-constrained noisy-OR generative process and inverts this model via Bayesian inference to obtain pathology probabilities from concept predictions; the resulting structure produces improved AUROC, calibration, and explanation quality on MIMIC-CXR while aligning concept-pathology links more closely with expert knowledge than non-causal CBMs or unconstrained causal ablations.

What carries the argument

The radiologist-guided noisy-OR generative model that encodes pathology-to-concept causal relationships and is inverted through Bayesian inference to predict pathologies from concepts.

If this is right

The model learns concept-pathology relationships that more closely match expert knowledge than data-driven alternatives.
Calibration improves, reducing overconfident pathology predictions.
Explanation quality rises because the generated concept activations follow clinically plausible pathways.
The same causal-inversion approach can be applied to other imaging modalities that admit similar pathology-to-finding generative models.

Where Pith is reading between the lines

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

Incorporating expert causal structure may reduce the volume of labeled data needed for training compared with purely discriminative models.
The framework could be extended to flag cases where observed concepts deviate from the modeled generative pathways, suggesting alternative diagnoses.
Similar radiologist-guided constraints might improve concept bottleneck models in non-radiology domains such as pathology slides or retinal images.

Load-bearing premise

The radiologist-curated concept-pathology associations accurately reflect the true clinical generative process and the noisy-OR approximation sufficiently captures how pathologies produce radiographic findings.

What would settle it

An independent set of radiologist annotations on the same images that contradict the curated associations, or a performance drop on a new chest X-ray dataset whose causal structure differs from the modeled pathways.

Figures

Figures reproduced from arXiv: 2605.07785 by Ajitha Rajan, Amy Rafferty, Rishi Ramaesh.

**Figure 1.** Figure 1: Comparison of model architectures. All models share a common InceptionV3 concept prediction model which maps chest X-rays to [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Figure 2: Example explanations from each approach for a CXR with [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Mean proportion of ground truth concepts captured by the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

read the original abstract

Concept Bottleneck Models (CBMs) in medical imaging aim to improve model interpretability by predicting intermediate clinical concepts before final diagnoses. However, most existing CBMs treat concepts as discriminative predictors of pathology labels, without explicitly modelling the underlying clinical generative process where diseases produce observable radiographic findings. We propose XpertCausal, a radiologist-guided causal CBM for chest X-ray interpretation which models pathology-to-concept relationships using a probabilistic noisy-OR framework. This generative model is then inverted via Bayesian inference to estimate pathology probabilities from predicted concepts. Radiologist-curated concept-pathology associations are used to constrain model structure to radiologist-defined clinically plausible reasoning pathways. We evaluate XpertCausal on MIMIC-CXR across pathology classification performance, calibration, explanation quality, and alignment with radiologist-defined reasoning pathways. Compared with both a non-causal CBM baseline and a causal ablation with unconstrained learned associations, XpertCausal achieves improved AUROC, calibration, and clinically relevant explanation quality, while learning concept-pathology relationships that more closely align with expert knowledge. These results demonstrate that incorporating clinically motivated causal structure and expert domain knowledge into CBMs can lead to more accurate, interpretable, and clinically aligned models for CXR interpretation.

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

XpertCausal adds radiologist-curated noisy-OR structure to CBMs for CXR and claims better alignment plus calibration, but the abstract gives no numbers or stats to judge the size of those gains.

read the letter

The paper's core move is to replace the usual discriminative concept-to-label step in CBMs with a generative noisy-OR model that runs from pathologies to radiographic concepts, then inverts it with Bayes to recover pathology probabilities. Radiologists supply the allowed edges, so the structure is not learned freely. That combination is new relative to prior CBM work on chest X-rays. The controls against a plain CBM and an unconstrained causal version are sensible, and the goal of producing explanations that track expert reasoning pathways is reasonable for this domain. If the full results hold up, the approach could be useful for teams that already have access to radiologist annotations and want models whose internal paths are easier to audit. The main soft spot is that the abstract reports AUROC, calibration, and alignment improvements without any deltas, confidence intervals, or test details, which makes it impossible to tell whether the gains are large enough to matter clinically or just small shifts. The noisy-OR independence assumption is also worth pressure: a single finding such as opacity can be produced by interacting pathologies, and violating conditional independence could bias the inverted posteriors even if the radiologist edges are correct. Without seeing the actual tables or sensitivity checks on the curated associations, it is hard to know how much the reported advantages depend on the specific approximation rather than the causal framing itself. This is the kind of paper that belongs in a reading group focused on interpretable medical models, mainly for the structural idea rather than as a finished method. It deserves a serious referee because the topic is high-stakes and the controls are in place, but the review will need to focus on the quantitative evidence and the robustness of the generative assumption.

Referee Report

3 major / 2 minor

Summary. The paper proposes XpertCausal, a radiologist-guided causal concept bottleneck model for chest X-ray interpretation. It constrains a noisy-OR generative model of pathology-to-concept relationships using radiologist-curated associations, inverts the model via Bayesian inference to obtain pathology posteriors from predicted concepts, and reports improved AUROC, calibration, explanation quality, and alignment with expert reasoning pathways on MIMIC-CXR relative to a non-causal CBM baseline and an unconstrained causal ablation.

Significance. If the performance and alignment gains hold under rigorous evaluation, the work would demonstrate that embedding expert-curated causal structure into CBMs can produce more clinically aligned and interpretable models for medical imaging, addressing a key limitation of purely discriminative concept-based approaches.

major comments (3)

[§4.2] §4.2 and Table 2: the abstract and results claim improved AUROC and calibration for XpertCausal, yet no quantitative deltas, confidence intervals, error bars, or statistical significance tests versus the two baselines are supplied; without these, the central empirical claim cannot be assessed for robustness.
[§3.2] §3.2, Eq. (3)–(5): the noisy-OR generative model assumes conditional independence of concept failures given pathologies, but the manuscript provides no analysis or ablation testing whether this approximation holds for multi-pathology interactions (e.g., overlapping opacities); violation would bias the Bayesian inversion and undermine the reported calibration and explanation-quality gains.
[§4.3] §4.3: the claim that learned concept-pathology relationships align more closely with expert knowledge is supported only by qualitative comparison to the radiologist-curated graph; no quantitative metric (e.g., edge overlap, causal effect correlation) or inter-rater validation of the curated associations is reported, leaving the alignment result ungrounded.

minor comments (2)

[Abstract] The abstract states performance gains but supplies no numerical values; the results section should include a concise table of key metrics with deltas in the abstract or introduction for readability.
[§3.3] Notation for the Bayesian inversion step (posterior P(pathology|concept)) is introduced without an explicit equation reference; adding a numbered equation would clarify the inference procedure.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights opportunities to strengthen the empirical and methodological rigor of our work. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [§4.2] §4.2 and Table 2: the abstract and results claim improved AUROC and calibration for XpertCausal, yet no quantitative deltas, confidence intervals, error bars, or statistical significance tests versus the two baselines are supplied; without these, the central empirical claim cannot be assessed for robustness.

Authors: We agree that the absence of quantitative deltas, confidence intervals, error bars, and statistical significance tests limits the interpretability of the reported gains. In the revised manuscript we will augment Table 2 with mean AUROC and ECE differences (with 95% bootstrap confidence intervals) relative to both baselines, include error bars on all bar plots, and report p-values from paired statistical tests (e.g., DeLong or Wilcoxon signed-rank) to establish robustness of the improvements. revision: yes
Referee: [§3.2] §3.2, Eq. (3)–(5): the noisy-OR generative model assumes conditional independence of concept failures given pathologies, but the manuscript provides no analysis or ablation testing whether this approximation holds for multi-pathology interactions (e.g., overlapping opacities); violation would bias the Bayesian inversion and undermine the reported calibration and explanation-quality gains.

Authors: The conditional-independence assumption is a standard modeling choice in noisy-OR formulations to maintain tractable inference. We acknowledge that its validity for co-occurring pathologies has not been explicitly tested. We will add a targeted ablation that (i) identifies cases with overlapping opacities, (ii) compares posterior calibration and explanation fidelity under the current independent noisy-OR versus a relaxed model that introduces limited dependence (via a small set of learned interaction terms), and (iii) quantifies any degradation attributable to the independence assumption. revision: yes
Referee: [§4.3] §4.3: the claim that learned concept-pathology relationships align more closely with expert knowledge is supported only by qualitative comparison to the radiologist-curated graph; no quantitative metric (e.g., edge overlap, causal effect correlation) or inter-rater validation of the curated associations is reported, leaving the alignment result ungrounded.

Authors: We will introduce quantitative alignment metrics, specifically the fraction of learned edges that exactly match the radiologist-curated graph and the Spearman correlation between learned and curated causal strengths. The associations were curated by a single board-certified radiologist; we will explicitly state this limitation in the revised text and note that multi-rater validation, while desirable, was outside the scope of the present study. The performance and calibration improvements provide indirect support for the clinical plausibility of the learned structure. revision: partial

Circularity Check

0 steps flagged

No significant circularity: derivation relies on external radiologist curation and standard Bayesian inversion

full rationale

The paper's core chain—predicting concepts then inverting a radiologist-constrained noisy-OR generative model via Bayesian inference to obtain pathology posteriors—draws its structure from external expert annotations and a standard probabilistic framework rather than from quantities fitted inside the paper. Performance gains are measured empirically against non-causal and unconstrained baselines on MIMIC-CXR, with no equations or claims reducing the reported AUROC, calibration, or alignment metrics to the model's own inputs by definition. No self-citations are invoked as load-bearing uniqueness theorems, and the noisy-OR approximation is presented as an explicit modeling choice whose validity is tested rather than assumed tautologically.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim depends on the validity of the noisy-OR generative assumption and the accuracy of expert-curated associations; no free parameters or new entities are explicitly introduced in the abstract.

axioms (2)

domain assumption Noisy-OR framework sufficiently models how pathologies generate observable radiographic concepts
Used to define the generative process that is then inverted
domain assumption Radiologist-curated concept-pathology associations are clinically accurate and complete
Used to constrain model structure to plausible reasoning pathways

pith-pipeline@v0.9.0 · 5523 in / 1236 out tokens · 44174 ms · 2026-05-15T06:22:38.440483+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

[1]

Ivanov, Alexey Ko- rnaev, and Ivan Titov

[Alukaevet al., 2023 ] Danis Alukaev, Semen Kiselev, Ilya Pershin, Bulat Ibragimov, Vladimir V . Ivanov, Alexey Ko- rnaev, and Ivan Titov. Cross-modal conceptualization in bottleneck models. InThe 2023 Conference on Empirical Methods in Natural Language Processing,

work page 2023
[2]

Structural causal bottleneck models,

[Binget al., 2026 ] Simon Bing, Jonas Wahl, and Jakob Runge. Structural causal bottleneck models,

work page 2026
[3]

Counterfactual concept bottle- neck models,

[Dominiciet al., 2025 ] Gabriele Dominici, Pietro Barbiero, Francesco Giannini, Martin Gjoreski, Giuseppe Marra, and Marc Langheinrich. Counterfactual concept bottle- neck models,

work page 2025
[4]

Causally reliable concept bottleneck models,

[Feliceet al., 2026 ] Giovanni De Felice, Arianna Casanova Flores, Francesco De Santis, Silvia Santini, Johannes Schneider, Pietro Barbiero, and Alberto Termine. Causally reliable concept bottleneck models,

work page 2026
[5]

Wichmann

[Geirhoset al., 2020 ] Robert Geirhos, J ¨orn-Henrik Jacob- sen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learn- ing in deep neural networks.Nature Machine Intelligence, 2(11):665–673, November

work page 2020
[6]

Physiobank, physiotoolkit, and physionet: components of a new re- search resource for complex physiologic signals.circu- lation, 101(23):e215–e220,

[Goldbergeret al., 2000 ] Ary L Goldberger, Luis AN Ama- ral, Leon Glass, Jeffrey M Hausdorff, Plamen Ch Ivanov, Roger G Mark, Joseph E Mietus, George B Moody, Chung-Kang Peng, and H Eugene Stanley. Physiobank, physiotoolkit, and physionet: components of a new re- search resource for complex physiologic signals.circu- lation, 101(23):e215–e220,

work page 2000
[7]

Setio, Florin C

[G¨undelet al., 2021 ] Sebastian G ¨undel, Arnaud A.A. Setio, Florin C. Ghesu, Sasa Grbic, Bogdan Georgescu, An- dreas Maier, and Dorin Comaniciu. Robust classification from noisy labels: Integrating additional knowledge for chest radiography abnormality assessment.Medical Im- age Analysis, 72:102087,

work page 2021
[8]

Semi-supervised concept bottleneck mod- els,

[Huet al., 2025 ] Lijie Hu, Tianhao Huang, Huanyi Xie, Xilin Gong, Chenyang Ren, Zhengyu Hu, Lu Yu, Ping Ma, and Di Wang. Semi-supervised concept bottleneck mod- els,

work page 2025
[9]

Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison

[Irvinet al., 2019 ] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpan- skaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. InPro- ceedings of the AAAI conference on artificial intelligence, volume 33, pages 590–597,

work page 2019
[10]

Concept bottleneck generative models

[Ismailet al., 2024 ] Aya Abdelsalam Ismail, Julius Ade- bayo, Hector Corrada Bravo, Stephen Ra, and Kyunghyun Cho. Concept bottleneck generative models. InThe Twelfth International Conference on Learning Represen- tations,

work page 2024
[11]

Young, Andrew Y

[Jainet al., 2021 ] Saahil Jain, Akshay Smit, Steven QH Truong, Chanh DT Nguyen, Minh-Thanh Huynh, Mudit Jain, Victoria A. Young, Andrew Y . Ng, Matthew P. Lun- gren, and Pranav Rajpurkar. Visualchexbert: addressing the discrepancy between radiology report labels and im- age labels. InProceedings of the Conference on Health, Inference, and Learning, ACM C...

work page 2021
[12]

Concept bottleneck models,

[Kohet al., 2020 ] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models,

work page 2020
[13]

Addressing the curse of imbalanced training sets: One-sided selection

[Kub´at and Matwin, 1997] Miroslav Kub ´at and Stan Matwin. Addressing the curse of imbalanced training sets: One-sided selection. InInternational Conference on Machine Learning,

work page 1997
[14]

Promises and pitfalls of black-box concept learning models,

[Mahinpeiet al., 2021 ] Anita Mahinpei, Justin Clark, Isaac Lage, Finale Doshi-Velez, and Weiwei Pan. Promises and pitfalls of black-box concept learning models,

work page 2021
[15]

[McDermottet al., 2020 ] Matthew B. A. McDermott, Tzu Ming Harry Hsu, Wei-Hung Weng, Marzyeh Ghassemi, and Peter Szolovits. Chexpert++: Approximating the chexpert labeler for speed,differentiability, and probabilis- tic output,

work page 2020
[16]

[Miller and Masarie, 1989] R. A. Miller and Fred E. Masarie. Use of the quick medical reference (qmr) program as a tool for medical education.Methods of Information in Medicine, 28:340 – 345,

work page 1989
[17]

Miller, Harry E

[Milleret al., 1982 ] Randolph A. Miller, Harry E. Pople, and Jack D. Myers. ¡i¿internist-i¡/i¿, an experimental computer-based diagnostic consultant for general internal medicine.New England Journal of Medicine, 307(8):468– 476,

work page 1982
[18]

To- wards multi-label concept bottleneck models in medical imaging: An exploratory survey,

[Mpindaet al., 2026 ] Berthine Nyunga Mpinda, Mehran Hosseinzadeh, Valay Bundele, and Hendrik Lensch. To- wards multi-label concept bottleneck models in medical imaging: An exploratory survey,

work page 2026
[19]

Label-free concept bottleneck models

[Oikarinenet al., 2023 ] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. InInternational Conference on Learn- ing Representations,

work page 2023
[20]

Causal inference and counterfactual prediction in machine learning for ac- tionable healthcare.Nature Machine Intelligence, 2:1–7, 07

[Prosperiet al., 2020 ] Mattia Prosperi, Yi Guo, Matt Sper- rin, James Koopman, Jae Min, Xing He, Shannan Rich, Mo Wang, Iain Buchan, and Jiang Bian. Causal inference and counterfactual prediction in machine learning for ac- tionable healthcare.Nature Machine Intelligence, 2:1–7, 07

work page 2020
[21]

Chest x-ray quality: Projection

[Radiology Masterclass, 2024] Radiology Master- class. Chest x-ray quality: Projection. https: //www.radiologymasterclass.co.uk/tutorials/chest/ chest quality/chest xray quality projection,

work page 2024
[22]

[Raffertyet al., 2025 ] Amy Rafferty, Rishi Ramaesh, and Ajitha Rajan

Ac- cessed: 2026-03-21. [Raffertyet al., 2025 ] Amy Rafferty, Rishi Ramaesh, and Ajitha Rajan. Leveraging expert input for robust and ex- plainable ai-assisted lung cancer detection in chest x-rays,

work page 2026
[23]

Explainability through human-centric de- sign for xai in lung cancer detection,

[Raffertyet al., 2026 ] Amy Rafferty, Rishi Ramaesh, and Ajitha Rajan. Explainability through human-centric de- sign for xai in lung cancer detection,

work page 2026
[24]

Reinhold, Aaron Carass, and Jerry L

[Reinholdet al., 2021 ] Jacob C. Reinhold, Aaron Carass, and Jerry L. Prince. A structural causal model for mr images of multiple sclerosis. In Marleen de Bruijne, Philippe C. Cattin, St ´ephane Cotin, Nicolas Padoy, Ste- fanie Speidel, Yefeng Zheng, and Caroline Essert, editors, Medical Image Computing and Computer Assisted Inter- vention – MICCAI 2021, ...

work page 2021
[25]

[Rodmanet al., 2023 ] Adam Rodman, Thomas A Buckley, Arjun K Manrai, and Daniel J Morgan

Springer International Publishing. [Rodmanet al., 2023 ] Adam Rodman, Thomas A Buckley, Arjun K Manrai, and Daniel J Morgan. Artificial intel- ligence vs clinician performance in estimating probabili- ties of diagnoses before and after testing.JAMA Network Open, 6(12):e2347075,

work page 2023
[26]

Towards causal representation learning,

[Sch¨olkopfet al., 2021 ] Bernhard Sch¨olkopf, Francesco Lo- catello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbren- ner, Anirudh Goyal, and Yoshua Bengio. Towards causal representation learning,

work page 2021
[27]

A probabilistic reformu- lation of the quick medical reference system.Proceedings / the

[Shweet al., 1990 ] Michael Shwe, Blackford Middleton, David Heckerman, Max Henrion, Eric Horvitz, Harold Lehmann, and Gregory Cooper. A probabilistic reformu- lation of the quick medical reference system.Proceedings / the ... Annual Symposium on Computer Application [sic] in Medical Care. Symposium on Computer Applications in Medical Care, 11

work page 1990
[28]

Artificial intelli- gence revolution in healthcare: transforming diagnosis, treatment, and patient care.Asian Journal of Advances in Research, 7(1):241–263,

[Singhet al., 2024 ] Ajit Pal Singh, Rahul Saxena, Suyash Saxena, and Neelesh Kumar Maurya. Artificial intelli- gence revolution in healthcare: transforming diagnosis, treatment, and patient care.Asian Journal of Advances in Research, 7(1):241–263,

work page 2024
[29]

Defining the unde- finable: the black box problem in healthcare artificial intel- ligence.Journal of Medical Ethics, 48(10):764–768,

[Wadden, 2022] Jordan Joseph Wadden. Defining the unde- finable: the black box problem in healthcare artificial intel- ligence.Journal of Medical Ethics, 48(10):764–768,

work page 2022
[30]

Nguyen, and Tengfei Ma

[Xuet al., 2026 ] Haotian Xu, Tsui-Wei Weng, Lam M. Nguyen, and Tengfei Ma. Graph concept bottleneck mod- els,

work page 2026
[31]

Robust and interpretable medical image clas- sifiers via concept bottleneck models,

[Yanet al., 2023 ] An Yan, Yu Wang, Yiwu Zhong, Zexue He, Petros Karypis, Zihan Wang, Chengyu Dong, Amil- care Gentili, Chun-Nan Hsu, Jingbo Shang, and Julian McAuley. Robust and interpretable medical image clas- sifiers via concept bottleneck models,

work page 2023
[32]

Language in a bottle: Language model guided concept bottlenecks for interpretable image clas- sification

[Yanget al., 2023 ] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image clas- sification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187–19197,

work page 2023
[33]

Post-hoc concept bottleneck models

[Yuksekgonulet al., 2023 ] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. InThe Eleventh International Conference on Learning Representations,

work page 2023
[34]

Concept embedding models: Beyond the accuracy-explainability trade-off, 2022

[Zarlengaet al., 2022 ] Mateo Espinosa Zarlenga, Pietro Bar- biero, Gabriele Ciravegna, Giuseppe Marra, Francesco Gi- annini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, Pietro Lio, and Mateja Jamnik. Concept embedding models: Beyond the accuracy-explainability trade-off, 2022

work page 2022