pith. sign in

arxiv: 2605.30275 · v1 · pith:4UDUHTQYnew · submitted 2026-05-28 · 💻 cs.LG · q-bio.QM

Digitally enriching a screening population for pancreatic cancer using routine blood-based measures and clinical histories

Chris Varghese , Leo Y. Li-Han , Richa Bisht , Ellen Larson , Frank Lee , Ryan M. Carr , Tanios S. Bekaii-Saab , Shounak Majumder
show 5 more authors
John D. Halamka Mark Truty Ajit H. Goenka Hojjat Salehinejad Cornelius A. Thiels
This is my paper

Pith reviewed 2026-06-29 08:32 UTC · model grok-4.3

classification 💻 cs.LG q-bio.QM
keywords pancreatic cancerrisk predictiontransformerscreeninglongitudinal datablood testsclinical historiesmachine learning
0
0 comments X

The pith

A Transformer model trained on sequences of routine blood tests and diagnosis codes predicts pancreatic cancer risk up to three years before diagnosis.

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

The paper establishes that longitudinal records of coded diagnoses and blood test values contain extractable signals that a neural network can use to forecast pancreatic cancer development years in advance. If correct, this would allow digital risk stratification to focus screening efforts on smaller high-risk groups, addressing the current barrier that broad population screening is not feasible given the disease's low incidence. External validation across sites shows solid performance at one-, two-, and three-year lead times, with risk outputs that remain calibrated and adjustable for different population settings. The result points toward a practical way to increase early detections that lead to curative treatment.

Core claim

The authors trained a custom Transformer neural network with multi-head attention on longitudinal sequences from 6,017 pancreatic cancer cases and 177,081 controls with a median 12 years of prior data. On leave-one-site-out external validation the model achieved mean AUCs of 0.837 (95% CI 0.827-0.848) at one year, 0.797 (0.782-0.813) at two years, and 0.760 (0.745-0.776) at three years prior to diagnosis. Estimated risks were well-calibrated (slope 1.08, intercept -0.077, Brier score 0.025), and a Bayesian prevalence update supports transportability across settings. At a >3.3% one-year risk threshold the diagnostic odds ratio reached 18.2, providing the basis for population-level digital enr

What carries the argument

The custom Transformer-based neural network with multi-head attention that ingests longitudinal sequences of coded diagnoses and blood test values to output calibrated multi-year pancreatic cancer risk scores.

If this is right

  • A one-year risk threshold above 3.3% delivers a diagnostic odds ratio of 18.2 for identifying cases.
  • Bayesian adjustment of the model's outputs with local prevalence makes risk estimates usable across different healthcare settings.
  • Risk stratification enables targeted screening that could increase the proportion of patients reaching curative-intent treatment.
  • Leave-one-site-out validation indicates the learned patterns generalize beyond individual institutions.

Where Pith is reading between the lines

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

  • Electronic health record systems could run the model in the background to surface alerts during routine visits without new data collection.
  • The same sequence-based approach could be tested for early detection of other low-incidence cancers where broad screening is impractical.
  • A next step would be a pragmatic trial measuring whether model-flagged patients actually receive and benefit from earlier diagnostic workups.
  • Adding genetic or family-history variables to the input sequences might further improve discrimination if those data become routinely available.

Load-bearing premise

The sequences of routine diagnoses and blood tests contain consistent latent patterns of future pancreatic cancer that the model can learn and that hold in new sites and time periods.

What would settle it

A prospective study in an independent population where the model's predicted risks show AUC below 0.70 or marked miscalibration against observed pancreatic cancer incidence.

Figures

Figures reproduced from arXiv: 2605.30275 by Ajit H. Goenka, Chris Varghese, Cornelius A. Thiels, Ellen Larson, Frank Lee, Hojjat Salehinejad, John D. Halamka, Leo Y. Li-Han, Mark Truty, Richa Bisht, Ryan M. Carr, Shounak Majumder, Tanios S. Bekaii-Saab.

Figure 1
Figure 1. Figure 1: (A) Visualization of time-bucketed data encoding matrices of longitudinal ICD diagnosis codes and laboratory tests for two subjects from case and control groups. (B) Architecture of the Transformer-based pancreatic cancer prediction model. (C) The model’s output includes a pancreatic cancer score and a binary prediction result. (D) Simulation of pancreatic cancer risk continuous monitoring in real-world ap… view at source ↗
read the original abstract

Earlier detection of pancreatic cancer is key to enabling wider access to curative treatment and reducing cancer deaths; however, screening is presently not viable. Latent indicators of pathology are evident in an individual's disease and blood test trajectories and may predict the development of pancreatic cancer. Longitudinal sequences of coded diagnoses and blood test values accrued by patients throughout their clinical interactions were used to train a custom Transformer-based neural network with a multi-head attention mechanism to predict risk of pancreatic cancer with a multi-year lead time and risk-stratify populations for targeted screening. The cohort comprised 6,017 adults with pancreatic cancer and 177,081 controls (overall median age 75, 45% female) with median 12 years (interquartile range 6.9-16.2) of medical history prior to pancreatic cancer diagnosis. External validation via leave-one-site-out, out-of-sample testing predicting pancreatic cancer 1-, 2-, and 3-years prior to diagnosis demonstrated mean area under the receiver operating characteristic of 0.837 (95% confidence interval 0.827-0.848), 0.797 (95% confidence interval 0.782-0.813), and 0.760 (95% confidence interval 0.745-0.776), respectively. Estimated pancreatic cancer risks were well-calibrated (calibration plot slope 1.08, intercept of -0.077; Brier score 0.025), and a Bayesian population pancreatic cancer prevalence update allows estimated cancer risk outputs to be transportable across settings. At testing, a screening threshold of >3.3% risk of pancreatic cancer in 1-year offered a diagnostic odds ratio of 18.2. Our work therefore lays the foundation for a first population-level digital enrichment tool to widen access to curative-intent management of pancreatic cancer.

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.

Referee Report

2 major / 2 minor

Summary. The paper claims that a custom Transformer model with multi-head attention, trained on longitudinal sequences of coded diagnoses and blood-test values (median 12 years of history) from 6,017 pancreatic-cancer cases and 177,081 controls, can predict pancreatic cancer 1–3 years prior to diagnosis. Leave-one-site-out external validation yields mean AUROCs of 0.837 (95% CI 0.827–0.848), 0.797 (0.782–0.813) and 0.760 (0.745–0.776) respectively, with reported calibration slope 1.08, intercept –0.077 and Brier score 0.025; a Bayesian prevalence update is proposed to make risk estimates transportable, and a >3.3% 1-year risk threshold gives a diagnostic odds ratio of 18.2.

Significance. If the performance and calibration hold under fuller methodological disclosure, the work supplies the first large-scale, externally validated demonstration that routine clinical trajectories can digitally enrich a screening population for pancreatic cancer, a disease for which population screening is currently infeasible. The leave-one-site-out design and explicit calibration metrics are strengths that directly address transportability.

major comments (2)
  1. [Methods] Methods (model description): the manuscript states only that a “custom Transformer-based neural network with a multi-head attention mechanism” was used, but supplies no architecture details (number of layers, heads, embedding dimension, positional encoding), training procedure (optimizer, learning-rate schedule, loss, regularization), feature encoding of diagnosis codes and lab values, or missing-data handling. These omissions are load-bearing for the central performance claim because the reported AUROCs and calibration cannot be interpreted or reproduced without them.
  2. [Results] Results (validation): while leave-one-site-out AUROCs and calibration metrics are presented, the paper does not report site-specific performance variation, the number of sites, or any analysis of whether performance degrades with shorter observation windows or differing coding practices across sites. This information is needed to assess whether the generalization claim is robust.
minor comments (2)
  1. [Abstract / Cohort description] Abstract and text: the cohort description gives overall median age and sex but does not break these down by case versus control status or by site; adding these tables would improve transparency.
  2. [Methods] The Bayesian prevalence-update step is mentioned but its implementation (prior specification, updating formula) is not shown; a short methods paragraph or appendix would clarify how transportability is achieved.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the opportunity to respond to the referee's comments. We appreciate the positive assessment of the work's significance and will address the methodological and validation concerns to enhance the manuscript's clarity and reproducibility.

read point-by-point responses

  1. Referee: [Methods] Methods (model description): the manuscript states only that a “custom Transformer-based neural network with a multi-head attention mechanism” was used, but supplies no architecture details (number of layers, heads, embedding dimension, positional encoding), training procedure (optimizer, learning-rate schedule, loss, regularization), feature encoding of diagnosis codes and lab values, or missing-data handling. These omissions are load-bearing for the central performance claim because the reported AUROCs and calibration cannot be interpreted or reproduced without them.

    Authors: We agree with the referee that additional details on the model architecture and training are necessary for reproducibility and proper interpretation of the results. In the revised version of the manuscript, we will expand the Methods section to include comprehensive descriptions of the Transformer architecture (including number of layers, heads, embedding dimension, and positional encoding), the training procedure (optimizer, learning-rate schedule, loss function, and regularization), feature encoding methods for diagnosis codes and blood test values, and the approach to handling missing data. revision: yes

  2. Referee: [Results] Results (validation): while leave-one-site-out AUROCs and calibration metrics are presented, the paper does not report site-specific performance variation, the number of sites, or any analysis of whether performance degrades with shorter observation windows or differing coding practices across sites. This information is needed to assess whether the generalization claim is robust.

    Authors: We acknowledge that providing site-specific performance metrics, the number of participating sites, and analyses regarding shorter observation periods or variations in coding practices would better support the robustness of the generalization claims. We will revise the Results section to include a table or figure showing site-specific AUROCs, state the number of sites used in the leave-one-site-out validation, and add sensitivity analyses examining performance with truncated histories and across different coding practices where feasible. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper's central derivation consists of training a Transformer model on longitudinal diagnosis and lab sequences from a multi-site cohort, followed by leave-one-site-out external validation that produces AUCs of 0.837/0.797/0.760 at 1/2/3-year horizons. This procedure fits parameters on training sites and evaluates on completely held-out sites, so the reported performance metrics are not equivalent to the inputs by construction, nor do they rely on self-citation chains, imported uniqueness theorems, or relabeling of fitted quantities as predictions. The validation directly tests generalization of any extracted patterns and remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Review performed on abstract only; full methods, architecture details, and data-processing steps are unavailable, preventing exhaustive enumeration of free parameters or background assumptions.

free parameters (1)
  • Transformer weights and attention parameters
    All neural-network parameters are fitted to the training sequences; exact count and regularization choices not stated.
axioms (1)
  • domain assumption Coded diagnoses and blood-test trajectories contain extractable latent signals of future pancreatic cancer
    Abstract premise that these routine data suffice for multi-year prediction.

pith-pipeline@v0.9.1-grok · 5939 in / 1479 out tokens · 37613 ms · 2026-06-29T08:32:11.671661+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

16 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    2 Stoop TF, Javed AA, Oba A, et al

    JAMA Netw Open 2021; 4: e214708. 2 Stoop TF, Javed AA, Oba A, et al. Pancreatic cancer. Lancet 2025; 405: 1182–202. 3 Siegel RL, Miller KD, Jemal A. Cancer statistics,

    2021

  2. [2]

    4 Neal RD, Johnson P, Clarke CA, et al

    CA Cancer J Clin 2018; 68: 7–30. 4 Neal RD, Johnson P, Clarke CA, et al. Cell-free DNA-based multi-cancer early detection test in an asymptomatic screening population (NHS-Galleri): Design of a pragmatic, prospective randomised controlled trial. Cancers (Basel) 2022; 14:

    2018

  3. [3]

    Long-term outcomes and risk of pancreatic cancer in intraductal papillary mucinous neoplasms

    5 de la Fuente J, Chatterjee A, Lui J, et al. Long-term outcomes and risk of pancreatic cancer in intraductal papillary mucinous neoplasms. JAMA Netw Open 2023; 6: e2337799. 6 Yachida S, Jones S, Bozic I, et al. Distant metastasis occurs late during the genetic evolution of pancreatic cancer. Nature 2010; 467: 1114–7. 7 Notta F, Chan-Seng-Yue M, Lemire M,...

    2023

  4. [4]

    Screening for pancreatic cancer: Updated evidence report and systematic review for the US Preventive Services Task Force

    10 Henrikson NB, Aiello Bowles EJ, Blasi PR, et al. Screening for pancreatic cancer: Updated evidence report and systematic review for the US Preventive Services Task Force. JAMA 2019; 322: 445–54. 11 Chari ST, Feng Z, Wu B, et al. Heuriskance: a novel paradigm for systematic earlier detection of sporadic pancreatic cancer. J Natl Cancer Inst 2025; : djaf...

    2019

  5. [5]

    27 18 Mukherjee S, Patra A, Khasawneh H, et al

    DOI:10.1053/j.gastro.2023.08.034. 27 18 Mukherjee S, Patra A, Khasawneh H, et al. Radiomics-based machine-learning models can detect pancreatic cancer on prediagnostic computed tomography scans at a substantial lead time before clinical diagnosis. Gastroenterology 2022; 163: 1435-1446.e3. 19 Foy BH, Petherbridge R, Roth MT, et al. Haematological setpoints...

    work page doi:10.1053/j.gastro.2023.08.034 2023

  6. [6]

    A readers’ guide to the interpretation of diagnostic test properties: clinical example of sepsis

    24 Fischer JE, Bachmann LM, Jaeschke R. A readers’ guide to the interpretation of diagnostic test properties: clinical example of sepsis. Intensive Care Med 2003; 29: 1043–51. 25 Elkan C. The foundations of cost-sensitive learning. Int Jt Conf Artif Intell 2001; : 973–8. 26 Mukherjee S, Antony A, Patnam NG, et al. Next-generation AI for visually occult pa...

    2003

  7. [7]

    27 Gordon-Dseagu VL, Devesa SS, Goggins M, Stolzenberg-Solomon R

    DOI:10.1136/gutjnl-2025-337266. 27 Gordon-Dseagu VL, Devesa SS, Goggins M, Stolzenberg-Solomon R. Pancreatic cancer incidence trends: evidence from the Surveillance, Epidemiology and End Results (SEER) population-based data. Int J Epidemiol 2018; 47: 427–39. 28 Siegel RL, Kratzer TB, Giaquinto AN, Sung H, Jemal A. Cancer statistics,

    work page doi:10.1136/gutjnl-2025-337266 2025

  8. [8]

    29 Samaan JS, Abboud Y, Oh J, et al

    CA Cancer J Clin 2025; 75: 10–45. 29 Samaan JS, Abboud Y, Oh J, et al. Pancreatic cancer incidence trends by race, ethnicity, age and sex in the United States: A population-based study, 2000-2018. Cancers (Basel) 2023; 15:

    2025

  9. [9]

    30 Mukherjee S, Korfiatis P, Patnam NG, et al. Assessing the robustness of a machine-learning model for early detection of pancreatic adenocarcinoma (PDA): evaluating resilience to variations in image acquisition and radiomics workflow using image perturbation methods. Abdom Radiol (NY) 2024; 49: 964–74. 31 Hewitt MJ, McPhail MJW, Possamai L, Dhar A, Vlav...

    2024

  10. [10]

    Validation of the Enriching New-Onset Diabetes for Pancreatic Cancer model in a diverse and integrated healthcare setting

    33 Chen W, Butler RK, Lustigova E, Chari ST, Wu BU. Validation of the Enriching New-Onset Diabetes for Pancreatic Cancer model in a diverse and integrated healthcare setting. Dig Dis Sci 2021; 66: 78–87. 28 34 National Lung Screening Trial Research Team, Aberle DR, Adams AM, et al. Reduced lung-cancer mortality with low-dose computed tomographic screening...

    2021

  11. [11]

    39 Gibson AD, White NM, Collins GS, Barnett AG

    DOI:10.1038/d41586-026-00697-4. 39 Gibson AD, White NM, Collins GS, Barnett AG. Evidence of unreliable data and poor data provenance in clinical prediction model research and clinical practice. medRxiv. 2026; published online Feb

    work page doi:10.1038/d41586-026-00697-4 2026

  12. [12]

    40 Aggarwal G, Kamada P, Chari ST

    DOI:10.64898/2026.02.24.26347028. 40 Aggarwal G, Kamada P, Chari ST. Prevalence of diabetes mellitus in pancreatic cancer compared to common cancers. Pancreas 2013; 42: 198–201. 41 Stott M, Stefanova I, Oldfield L, et al. Prevalence of new-onset diabetes in patients undergoing pancreatic surgery and the association of glucose dysregulation with complicati...

    work page doi:10.64898/2026.02.24.26347028 2026

  13. [13]

    48 Bahdanau D, Cho K, Bengio Y

    https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa- Abstract.html. 48 Bahdanau D, Cho K, Bengio Y. Neural machine translation by jointly learning to align and translate. arXiv [cs.CL]. 2014; published online Sept

    2017

  14. [14]

    Neural Machine Translation by Jointly Learning to Align and Translate

    http://arxiv.org/abs/1409.0473. 49 Lin T-Y, Goyal P, Girshick R, He K, Dollar P. Focal Loss for dense object detection. IEEE Trans Pattern Anal Mach Intell 2020; 42: 318–27. 29 50 Roesler MW, Wells C, Schamberg G, et al. Class imbalance correction in artificial intelligence models leads to miscalibrated clinical predictions: a real-world evaluation. medRx...

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  15. [15]

    52 Saerens M, Latinne P, Decaestecker C

    DOI:10.1002/sim.10320. 52 Saerens M, Latinne P, Decaestecker C. Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure. Neural Comput 2002; 14: 21–41. 53 Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Advances in neural information processing systems 2017

    work page doi:10.1002/sim.10320 2002

  16. [16]

    54 Naeini MP, Cooper GF, Hauskrecht M

    https://proceedings.neurips.cc/paper/2017/hash/8a20a8621978632d76c43dfd28b67767- Abstract.html. 54 Naeini MP, Cooper GF, Hauskrecht M. Obtaining well calibrated probabilities using Bayesian Binning. Proc Conf AAAI Artif Intell 2015; 2015: 2901–7. 55 Niculescu-Mizil A, Caruana R. Predicting good probabilities with supervised learning. In: Proceedings of th...

    2017