Skip to main content
Tessa NewsLink
Paste a health news link, or browse
Latest checks

Source study found

Story checked

New AI model detects hidden diseases in routine CT scans (opens in a new tab)

news-medical.net · 2026-10-20

Short answerEvidenceSource

Short answer

Mostly not supported

Mostly not supported.

The claims we could check match the study, but some claims were not covered by the evidence reviewed.

  • 1 supported
  • 4 not covered

Checked against the study summary. The full text wasn't available, so some details couldn't be settled either way.

Share this check

Follow the evidence trail
1
2

NewsLink checks it

Mostly not supported

The one claim we could check holds up. One of five claims matches the study. This overall rating is based only on the claims we could check. Four claims the study doesn't address.

  • 1 supported
  • 4 not covered
Open claim evidence
3
Then inspect each claim

Evidence layer

Claim by claim

Each claim gets a verdict. Expand it to see the evidence directly below.

5 claims in this story

Showing all 5 claimsChoose a verdict to focus the list.

Claim 1 of 5Not coveredThe model was built using data from the Penn Medicine BioBank, including information from more than 350,000 people and roughly 400,000 CT scans from about 50,000 patients.View evidenceHide evidence

As statedmore than 350,000 people; roughly 400,000 CT scans; approximately 50,000 patients

Why this verdict

The abstract supports that Percival was trained on more than 400,000 CT-report pairs from the Penn Medicine BioBank. However, the additional figures in the story — more than 350,000 people and roughly 50,000 CT patients — are not present in the supplied abstract-level profile, so the full numerical claim cannot be verified at this evidence depth.

1

Study evidence

Developed and pretrained Percival, a CT-native vision-language model trained on >400,000 CT–report pairs from the Penn Medicine BioBank using a dual-encoder symmetric contrastive framework.

“We introduce Percival, a CT-native VLM trained on more than 400,000 CT-report pairs from the Penn Medicine BioBank using a dual-encoder symmetric contrastive framework.”
Claim 2 of 5Not coveredPercival was found to uncover traces of several chronic conditions, including atrial fibrillation and chronic kidney disease, in scans.View evidenceHide evidence

Why this verdict

The abstract-level profile supports broad phenome-wide EHR associations and disease-classification evaluation, but it does not identify atrial fibrillation, chronic kidney disease, or specific chronic-condition traces in scans. The specificity of these examples is therefore not verifiable from the supplied abstract profile.

1

Study evidence

Percival's latent space aligns with demographic, physiological, and laboratory variation across a held-out cohort of over 20,000 participants.

“Across over 20,000 held-out participants, Percival’s latent space aligns with demographic, physiological, and laboratory variation”
2

Study evidence

Percival's latent space aligns with demographic, physiological, and laboratory variation across a held-out set of >20,000 participants.

“...and supports phenome-wide associations across the electronic health record.”
Claim 3 of 5Not coveredCompared with a similar AI model called TotalSegmentator, Percival was found to flag pancreatic cancer, cirrhosis, and bacterial pneumonia more reliably.View evidenceHide evidence

Why this verdict

The abstract supports a benchmark against alternative paradigms, including multi-organ segmentation pretraining, and says vision-language pretraining captured clinical information for disease classification and risk modeling. But the profile does not name TotalSegmentator, does not list pancreatic cancer, cirrhosis, or bacterial pneumonia, and provides no comparative reliability metrics. The story’s specific comparator and disease-specific performance claim are not verifiable at abstract depth.

1

Study evidence

Vision-language pretraining with Percival captures clinical information that is not fully accessible to vision-only contrastive or multi-organ segmentation pretraining, improving representation alignment with clinical, physiological, and prognostic variation for disease classification and longitudinal risk modeling.

“We evaluate Percival against alternative foundation-model paradigms, including vision-only contrastive and multi-organ segmentation”
2

Study evidence

Vision-language pretraining (Percival) captures clinical information not fully accessible to vision-only pretraining alternatives on downstream disease classification and longitudinal risk modeling.

“...demonstrating that vision-language pretraining captures clinical information not fully accessible to vision-only alternatives across disease classification...”
Claim 4 of 5Not coveredThe researchers say the model is still being tested preclinically and hope to test it early next year with an eye toward bringing it into the clinic for regular use.View evidenceHide evidence

As statedearly next year

Why this verdict

The abstract-level profile describes secondary-data model development, held-out evaluation, disease classification, and longitudinal risk modeling, but it does not state that the model is being tested preclinically, give a planned early-next-year clinical-testing timeline, or discuss a path to regular clinical use. The non-deployment implication is consistent with the abstract’s research-evaluation framing, but the specific timeline and clinical-use plan are not verifiable at this depth.

1

Study evidence

Vision-language pretraining (Percival) captures clinical information not fully accessible to vision-only pretraining alternatives on downstream disease classification and longitudinal risk modeling.

“...demonstrating that vision-language pretraining captures clinical information not fully accessible to vision-only alternatives across disease classification...”
2

Study evidence

Percival’s latent space aligns with demographic, physiological, and laboratory variation and supports phenome-wide associations across the electronic health record.

“...across disease classification and longitudinal risk modeling.”
Claim 5 of 5SupportedA new artificial intelligence model developed at Penn Medicine, called Percival, can uncover nearly hidden patterns in disease and potential future health risks from routine CT scans.View evidenceHide evidence

Why this verdict

At abstract depth, the paper supports the broad, hedged claim that Percival can reveal CT-derived latent structure associated with clinical, physiological, laboratory, phenome-wide EHR, and prognostic/longitudinal-risk variation. The headline framing is somewhat journalistic, but because the story frames this as potential/associational rather than proven clinical diagnosis or causation, it is broadly supported. The abstract does not specifically verify the phrase about findings being beyond the original scan indication.

1

Study evidence

Percival's latent space aligns with demographic, physiological, and laboratory variation across a held-out cohort of over 20,000 participants.

“Across over 20,000 held-out participants, Percival’s latent space aligns with demographic, physiological, and laboratory variation”
2

Study evidence

Percival's latent space aligns with demographic, physiological, and laboratory variation across a held-out set of >20,000 participants.

“...and supports phenome-wide associations across the electronic health record.”
Then look for missing context

Context layer

What the story left out

Important study details the story did not include.

  • The paper evaluates Percival in over 20,000 held-out participants and reports that its latent space aligns with demographic, physiological, and laboratory variation.

    The story gives a broad account of hidden disease patterns and future risks but does not report the held-out evaluation cohort or the demographic, physiological, and laboratory alignment findings.

    From secondary-data observational embedding analysis

  • Percival embeddings support phenome-wide associations across the electronic health record.

    The story translates the work into examples of disease detection/risk, but it does not explain the EHR-wide phenome-wide association analysis that the abstract identifies as a key result.

    From PheWAS-style EHR association

  • The abstract does not provide quantitative performance metrics, effect sizes, confidence intervals, p-values, or detailed statistical comparisons for the disease-classification and longitudinal-risk results.

    This limitation is material to claims that the model performed 'more reliably.' The story mentions preclinical status, but it does not acknowledge that the abstract-level evidence lacks numerical performance metrics or statistical uncertainty.

    From comparative model benchmarking; secondary_data; secondary_data

3 things the story did carry across
  • Percival is a CT-native vision-language foundation model trained on more than 400,000 CT-report pairs from the Penn Medicine BioBank using a dual-encoder symmetric contrastive framework.
  • The paper benchmarks Percival against alternative foundation-model paradigms, including vision-only contrastive and multi-organ segmentation pretraining, for disease classification and longitudinal risk modeling.
  • The work is presented in the profile as secondary-data model development and evaluation rather than completed clinical deployment.
Then read the study layer

Study layer

Study at a glance

Scan the study first. Expand only the parts you want to inspect.

Pieces of work

6

Evidence read

study summary

Lead result

secondary data

1Lead resultsecondary dataDevelop and train Percival, a CT-native vision-language foundation model on >400,000 CT–report pairs using a dual-encoder symmetric contrastive framework.vision-language dual-encoder pretrainingExpand

In plain English

Introduces Percival, a CT-native vision-language foundation model pretrained on more than 400,000 paired CT images and radiology reports from the Penn Medicine BioBank using a dual-encoder symmetric contrastive learning framework.

Key findings

  • Developed and pretrained Percival, a CT-native vision-language model trained on >400,000 CT–report pairs from the Penn Medicine BioBank using a dual-encoder symmetric contrastive framework.
“We introduce Percival, a CT-native VLM trained on more than 400,000 CT-report pairs from the Penn Medicine BioBank using a dual-encoder symmetric contrastive framework.”
What this piece can’t prove

2 further details could not be confirmed from the summary.

2secondary dataShow that Percival’s learned latent space aligns with real-world population variation (demographic, physiological, laboratory) and enables phenome-wide associations using EHR data across >20,000 held-out participants.secondary-data observational embedding analysisExpand

In plain English

The authors evaluate Percival, a CT-native vision-language foundation model trained on >400,000 CT–report pairs, by extracting embeddings for over 20,000 held-out participants and testing whether the model's latent space aligns with real-world population variation. Using linked EHR attributes, they report that Percival embeddings align with demographic, physiological, and laboratory variation and enable phenome-wide associations across the electronic health record.

Key findings

  • Percival's latent space aligns with demographic, physiological, and laboratory variation across a held-out cohort of over 20,000 participants.
  • Percival embeddings enable phenome-wide associations across the electronic health record.
“Across over 20,000 held-out participants, Percival’s latent space aligns with demographic, physiological, and laboratory variation”
What this piece can’t prove
  • Abstract does not report effect sizes, confidence intervals, p-values, or details on confounding adjustment.

2 further details could not be confirmed from the summary.

3secondary dataShow that Percival’s learned latent space aligns with real-world population variation (demographic, physiological, laboratory) and enables phenome-wide associations using EHR data across >20,000 held-out participants.PheWAS-style EHR associationExpand

In plain English

Using Percival, a CT-native vision-language model trained on >400,000 CT-report pairs, the authors report that the model's latent space aligns with demographic, physiological, and laboratory variation and that these CT-derived embeddings support phenome-wide association analyses across the electronic health record in a held-out set of over 20,000 participants.

Key findings

  • Percival's latent space aligns with demographic, physiological, and laboratory variation across a held-out set of >20,000 participants.
  • CT-derived embeddings from Percival support phenome-wide associations across the electronic health record.
“...and supports phenome-wide associations across the electronic health record.”
What this piece can’t prove

3 further details could not be confirmed from the summary.

4secondary dataBenchmark Percival against alternative foundation-model paradigms (vision-only contrastive; multi-organ segmentation pretraining) on downstream disease classification and longitudinal risk modeling, showing VLM pretraining captures clinical information not accessible to vision-only alternatives.comparative model benchmarkingExpand

In plain English

The paper reports a comparative evaluation benchmarking the CT-native vision-language model Percival against two alternative foundation-model paradigms — a vision-only contrastive pretraining baseline and a multi-organ segmentation pretraining baseline — on downstream disease classification and longitudinal risk modeling tasks. The authors state that vision-language pretraining with Percival captures clinical information not fully accessible to the vision-only alternatives.

Key findings

  • Vision-language pretraining with Percival captures clinical information that is not fully accessible to vision-only contrastive or multi-organ segmentation pretraining, improving representation alignment with clinical, physiological, and prognostic variation for disease classification and longitudinal risk modeling.
“We evaluate Percival against alternative foundation-model paradigms, including vision-only contrastive and multi-organ segmentation”
What this piece can’t prove

3 further details could not be confirmed from the summary.

5secondary dataBenchmark Percival against alternative foundation-model paradigms (vision-only contrastive; multi-organ segmentation pretraining) on downstream disease classification and longitudinal risk modeling, showing VLM pretraining captures clinical information not accessible to vision-only alternatives.secondary dataExpand

In plain English

Benchmarked Percival (a CT-native vision-language foundation model) against vision-only contrastive and multi-organ segmentation pretrained alternatives on downstream disease classification and longitudinal risk modeling using representations learned from CT. Percival was pretrained on >400,000 CT-report pairs with a dual-encoder symmetric contrastive framework and evaluated on >20,000 held-out participants; the authors report that Percival’s latent space aligns with demographic, physiological, and laboratory variation and that vision-language pretraining captures clinical information not fully accessible to vision-only alternatives for disease classification and longitudinal risk modeling.

Key findings

  • Vision-language pretraining (Percival) captures clinical information not fully accessible to vision-only pretraining alternatives on downstream disease classification and longitudinal risk modeling.
  • Percival’s latent space aligns with demographic, physiological, and laboratory variation and supports phenome-wide associations across the electronic health record.
“...demonstrating that vision-language pretraining captures clinical information not fully accessible to vision-only alternatives across disease classification...”
What this piece can’t prove

5 further details could not be confirmed from the summary.

6secondary dataBenchmark Percival against alternative foundation-model paradigms (vision-only contrastive; multi-organ segmentation pretraining) on downstream disease classification and longitudinal risk modeling, showing VLM pretraining captures clinical information not accessible to vision-only alternatives.secondary dataExpand

In plain English

The abstract reports a benchmark comparing Percival, a CT-native vision-language foundation model (VLM), to alternative foundation-model paradigms (vision-only contrastive and multi-organ segmentation pretraining) on downstream disease classification and longitudinal risk modeling. Percival was trained with a dual-encoder symmetric contrastive framework on >400,000 CT–report pairs and evaluated across >20,000 held-out participants. The authors state that Percival’s latent space aligns with demographic, physiological, and laboratory variation, supports phenome-wide EHR associations, and that vision-language pretraining captures clinical information relevant to disease classification and prognostic/longitudinal risk prediction that is not fully accessible to vision-only alternatives.

Key findings

  • Percival’s latent space aligns with demographic, physiological, and laboratory variation and supports phenome-wide associations across the electronic health record.
  • Vision-language pretraining (Percival) captures clinical information relevant to disease classification and longitudinal risk modeling that is not fully accessible to vision-only contrastive or multi-organ segmentation pretraining.
“...across disease classification and longitudinal risk modeling.”
What this piece can’t prove
  • Abstract does not specify the exact longitudinal/prognostic modeling approach (time-to-event vs. other temporal prediction), outcome definitions, follow-up duration, censoring handling, or covariate adjustment.
  • Unclear whether evaluations included external validation, subgroup analyses, or calibration assessments.

2 further details could not be confirmed from the summary.

Finally, the search trail

Method layer

NewsLink found the paper. Tessa takes you deeper.

NewsLink checks the story. Tessa is where you inspect the paper, authors, evidence, and research context.

Open the paper in Tessa

Generalizable CT vision-language modeling for population health and disease risk

npj Digital Medicine · 2026

Source record (opens in a new tab)

Why this one

Near certain

NewsLink found the paper. Tessa is where you inspect it deeply.

Papers considered

The selected paper, plus nearby candidates.

Crossref, Europe PMC, PubMed · 41 candidate papers

Selected

Generalizable CT vision-language modeling for population health and disease risk

Npj Digital Medicine · 2026 · Crossref

Source (opens in a new tab)

And 35 more candidates considered.