Cancer-fever model travelled to a second hospital—but still needs a treatment trial
A model developed on 987 German hospital stays and tested on 7,409 US stays predicted whether fever would persist after antibiotics, with external AUROC 0.78. Retrospective opportunity counts do not show that withholding escalation is safe.
The short version
- What happened
- The retrospective cohorts contained 987 eligible hospital stays at University Medicine Essen and 7,409 at Brigham and Women's Hospital; persistent fever occurred in 54.0% and 46.5%, respectively.
- Who it affects
- For clinicians, the model could eventually add context during a difficult antibiotic review. For patients with cancer, the priority is proving that fewer escalations or scans do not delay treatment of serious infection.
- What remains unproven
- Retrospective two-centre study of hospital stays; repeat admissions by the same patient may appear more than once.
Keep in mind
Retrospective two-centre study of hospital stays; repeat admissions by the same patient may appear more than once.
Outcome required a recorded temperature 48–72 hours after antibiotics, excluding stays without that measurement.
External calibration shifted with prevalence and care patterns; local recalibration would be required.
No model-guided decision, antimicrobial reduction, diagnostic yield, ICU transfer, mortality or patient-reported outcome was tested.
Public and charitable grants supported authors; the publisher record declares no competing interests.
Source analysis · 1 linked source. Primary analysed source date: 31 October 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 3 min The evidence, methods, implications and limitations
Routine measurements carried a portable signal
A peer-reviewed two-hospital study found that routine data could estimate whether an adult with cancer would remain febrile 48–72 hours after starting intravenous antibiotics. Researchers developed models on 987 eligible hospital stays at University Medicine Essen and externally tested the locked models on 7,409 stays at Brigham and Women's Hospital. Persistent fever occurred in 533 German and 3,442 US stays.
The strongest model, TabPFN, reached an AUROC of 0.82 in nested five-fold internal validation and 0.78 in the external cohort. Its inputs included cancer type, comorbidities, laboratory values, vital signs and temperature patterns. Temperature-related features dominated. The US cohort differed substantially in cancer mix and measurement frequency, making external validation more informative than another split of the same hospital's records.[1]
Prediction is not permission to reduce treatment
Retrospectively, roughly one quarter of antibiotic escalations and chest CT scans occurred among patients the model correctly classified as unlikely to remain febrile. That identifies a possible target for a trial; it does not prove those interventions were unnecessary. A scan may reveal another problem, and antimicrobial decisions in immunocompromised patients depend on more than temperature.
External prevalence fell from 54.0% to 46.5%, and most models underpredicted risk at the US site, indicating that recalibration would be needed. The study excluded admissions without later temperature measurements and counted hospital stays rather than unique people. A prospective study should predefine a safe action, preserve clinician override and measure missed infection, intensive-care transfer, mortality, antimicrobial exposure and imaging—not only model accuracy.[1]
The distinction that matters
What the two-site validation established
The model's discrimination transferred reasonably; treatment safety did not form part of the experiment.
Germany
987 eligible stays; 54.0% persistent fever
Nested-validation AUROC 0.82
United States
7,409 eligible stays; 46.5% persistent fever
Locked-model AUROC 0.78
Next test
Model-guided reassessment
Must measure missed harm and patient outcomes
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring clinical ai: compare the finding and its evidence limits.
Can an MRI model predict cancer progression? A 33-patient holdout is promising, not treatment guidance
A Chinese single-centre model separated progression risk after chemoradiotherapy in 113 patients, but the holdout contained only 33 people and no external hospital. The paper's feature count is also internally inconsistent.
Read the next reportExplore the full source and evidence record
What this means for people
- For clinicians, the model could eventually add context during a difficult antibiotic review. For patients with cancer, the priority is proving that fewer escalations or scans do not delay treatment of serious infection.
Global context
Antimicrobial stewardship in cancer care must balance resistance and unnecessary testing against the high cost of missing infection. Hospitals differ in monitoring, pathogens, cancer populations and escalation protocols, so broader prospective validation is essential.
What the evidence does not yet show
- Retrospective two-centre study of hospital stays; repeat admissions by the same patient may appear more than once.
- Outcome required a recorded temperature 48–72 hours after antibiotics, excluding stays without that measurement.
- External calibration shifted with prevalence and care patterns; local recalibration would be required.
- No model-guided decision, antimicrobial reduction, diagnostic yield, ICU transfer, mortality or patient-reported outcome was tested.
- Public and charitable grants supported authors; the publisher record declares no competing interests.
What to watch next
- Prospective impact trial
- Safety outcomes after reduced escalation
- Calibration across hospitals and countries
- Unique-patient and repeat-admission analysis
- Comparison with simple temperature rules
- Clinician uptake and override patterns
Living evidence record
Impact record IAI-04VFTP0
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
31 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what npj Digital Medicine published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 31 October 2026
Research question, topics and regions
Whether routine hospital data can predict persistent fever in adults with cancer 48–72 hours after intravenous antibiotics
Topics: Clinical AI · Cancer care · Antibiotics · Fever · External validation
Regions: Europe · Germany · North America · United States
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.