An AI smear screen reached 0.81 AUC for an AML mutation—but missed about four in ten cases
A four-centre study suggests bone-marrow images could help prioritise IDH1 testing. Internal validation, an enriched sample and weak sensitivity mean it cannot replace genetic confirmation.
The short version
- What happened
- Researchers retrospectively assembled 525 digitised bone-marrow smear slides from 436 adults with acute myeloid leukaemia at four centres in Germany and the United States.
- Who it affects
- A faster prompt could help a patient reach scarce molecular testing sooner, but a missed flag must not close that route. Genetic confirmation remains necessary before any IDH1-targeted treatment decision.
- What remains unproven
- Retrospective study
Keep in mind
Retrospective study
Internal cross-validation only
Mutation-enriched cohort
Moderate sensitivity
No workflow or patient outcome
No external scanner or stain validation
Confidential dataset
Industry funding and disclosed commercial interests
Source analysis · 2 linked sources. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
A possible prompt for testing, not a genetic result
A deep-learning system found a signal associated with an IDH1 mutation in routine bone-marrow smear images from adults with acute myeloid leukaemia, according to a peer-reviewed paper published on 29 October. Across three internal test folds, the mean area under the receiver-operating curve was 0.806. That score describes how well the model ranked mutated and non-mutated cases inside this dataset; it is not the percentage of patients correctly diagnosed and it is not evidence that treatment outcomes improved.
The clinically important weakness was sensitivity. The model detected 54.8% to 60.0% of mutated cases across folds, meaning it missed roughly four in ten. Specificity was higher, at 82.1% to 94.7%. For a haematology team, the plausible use is therefore an additional prompt to prioritise molecular testing when laboratory access is constrained—not a rule-out test and never a substitute for sequencing before selecting targeted treatment.[1][2]
What the four-centre experiment actually measured
The retrospective cohort contained 436 adults and 525 whole-slide images from Dresden, Munich and Ohio: 186 patients had an IDH1 mutation and 250 did not. Mutation status came from next-generation or targeted sequencing with a 2% variant-allele-frequency threshold. After slide segmentation, the pipeline retained 40 image tiles per patient, extracted features with a vision transformer and used attention-based multiple-instance learning to produce a patient-level prediction. All images from one patient stayed within the same split, reducing patient leakage.
Researchers used three-fold cross-validation rather than an independent external cohort. The study deliberately enriched mutated cases to create a nearly balanced sample, although IDH1 mutations occur in about 8% of AML. When the authors projected performance to that prevalence, positive predictive value ranged from 22.5% to 47.5%; many flagged patients would therefore be false positives. Negative predictive value approached 96%, but the modest sensitivity still makes a negative screen unsafe as a reason to withhold testing.[1][2]
The distinction that matters
The same model can look different at clinical prevalence
Three figures put the internal ranking score into a patient-facing context.
Ranking
Three-fold internal validation
Mean ROC-AUC 0.806
Detection
Count mutated cases found
Sensitivity 54.8%–60.0%
Likely clinic mix
Project to 8% mutation prevalence
Positive predictive value 22.5%–47.5%
Why the result needs a prospective test
The system has not been tested prospectively in a diagnostic workflow, on an external hospital cohort or across additional scanners and staining practices. Attention maps highlighted immature-looking regions and reduced mature granulopoiesis, but the authors did not identify a specific cellular feature that explains the mutation signal. The confidential slides are available only on reasonable request, limiting independent reproduction. The paper appropriately calls the result hypothesis-generating.
Servier partly funded the work, although the authors say the funder had no role in study design, analysis or publication. Several authors disclosed shares, employment or financial relationships involving diagnostic businesses and pharmaceutical companies, including organisations connected to the study. The assessment would change with preregistered external validation in a representative consecutive cohort, a comparison with existing clinical triage, and evidence that using the model shortens access to sequencing without missed mutations, inappropriate treatment or unequal performance across patient groups.[1][2]