Investigators trained two distinct multimodal fusion models to predict best response to venetoclax and azacitidine in newly diagnosed acute myeloid leukemia. One model used a custom ResNet-34 backbone; the other was a pathology foundation model fine-tuned on bone marrow smear images. The training cohort comprised 193 patients who received ven-aza; that is the sample size reported by the authors. In a held-out test set, the ResNet-based fusion model produced an area under the ROC curve of 0.79, while the fine-tuned foundation model achieved 0.82. External validation performance was also reported: AUROCs of 0.76 for the ResNet-based model and 0.73 for the fine-tuned foundation model. The preprint notes these values without providing further split or calibration statistics in the abstract.
The study team ran a sensitivity analysis comparing fused models with models trained on single data types. Data indicated that combining image information with clinical and genetic variables improved predictive performance compared with training on either modality alone. The authors reported this effect qualitatively and through the comparative AUROC results, which served as the primary metric described. The abstract does not list additional performance measures or per-class metrics; those details were not reported in the provided summary.
Univariable and multivariable analyses were performed to position the image-based classifier among established predictors. The investigators reported that image-based classification remained an independent predictive marker when included in multivariable models. The effect size of the image-derived predictor ranked second after IDH1/2 mutation status, and it outperformed other individual genetic alterations according to the authors. The abstract does not provide effect estimates, confidence intervals, or exact model covariates; those particulars were not reported in the source text.
The study team applied image explainability methods to localize model attention. Explainability outputs showed concentrated attention on myeloblasts in the bone marrow smears. The authors interpret this as evidence that the models identify cell-specific morphological patterns associated with response to ven-aza that are not captured by genetic profiling alone. The abstract does not detail the specific explainability algorithm used or quantify attention scores; those methodological elements were not provided in the source summary.
Training and internal testing data were deemed exempt from IRB review by the UC Denver IRB, as stated by the authors. External validation used retrospective registry data from the German Ven/Aza registry and received approval from the local Ethics Review Committee (ethical vote No.7972_BO_K_2018) at Hannover Medical School. The authors declare competing interests, with named investigators reporting consulting fees, honoraria, institutional research grants, and shareholdings; the abstract lists those declarations. The preprint states that all necessary patient consent and institutional approvals were handled and that patient identifiers were not shared outside the research group. The authors also note that all data are available upon reasonable request to the corresponding author.
This report is a preprint and has not completed peer review; the authors explicitly note that the manuscript should not guide clinical practice until formally evaluated. The abstract does not provide details on randomization, prospective validation, or model calibration metrics, and it does not list subgroup performance by age, cytogenetic risk groups, or prior therapies—information that was not reported in the supplied summary. The authors made no claims in the abstract about deployment, regulatory status, or prospective clinical testing; those next steps were not documented in the source text. Key unanswered questions remaining in the presented material include how the models perform across clinically relevant subgroups and whether image-based signals replicate in independent prospective cohorts. The manuscript makes data and code available pathways available through supplementary material and on reasonable request, but formal external prospective validation and peer-review results remain outstanding.