Deep Learning Predicts Response to Venetoclax and Azacitidine in Patients with Newly Diagnosed Acute Myeloid Leukemia Solely Based on Bone Marrow Smear Image Data

Venetoclax (ven) combined with azacitidine (aza) is a therapeutic standard for newly diagnosed patients with acute myeloid leukemia (AML) who are unfit to receive intensive chemotherapy. Recent studies also suggest that patients eligible for intensive chemotherapy may benefit from ven/aza, reporting comparable rates of complete remission (CR), fewer adverse events, and shorter durations of hospitalization. However, while most patients respond to this regimen, a substantial proportion are refractory to ven/aza treatment; these patients have very poor outcomes. Investigations into patient features predicting response to ven/aza have suggested molecular alterations in NPM1 and IDH1/2 to be associated with higher response rates, whereas NRAS, KRAS, FLT3–ITD, and TP53 mutations were associated with poorer outcomes. However, significant time can be required for mutational data to become available and other factors such as co-mutations and differentiation stage of the leukemic cells should be considered as they may influence the interpretation of these data.

We hypothesized that digital biomarkers in baseline, pre-treatment bone marrow aspirate smear images may predict treatment response to this therapy. Whole-slide-images (WSI) of bone marrow smears from 170 newly diagnosed AML patients who received frontline treatment with ven/aza were digitized at 63× magnification. Median age was 71 years (interquartile range [IQR]: 65-76) and 41.8% of patients were female. Median bone marrow blast count at initial diagnosis was 58% (IQR: 33-80%). According to ELN2022, 14.7% of patients had favorable risk, 16.5% had intermediate risk, and 68.8% had adverse risk. CR/CRi as best response was achieved in 80.6% with a median treatment duration until best response of 35 days (IQR: 30-49). Any response less than CRi was defined as a non-response. Eight regions-of-interest (ROI) were captured per WSI and used as input for a deep learning pipeline. Pre-processing included balancing the dataset to address the imbalance between responders (n=137) and non-responders (n=33). To accommodate for the small sample size, standard techniques for image augmentation like geometric transformations were used. A wide-resnet-50-2 model was trained using the High-Performance Computing Cluster at TU Dresden with eight Nvidia A100 GPUs, including a hyperparameter search over 100 runs to optimize model configurations.

Model performance was measured using fivefold cross-validation (train-test-split 80:20), ensuring that each patient appeared in the test set exactly once to account for effect size variation due to the small sample size. Across the five folds, our model achieved an average area-under-the-receiver-operating-characteristic (AUROC) of 0.84 (range: 0.77-0.86) in distinguishing between responders and non-responders. Confusion matrices showed accuracies ranging from 0.84 to 0.90, F1-scores from 0.77 to 0.81, sensitivity from 0.93 to 0.96, specificity from 0.49 to 0.67, positive predictive values from 0.88 to 0.92, and negative predictive values from 0.62 to 0.79 across the five folds for the prediction of responses to ven/aza.

In summary, our deep learning model accurately identified responders (CR/CRi) to ven/aza first-line therapy in newly diagnosed AML patients using only bone marrow aspirate smear images. Our findings demonstrate the ability of deep learning to detect novel layers of information in microscopic images, agnostic to typical response predictors such as gene mutations and other biological data. As the field works to develop alternative therapies for patients unlikely to respond to ven/aza, such a model could serve as a rapid “companion prognostic” tool at initial diagnosis, guiding patient enrollment in clinical trials investigating ven/aza combined with novel therapies or non-venetoclax-based regimens. Improving outcomes for venetoclax non-responders is an urgent need for newly diagnosed AML patients, and the first step to accomplishing this goal will be to reliably predict, a priori, who these patients are.