Deep Learning-Driven 24-Hour Mortality Risk Assessment through Peripheral Blood Smear Morphology in Hospitalized Cancer Patients

Introduction: Mortality in cancer pateints is often preceded by catastrophic hematologic and inflammatory syndromes such as the systemic inflammatory response syndrome/sepsis, disseminated intravascular coagulation (DIC), microangiopathic hemolytic anemia, hemophagocytic lymphohistiocytosis, and cytokine release syndrome. These conditions alter peripheral blood cells, causing changes in the morphology of red blood cells (RBC), white blood cells (WBC) and platelets in peripheral blood smears (PBS). We hypothesize that attention-based deep learning models using PBS feature embeddings can predict short-term mortality risk and quantify the underlying biology contributing to that risk.

Methods: We utilized 790,000 PBS from Memorial Sloan Kettering Cancer Center (2017-2024), imaged by CellaVision, with corrected cell level clinical annotations. This dataset included 790,000 RBC/platelet and 100 million WBC single cell morphology images. We selected 20,000 RBC/platelet and 20,000 WBC images to train two ResNext50 models that generate feature embeddings. Our cohort consisted of 627 PBS from patients who died within 24 hours and 15,863 control PBS, split 60/20/20% respectively for training, validation, and testing. These slides came from 1,565 unique patients with no overlap between the test and training and validationsets. We extracted features and trained a multiple instance deep neural network with self-attention to predict 24-hour mortality risk. A logistic regression classifier using 8 CBC parameters and patient age served as the baseline model.

We analyzed the top 10% of highest attention cells from correctly predicted mortality events (mortality-wbc, mortality-rbc) and control cases (control-wbc, control-rbc) to identify morphologies linked to increased mortality risk. A pathologist performed a qualitative analysis, and we used the two ResNext classifiers to perform a quantitative analysis of class distribution changes between the two sets using a two-proportion z-test.

Results

Performance: When run on a test set with 3,172 slides from control patients, and 134 slides from deceased patients, the trained MIL model achieved an AUC = 0.83. The logistic regression model using CBC and age had an AUC = 0.71.

Attention-Based Pathologist Morphology Analysis:

mortality-rbc: “anisopoikilocytosis with increased echinocytes, schistocytes, tear drops, and spherocytes, increased polychromatophilic cells.

control-rbc: Unremarkable.

mortality-wbc: Increased erythroblasts, often with abnormal/dysplastic nuclei, increased large/giant platelets, increased smudge cells, some of which have neutrophil extracellular trap (NET) morphology.”

control-WBC: “Normal neutrophils, lymphocytes or monocytes.”

Computational Quantitative Analysis:

When comparing the top 10% highest attention RBC patches, the mortality group showed an increase in echinocytes (54%:0%) and schistocytes (29%:3%), and a decrease in normal RBC patches (4%:72%) (all p<10^-10). For WBCs, the mortality group had more erythroblasts (44%:1%), smudge cells (14%:1%), and giant thrombocytes (12%:1%), with fewer segmented neutrophils (5%:44%), band neutrophils (6%:15%), and lymphocytes (4%:13%) (all p<10^-10).

Conclusions: Deep learning-based analysis of PBSs could provide clinical decision support in hospitalized patients by assessing their 24-hour mortality risk. These patients may require escalation of care and specific interventions depending on the nature of the risk. Qualitative and quantitative analyses of the model’s attention scores suggest it relied heavily on echinocytes, erythroblastosis, schistocytes, large platelets, smudge cells and NETs. The elevation of erythroblasts, giant platelets, and schistocytes may signal a profound marrow response to severe hemolysis, tissue hypoxia, and consumptive coagulopathy as seen in severe DIC.