Type: Oral
Session: 803. Emerging Tools, Techniques, and Artificial Intelligence in Hematology: New Approaches to Predicting Patient Outcomes in Hematologic Malignancies
Hematology Disease Topics & Pathways:
Research, Clinical Research, Diseases, Immune Disorders, Technology and Procedures, Study Population, Human
Methods: Baseline characteristics were collected from patients (n=441) enrolled in EXPLORER, PATHFINDER, and PIONEER. Diagnoses of ISM (n=265) and AdvSM (n=176, including aggressive SM [n=29], mast cell leukemia [n=28], and SM with associated hematologic neoplasm [n=119]) were determined at trial enrollment by expert review according to WHO 2016 diagnostic criteria. The dataset was randomly split (60:40) into training and test cohorts. A random forest algorithm identified which of the following parameters were independently predictive of ISM vs AdvSM status: age, race, sex, country of origin, alanine transaminase, albumin, alkaline phosphatase (ALP), presence of ascites, aspartate transaminase, total bilirubin, creatinine, absolute neutrophil count (ANC), absolute basophil count, absolute eosinophil count, absolute lymphocyte count, absolute monocyte count, platelets, hemoglobin, KIT D816V variant allele frequency (VAF) in peripheral blood, history of anaphylaxis, presence of pleural effusion, splenomegaly, mast cell percentage on bone marrow biopsy, and serum tryptase. The most important variables identified were then used in stepwise logistic regression to build the final models.
Results: Initial iterations identified splenomegaly and bone marrow biopsy mast cell percentage as important. However, these parameters can have low inter-rater reliability, limiting applicability of the model, and were excluded from model design. A model (Model 1) was developed that predicted AdvSM vs ISM with an area under the curve (AUC) of 0.97 in the test cohort, using the following objective parameters: age, platelets, absolute monocyte count, hemoglobin, ALP, tryptase, and total bilirubin. As C-findings such as thrombocytopenia and anemia are already used for AdvSM diagnosis, an additional model (Model 2) was developed removing the following C-findings from consideration: pleural effusion, ascites, hemoglobin, ANC, and platelets. Model 2 was still able to predict AdvSM vs ISM with an AUC of 0.95 in the test cohort and employed the following parameters: age, ALP, tryptase, total bilirubin, albumin, absolute monocyte count, and absolute lymphocyte count. By Model 1, 31/441 patients were misclassified (of these, 14 were clinically diagnosed with ISM, 17 with AdvSM). By Model 2, 33/441 patients were misclassified (of these, 12 were clinically diagnosed with ISM, 21 with AdvSM). Interestingly, many of the patients with clinically diagnosed ISM who were misclassified by our models as having AdvSM had high risk disease characteristics, with over half of these patients possessing KIT D816V VAFs >6% and tryptase >100 ng/mL. Application of the model to an independent validation set of patients with SM from Dana-Farber Cancer Institute will be presented.
Conclusion: We describe two predictive models, each using age plus a combination of objective, easily measured parameters in peripheral blood, that can distinguish between AdvSM and ISM with high accuracy. Patients aberrantly classified by these models may identify current limitations in our understanding of SM biology.
Disclosures: Lampson: Blueprint Medicines Corporation: Current Employment, Other: Shareholder. Zakharyan: Blueprint Medicines Corporation: Current Employment, Current holder of stock options in a privately-held company. Shi: Blueprint Medicines Corporation: Current Employment, Current holder of stock options in a privately-held company. Deangelo: Amgen: Consultancy; Autolos: Consultancy; Agios: Consultancy; Blueprint: Consultancy, Research Funding; Forty-seven: Consultancy; Gilead: Consultancy; incyte: Consultancy; Jazz: Consultancy; Novartis: Consultancy, Research Funding; Pfizer: Consultancy; servier: Consultancy; Takeda: Consultancy; Abbvie: Research Funding; Glycomimetics: Research Funding.