2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Sunday, December 13, 2026
06:00 p.m. - 08:00 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Molecular Testing, Emerging Technologies, Profiling, Treatment Considerations, Combination Therapy, Drug Development, Supportive Care, Therapy Sequence, Non-Biological Therapies, Workforce, Adverse Events, Artificial Intelligence (AI), Generative AI, Machine Learning, Biological Processes, Apoptosis, Pathogenesis, Genomics, Hematopoiesis, Immune Mechanism, Metabolism, Molecular Biology, Clinical Practice (Health Services and Quality), Clinical Guidelines, Diseases, Bleeding and Clotting, Platelet Disorders, Genetic Disorders, Immunology, Autoimmune Disorders, Infectious Diseases, Myeloid Malignancies, Acute Myeloid Malignancies, Chronic Myeloid Malignancies, Measurable Residual Disease, Research, AML, Translational Research, MDS, MPN, CMML, Clinical Research, Clinical Trials, Epidemiology, Health Outcomes Research, Real-World Evidence, Study Population, Human, Adult, Elderly, Young Adult
PAPER

3857 A machine learning model for noninvasive diagnosis of ITP vs MDS in older patients with isolated thrombocytopenia

Howard S Oster, MD PhD, Vasileios Papadopoulos, MD2, Zmira Silman, MS3, Gil Shenberg, MD4, Noa Goldschmidt, MS5, Sandra Huber, PhD6, Konstantinos Liapis, MD7, George Vrachiolias8, Jacqueline Ferrari, MD9, Martina Boldini, MD10, Sara Pozzi11, Anna Gallì, Bsc12, Chiara Elena, MD13, Torsten Haferlach, MD, PhD14, Luca Malcovati, MD15, Moshe Mittelman, MD16, Ioannis Kotsianidis, MD PhD17
1 Ichilov Sourasky Hospital, Tel Aviv-Yafo, Israel, Tel Aviv University, Tel Aviv-Yafo, Israel
2 Democritus University Of Thrace, Alexandroupolis, Anatoliki Makedonia Thraki, Greece
3 Independent Consultant, Netanya, Israel
4 Ichilov Sourasky Hospital, Tel Aviv-Yafo, Israel
5 Ichilov Sourasky Hospital, Tel Aviv-Yafo, Israel
6 MLL Munich Leukemia Laboratory, München-Hadern, Germany
7 Democritus University Of Thrace, Alexandroupolis, Anatoliki Makedonia Thraki, Greece
8 Democritus University Of Thrace, Alexandroupolis, Anatoliki Makedonia Thraki, Greece
9 Fondazione Irccs Policlinico San Matteo, Pavia, Italy
10 Fondazione Irccs Policlinico San Matteo, Pavia, Italy
11 Fondazione Irccs Policlinico San Matteo, Pavia, Italy
12 Fondazione Irccs Policlinico San Matteo, Pavia, Italy
13 Fondazione Irccs Policlinico San Matteo, Pavia, Italy
14 MLL Munich Leukemia Laboratory, München-Hadern, Germany
15 Fondazione Irccs Policlinico San Matteo, Pavia, Italy
16 Ichilov Sourasky Hospital, Tel Aviv-Yafo, Israel, Tel Aviv University, Tel Aviv-Yafo, Israel
17 Democritus University Of Thrace, Alexandroupolis, Anatoliki Makedonia Thraki, Greece

Background:

Half of the patients with myelodysplastic syndrome/neoplasms (MDS) have pancytopenia, and 5-10% have isolated thrombocytopenia. Immune thrombocytopenic purpura (ITP), although common in young patients may occur in the elderly. Older patients with isolated thrombocytopenia are particularly challenging, and the correct diagnosis is crucial since treatment and prognosis are different.  Bone marrow examination is invasive and often inconclusive. We have previously published our model to diagnose MDS vs non-MDS in patients with unexplained anemia/cytopenia (Oster et al. Bld Adv 2021). Here we present a non-invasive machine learning algorithm using 5 simple parameters that can assist in diagnosis to differentiate between ITP and MDS.

Methods:

Two cohorts, proven ITP and MDS from Greece, served for the development of the models (80% training). Using variables age, sex, hemoglobin (HB), white blood cell count (WBC), platelet count (PLT), we evaluated 4 models: Logistic Regression, Decision Tree, Random Forest and XGBoost, comparing via 5-fold cross-validation. A score (0 – 1) predicted diagnosis (≥0.7 ITP; <0.3 MDS, 0.3-0.7 defined as indeterminate). Models were evaluated by assessing sensitivity (correct identification of ITP) and specificity (MDS), as well as positive and negative predictive values (PPV, NPV, proportion correctly predicted ITP or MDS respectively) and the area under the ROC curve (AUC). Subpopulation analysis was conducted on elderly patients with isolated thrombocytopenia. Finally, we performed internal validation, using the reserved 20%, and external validation using data of three cohorts from Israel, Italy and Germany.

Results:

The original two cohorts included 3790 patients. After restricting age≥18y, PLT≤100 and excluding those with incomplete data, the training set (80%) included 1202 patients (ITP 544, MDS 658). Variables were different for ITP and MDS groups respectively: Male sex 41.7% vs 70.7%; Age (mean [95% CI]): 57.1 [55.0, 59.1] vs 72.5 [71.7, 73.3]; HB 12.8 [12.7, 13.0] vs 9.8 [9.7, 10.0]; WBC 9.0 [8.2, 9.8] vs 5.0 [4.5, 5.5]; PLT 24.3 [22.3, 26.3] vs 58.3 [56.3, 60.3]; All with p<0.001. XGBoost was the best model and was used for the rest of the study. The performance of the model training group (mean [95%CI]) was:  AUC=0.977 [0.969-0.983]; sensitivity (ITP) 0.923, specificity (MDS) 0.912, PPV (predicted ITP, pITP → truly ITP) 0.896, NPV (predicted MDS pMDS → truly MDS) 0.935. Adding Neutrophil count to the model had no significant impact.

Internal validation: sensitivity (ITP) 0.897, specificity (MDS) 0.939, PPV (predicted ITP, pITP → truly ITP) 0.924, NPV (predicted MDS pMDS → truly MDS) 0.917. External validation with ITP (n=549) and MDS (n=147): sensitivity (ITP) 0.816; specificity (MDS) 0.918; PPV (pITP → truly ITP) 0.974; NPV (pMDS → truly MDS) 0.572. All measures were excellent except for correctly predicting MDS in external validation.

A web APP for this model that can be used for any patient in the community is provided and the link can be found at https://itp-mds-calculator.netlify.app.

The quality of the model in older patients with isolated thrombocytopenia (age≥65y, HB>12, WBC>4, PLT<100) was confirmed. Internal Validation (n=55): sensitivity 0.957, specificity 0.625, AUC 0.896 . External Validation (n=123): sensitivity 0.822, specificity 0.800, AUC 0.871.

Examples in this subgroup of interest:

A 65 year old woman with HB 14.3, WBC 6.0, PLT 35. Score = 0.994 (99.4% in APP), correctly predicting ITP. A 73 year old man with HB 12, WBC 10.1, PLT 67. Score = 0.054 (5.4% in APP), correctly predicting MDS.

Conclusions:

We have developed an XGBoost model that can differentiate ITP from MDS. Internal and external data sets validated the model. The model is especially useful in elderly patients with thrombocytopenia. The model is a non-invasive simple and inexpensive screening tool that can assist the community physician in diagnosis. ITP prediction is accurate. Especially in MDS diagnosis, some patients may need further diagnostic workup including our MDS prediction model (see above), bone marrow examination and mutational profile.
Howard S Oster, MD PhD: No relevant conflicts of interest to declare.