Related Experiment Video
Updated: Jan 10, 2026
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
7.3K
Predicting secondary myeloid neoplasms in acquired aplastic anemia using machine learning models
Ahmet Celal Toprak1, Mutlu Mete2, Julia J Shi1
1Division of Hematology and Oncology, Department of Internal Medicine, University of Texas Southwestern Medical Center, Dallas, TX.
Blood Neoplasia
|November 24, 2025
Summary
Machine learning models predict secondary myeloid neoplasms in aplastic anemia patients treated with immunosuppressive therapy. These tools can guide hematopoietic stem cell transplantation decisions and surveillance strategies.
Area of Science:
- Hematology
- Oncology
- Machine Learning in Medicine
Background:
- Acquired aplastic anemia (AA) patients treated with immunosuppressive therapy (IST) have a significant risk (up to 20%) of developing secondary myeloid neoplasms (sMNs).
- Hematopoietic stem cell transplantation (HSCT) is curative for AA and prevents sMNs, but historically, IST was the first-line treatment for older patients or those without donors.
- Improved HSCT outcomes necessitate better tools to identify high-risk patients for early transplant consideration or tailored surveillance.
Purpose of the Study:
- To develop and validate machine learning models for predicting the development of sMNs in adult patients with acquired AA.
- To identify key clinical predictors of sMN development at diagnosis and after IST response.
Main Methods:
- Analysis of clinical data from 275 adult AA patients treated between 1975 and 2023 across three major medical centers.
- Collection of 79 clinical variables, including demographics, somatic mutations, and treatment response.
- Development of two binary machine learning models (neural networks) using leave-1-out cross-validation to predict sMN risk at diagnosis and 6 months post-IST response.
Main Results:
- Both models demonstrated strong predictive performance (AUC=0.82, sensitivity=0.82, specificity=0.73).
- Key predictors for sMN development included *DNMT3A* mutation, *CUX1* mutation, total mutation count, and patient age.
- *TET2* mutation predicted risk at diagnosis (model 1), while paroxysmal nocturnal hemoglobinuria clone presence predicted risk after IST response (model 2).
- High-risk classification was significantly associated with poorer overall survival (P < .0001).
Conclusions:
- Machine learning models show feasibility for predicting sMN risk in AA patients.
- These models can aid in personalized decision-making regarding HSCT and post-IST surveillance.
- Further validation with larger datasets is warranted to refine these predictive tools.
