Artificial intelligence-based prediction models for acute myeloid leukemia using real-life data: A DATAML registry...

Area of Science:

• Hematology
• Artificial Intelligence
• Oncology

Background:

• Acute myeloid leukemia (AML) management involves complex treatment decisions.
• Large datasets from patient registries pose challenges for clinical interpretation.
• Predictive models are needed to aid in prognosis and treatment selection for AML patients.

Purpose of the Study:

• To develop and validate artificial intelligence-based prediction models (AIPM) for overall survival (OS) in AML patients.
• To identify key diagnostic variables for accurate OS prediction in AML.
• To assess the utility of AIPM in guiding treatment choices for intensive chemotherapy (IC) and azacitidine (AZA) therapies.

Main Methods:

• Utilized the DATAML registry data from 3687 AML patients.
• Developed a multilayer perceptron (MLP) neural network for OS prediction.
• Employed the Boruta algorithm to select the most impactful diagnostic variables.

Main Results:

• MLP models achieved 68.5% OS prediction accuracy in the IC cohort and 62.1% in the AZA cohort.
• The Boruta algorithm identified 13 key features for the IC cohort and 7 for the AZA cohort without reducing prediction accuracy.
• Selected features included age, cytogenetics, blood counts, specific biomarkers, and mutations.

Conclusions:

• AIPM can accurately predict OS in AML patients treated with IC or AZA.
• Feature selection algorithms effectively reduce the number of diagnostic variables needed for prediction.
• These models can assist hematologists in managing AML data and optimizing treatment strategies.