Development of a Dynamic Counterfactual Risk Stratification Strategy for Newly Diagnosed Patients With AML Treated...

Area of Science:

• Hematology
• Machine Learning in Medicine
• Clinical Informatics

Background:

• Acute myeloid leukemia (AML) treatment requires precise risk stratification.
• Venetoclax plus azacitidine (ven/aza) is a key AML therapy.
• Existing risk models may not fully address real-world data (RWD) complexities or treatment specificity.

Purpose of the Study:

• To develop a flexible risk stratification strategy for AML tailored to venetoclax plus azacitidine (ven/aza) therapy.
• To create a strategy that accommodates real-world data (RWD) challenges and is adaptable to various use cases.
• To generate tunable risk models (RMs) using machine learning (ML).

Main Methods:

• Utilized a dynamic counterfactual machine learning (ML) strategy to generate tunable risk models (RMs).
• Employed features from diagnostic AML samples and tested models on a cohort of 316 newly diagnosed patients treated with ven/aza.
• Validated RM performance using diverse model assumptions, data elements, endpoints, and an external real-world cohort (RWC).

Main Results:

• Identified favorable, intermediate, and adverse risk groups using ML-based RMs with genetic and/or phenotypic data.
• Achieved equitable patient distribution (20%-40% per group) and significant risk strata separation (P <0.001).
• Demonstrated strong predictability (AUC 0.60-0.70) and comparable performance to European Leukemia Net 2022 on an external RWC.

Conclusions:

• The proposed ML strategy effectively addresses RWD considerations for AML risk stratification.
• The strategy is tunable via coding and parameter updates, enhancing adaptability for different contexts and use cases.
• This novel approach offers a more effective method for developing risk models in AML and potentially other diseases.