Identification of relevant features using SEQENS to improve supervised machine learning models predicting AML...

Area of Science:

• Computational biology
• Medical informatics
• Oncology

Background:

• Acute myeloid leukemia (AML) poses a significant challenge in clinical decision-making.
• Predicting complications at diagnosis is crucial for effective therapeutic strategies.
• Existing methods may not fully leverage comprehensive patient data for risk stratification.

Purpose of the Study:

• To evaluate the SEQENS algorithm for feature selection in predicting AML complications.
• To validate machine learning models for predicting AML patient risk at 90 days, six months, and one year post-diagnosis.
• To develop a tool supporting clinical decisions and understanding AML risk factors.

Main Methods:

• Utilized a dataset of 568 patients with demographic, clinical, genetic, and cytogenetic data.
• Applied an enhanced SEQENS algorithm for feature selection tailored to each prediction time point.
• Compared four machine learning classifiers (XGBoost, MLP, Logistic Regression, Decision Tree) to assess feature selection impact.

Main Results:

• SEQENS identified key features like Age, TP53, -7/7Q, and EZH2 as consistently relevant across all time points.
• XGBoost achieved the highest ROC-AUC scores: 0.81 (90-day), 0.84 (6-month), and 0.82 (1-year).
• Feature selection generally maintained or improved model performance, validated on an external set.

Conclusions:

• Developed predictive models with performance suitable for clinical decision support at various post-diagnosis intervals.
• Selected features align with European LeukemiaNet (ELN) 2022 risk classification.
• SEQENS effectively reduced feature dimensionality while preserving predictive accuracy for AML complications.