Enhancing Severe Neutropenia Prediction: PKPD-Informed Labeling for Machine Learning Models Trained on Real-World...

Area of Science:

• Pharmacology and Machine Learning
• Clinical Data Science
• Computational Biology

Background:

• Real-world data (RWD) labeling for machine learning (ML) is hindered by sparsity and imbalances.
• Accurate prediction of docetaxel-induced neutropenia is crucial for patient safety.
• Existing methods struggle with the nuances of RWD for predictive modeling.

Purpose of the Study:

• To develop and evaluate a pharmacokinetic-pharmacodynamic (PKPD)-informed labeling strategy.
• To enhance the risk prediction of docetaxel-induced neutropenia using ML.
• To address data sparsity and imbalance issues in RWD for clinical outcome prediction.

Main Methods:

• Developed a PKPD-informed labeling strategy using semi-mechanistic model simulations to determine neutrophil nadir.
• Trained three ML models (logistic regression, XGBoost, TabPFN) on RWD from 4,248 patients.
• Compared PKPD-informed labeling against a naive labeling method using only neutrophil observations.

Main Results:

• The PKPD labeling approach increased labeled patient instances by 3.4 times (7,719 vs. 2,283).
• ML models trained with PKPD-informed labels showed significantly superior predictive performance (AUC-ROC, AUC-PR) across all architectures.
• Performance gains were consistent even when training set sizes were matched.

Conclusions:

• PKPD-informed labeling effectively overcomes RWD sparsity and imbalance limitations.
• This strategy enhances both the quantity and quality of labels for ML model training.
• The methodology provides a robust and generalizable framework for improving clinical outcome prediction in ML.