Navigating Severe Class Imbalance in Population Cohort Data.

Area of Science:

• Machine Learning
• Medical Informatics
• Genomics

Background:

• Class imbalance poses significant challenges for predictive modeling of rare disease outcomes in large population studies.
• Traditional machine learning algorithms often exhibit biased performance and poor generalizability on imbalanced datasets.

Purpose of the Study:

• To systematically evaluate various methods for mitigating class imbalance in predicting multiple myeloma using proteomic and clinical data.
• To compare standard classifiers, resampling techniques, anomaly detection, and a foundation model for rare disease prediction.

Main Methods:

• Compared XGBoost, logistic regression, SMOTE, isolation forests, local outlier factors, one-class SVM, autoencoders, and TabPFN.
• Introduced a sequential XGBoost ensemble (SeqXGB) to prioritize precision.
• Evaluated models using standard classification performance metrics on UK Biobank data.

Main Results:

• Anomaly detection models demonstrated superior generalization compared to conventional classifiers (XGBoost, logistic regression).
• SMOTE did not improve, and potentially worsened, predictive performance.
• SeqXGB significantly reduced false positives but also substantially decreased sensitivity.

Conclusions:

• No single method universally addresses class imbalance; model selection must align with clinical application and the trade-off between false positives and negatives.
• Findings emphasize the need to consider the specific clinical purpose (e.g., screening vs. diagnosis) when evaluating machine learning models for rare diseases.