Imbalanced prediction in epidemiological study: A machine learning-based analysis

Area of Science:

• Epidemiology
• Data Science
• Machine Learning

Background:

• Class imbalance is a prevalent challenge in epidemiological research, potentially compromising predictive model accuracy.
• Existing methods for handling imbalanced data in epidemiological forecasting lack comprehensive evaluation.
• Stroke prediction is a critical area where accurate forecasting is essential for timely intervention.

Purpose of the Study:

• To evaluate the efficacy of various machine learning techniques in managing class imbalance for epidemiological forecasting.
• To explore the potential of multiple machine learning algorithms in improving stroke prediction models.
• To compare the performance of different imbalance-handling strategies in a real-world epidemiological context.

Main Methods:

• Utilized data from 11,140 participants (aged 45+) from the China Health and Retirement Longitudinal Study (CHARLS).
• Developed sex-specific stroke prediction models using 15 predictors and 3-year follow-up data (2015-2018).
• Applied six machine learning algorithms combined with data resampling, threshold tuning, cost-sensitive learning, ensemble learning, and anomaly detection.

Main Results:

• Stroke incidence was 5.9% for men and 5.6% for women over 3 years.
• Initial models on imbalanced data showed suboptimal performance.
• Machine learning techniques significantly improved model performance, with anomaly detection (Local Outlier Factor) yielding high sensitivity, PPV, and G-mean.

Conclusions:

• Machine learning approaches demonstrate significant potential for addressing class imbalance in epidemiological studies.
• These techniques can substantially enhance the performance and reliability of predictive models for diseases like stroke.
• The findings support the integration of advanced machine learning strategies into epidemiological forecasting.