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

Class imbalance is common in epidemiological studies. To date, no comprehensive investigation has been conducted to evaluate the efficacy of various class-imbalance handling strategies for epidemiological forecasting. Therefore, this study aimed to explore the potential of multiple machine learning techniques in addressing class imbalance through a stroke prediction case study.

A total of 11140 eligible participants (5136 males and 6004 females) aged 45 or above were included from the China Health and Retirement Longitudinal Study (CHARLS). Using 15 predictors, we constructed stroke prediction models based on 3-year follow-up data (2015-2018). The outcome was self-reported doctors' diagnosis of stroke. Sequential forward selection was used for variable selection. Six machine learning algorithms combined with data resampling, threshold tunning, cost-sensitive learning, ensemble learning, and anomaly detection were used to construct sex-specific stroke prediction models. Accuracy, sensitivity, positive predictive value (PPV), G-mean, and area under the ROC curve (AUROC) were applied to evaluate model performance.

The incidence of stroke over a 3-year period was 5.9 % and 5.6 % for men and women, respectively. All models demonstrated suboptimal performance on imbalanced dataset. After using machine learning techniques to address class imbalance, the performance improved significantly, especially for local outlier factor from anomaly detection, with its sensitivity, PPV, and G-mean reaching 0.98, 0.59 and 0.92 for male and 0.93, 0.63, and 0.91 for female.

Machine learning showed potential in addressing class imbalance, which would be beneficial for epidemiological prediction studies.