A unified Foot and Mouth Disease dataset for Uganda: evaluating machine learning predictive performance degradation under varying distributions.

Abstract

Listen

Translate article

In Uganda, the absence of a unified dataset for constructing machine learning models to predict Foot and Mouth Disease outbreaks hinders preparedness. Although machine learning models exhibit excellent predictive performance for Foot and Mouth Disease outbreaks under stationary conditions, they are susceptible to performance degradation in non-stationary environments. Rainfall and temperature are key factors influencing these outbreaks, and their variability due to climate change can significantly impact predictive performance. This study created a unified Foot and Mouth Disease dataset by integrating disparate sources and pre-processing data using mean imputation, duplicate removal, visualization, and merging techniques. To evaluate performance degradation, seven machine learning models were trained and assessed using metrics including accuracy, area under the receiver operating characteristic curve, recall, precision and F1-score. The dataset showed a significant class imbalance with more non-outbreaks than outbreaks, requiring data augmentation methods. Variability in rainfall and temperature impacted predictive performance, causing notable degradation. Random Forest with borderline SMOTE was the top-performing model in a stationary environment, achieving 92% accuracy, 0.97 area under the receiver operating characteristic curve, 0.94 recall, 0.90 precision, and 0.92 F1-score. However, under varying distributions, all models exhibited significant performance degradation, with random forest accuracy dropping to 46%, area under the receiver operating characteristic curve to 0.58, recall to 0.03, precision to 0.24, and F1-score to 0.06. This study underscores the creation of a unified Foot and Mouth Disease dataset for Uganda and reveals significant performance degradation in seven machine learning models under varying distributions. These findings highlight the need for new methods to address the impact of distribution variability on predictive performance.