Prediction and interpretation of pathogenic bacteria occurrence at a recreational beach using data-driven algorithms

초록

Recreational beaches face a threat from pathogenic bacteria that harbor antibiotic resistance genes (ARGs). To predict bacterial occurrence and comprehend their non-linear relationship with hydrometeorological features, advanced machine- and deep-learning algorithms were employed. These algorithms include regression trees (RT), as well as interpretable deep-learning algorithms such as the 'Input Attention-Long Short-Term Memory (IA-LSTM)' and 'Temporal Fusion Transformer (TFT)'. Our focus was on predicting the occurrence of Prevotella, a prevalent pathogenic bacterium found at the beaches. Utilizing model-dependent and model-agnostic interpretation methods, which encompass sensitivity analysis, permutation, and the SHapley Additive exPlanations (SHAP) importance, we evaluated model behavior. RT-based algorithms exhibited predictive capabilities comparable to those of IA-LSTM and TFT, achieving validation Nash-Sutcliffe efficiencies of 0.93, 0.94, and 0.96, respectively. However, the deep-learning algorithms (IA-LSTM and TFT) are surpassed in terms of interpretability. The model-dependent interpretation method identified heavy precipitation as a pivotal hydrometeorological feature linked to increased Prevotella occurrence. Notably, the IA-LSTM identified Prevotella as a potential host for the sulfonamide resistance gene (sul1), suggesting the potential of Prevotella as an indicator for sul1. This research, leveraging interpretable data-driven models, advances our understanding of the hydrometeorological features influencing the occurrence of pathogenic bacteria and the prevalence of ARGs at the beach, and enhances predictive capabilities for bacterial occurrence.

키워드