AI-driven forecasting of Vibrio vulnificus in the Southern Baltic Sea using high-resolution data

Vibrio vulnificus poses a growing public health risk in the brackish Baltic Sea, where rising summer sea temperatures can create optimal conditions for this potentially lethal pathogen. V. vulnificus can cause severe infections through open wounds, and identifying periods of elevated risk is key to mitigating rising case numbers. However, current early warning tools at bathing sites remain limited. To facilitate accurate forecasting of Vibrio vulnificus occurrence at future time points, we integrated comprehensive spatio-temporal V. vulnificus datasets with machine learning approaches. These datasets were generated from a twice-weekly sampling campaign (2022-2023) conducted across 15 locations in the Baltic Sea and the Warnow Estuary, from which V. vulnificus abundance was assessed using a multi-method strategy comprising droplet digital PCR, agar-based cultivation, and species-level 16S rRNA gene sequencing. Random Forest, Random-Forest time-lag, and Long Short-Term Memory models trained on concurrently collected physicochemical, biological, and satellite-derived data were used to classify and forecast V. vulnificus presence. In the Baltic Sea, Random Forest models using 16S ribosomal RNA gene community profiles achieved strong predictive performance (Area under the precision-recall curve (AUPRC) = 0.78), whilst Long short-term memory models successfully forecast V. vulnificus up to 4-5 weeks ahead with peak AUPRC values of 0.72-0.78. Operational early-warning tests showed high classification power, demonstrating feasibility for real-time risk prediction, especially with satellite data. The application of prokaryotic, eukaryotic, and satellite-derived datasets as predictors highlights microbial succession as a precursor to Vibrio vulnificus blooms, demonstrating the feasibility of machine learning-based forecasting as an early warning tool to support pathogen surveillance in coastal ecosystems under climate stress.