Legionellosis or Legionnaires' disease is a serious form of pneumonia caused by Legionella bacteria, and is an emerging One Health problem in Ontario and globally. Forecasting is an important aspect of disease surveillance since forecasts of the expected incidences provide a baseline against which observations can be assessed in support of an early warning system for

disease outbreaks.

Traditionally, forecasts were derived from statistical time series analysis, the most prominent being the SARIMA model. Machine learning and artificial intelligence (AI) offer additional solutions, including the application of neural network algorithms and, more recently, pretrained foundational models for forecasting.

This case study analyzed monthly Legionella incidence reports from Ontario, 2013-2024, using four time series forecasting methods, including automated SARIMA modeling as a benchmark and zero-shot forecasting using TimeGPT. The fully pre-trained transformer based TimeGPT model is a foundational time series model based on the GPT large language model. Forecasts were visualized and further compared using two accuracy measures: RMSE and MAD.

Time series exploration of the legionellosis incidence by STL revealed an increasing trend and a strong seasonal pattern with summer peaks. The legionellosis incidence increased with rising ambient temperatures.

In conclusion, zero-shot forecasting using a foundational AI model is not necessarily a golden bullet that offers a window into the future with less uncertainty. There is no best model or method that is consistently outperforming its competitors. However, it is certain that legionellosis is an emerging One Health issue of public health concern requiring sustained environmental and public health surveillance.

Keywords: Legionnaires' disease; forecasting; artificial intelligence; surveillance; evaluation