Artificial intelligence vs conventional models to predict risk contamination of Legionella in the water network of an Italian hospital pavilion
More details
Hide details
University of Bari Aldo Moro, Bari, Italy, Interdisciplinary Department of Medicine, University of Bari Aldo Moro Piazza G. Cesare 11, 70124 Bari, Italy.
University of Bari Aldo Moro, Bari, Italy
University of Bari Aldo Moro, Bari, Italy, Department of Precision and Regenerative Medicine and Ionian Area (DiMePRe-J), Italy
University of Bari Aldo Moro, Bari, Italy, Interdisciplinary Department of Medicine, University of Bari Aldo Moro, Italy
A.O.U Policlinico of Bari, Italy
A.O.U Policlinico of Bari, Italy, U.O.C Hygiene - Environmental and Food Hygiene Laboratory, Bari Policlinico University Hospital, Italy
A.O.U Policlinico of Bari, Italy, Health General Management, Bari Policlinico University Hospital, Italy
Publication date: 2023-04-26
Popul. Med. 2023;5(Supplement):A944
Background and Objective:
Periodic monitoring of Legionella in hospital water network allows to take preventive actions to be taken to avoid legionellosis risk of patients and health professionals. The aim of study was to verify the effectiveness of innovative vs conventional model to predict Legionella contamination risk in a hospital water supply.

In the period February 2021- October 2022, water samplings for Legionella spp were carried out in the rooms of a hospital pavilion (89.9%), located in Apulia region, southern Italy. Fifty-six different parameters regarding structural aspects, water supply points, network maintenance, sampling method, atmospheric temperature were collected. Machine learning and Poisson regression models were tested, in order to predict water network contamination by Legionella. Models were built with 70% of the dataset and tested with the remaining 30%, evaluating accuracy, sensitivity and specificity. All models have been developed with R program.

Overall, 1,053 water samples were analysed and 57 (5.4%) resulted positive for Legionella. Among different tested machine learning models, the most efficient had an input layer (56 neurons), a hidden layer (30 neurons), and an output layer (2 neurons). The accuracy was 93.4%, sensitivity 43.8%, and specificity 96%. Regression model had accuracy 82.9%, sensitivity 20.3% and specificity 97.3%. Combination of the models resulted in 79.1% accuracy, 22.4% sensitivity, 98.4% specificity. The combination models included 79% of the “really” positive samples in prediction of Legionella presence. The most important parameters influencing the model resulted: type of water network (hot/cold), replacement of valve filters and atmospheric temperature.

Tested models gave excellent results of accuracy and specificity, with low sensitivity. The combination of models have a good reliability in setting controls of the hospital water network. Future studies are required to extend the dataset including analysis in other pavilions of the hospital, in order to improving model’s performance.

Journals System - logo
Scroll to top