Public health agencies are increasingly relying on genomics during Legionnaires' disease investigations. However, the causative bacterium () has an unusual population structure, with extreme temporal and spatial genome sequence conservation. Furthermore, Legionnaires' disease outbreaks can be caused by multiple genotypes in a single source. These factors can confound cluster identification using standard phylogenomic methods. Here, we show that a statistical learning approach based on core genome single nucleotide polymorphism (SNP) comparisons eliminates ambiguity for defining outbreak clusters and accurately predicts exposure sources for clinical cases. We illustrate the performance of our method by genome comparisons of 234 isolates obtained from patients and cooling towers in Melbourne, Australia, between 1994 and 2014. This collection included one of the largest reported Legionnaires' disease outbreaks, which involved 125 cases at an aquarium. Using only sequence data from cooling tower isolates and including all core genome variation, we built a multivariate model using discriminant analysis of principal components (DAPC) to find cooling tower-specific genomic signatures and then used it to predict the origin of clinical isolates. Model assignments were 93% congruent with epidemiological data, including the aquarium Legionnaires' disease outbreak and three other unrelated outbreak investigations. We applied the same approach to a recently described investigation of Legionnaires' disease within a UK hospital and observed a model predictive ability of 86%. We have developed a promising means to breach genetic diversity extremes and provide objective source attribution data for outbreak investigations. Microbial outbreak investigations are moving to a paradigm where whole-genome sequencing and phylogenetic trees are used to support epidemiological investigations. It is critical that outbreak source predictions are accurate, particularly for pathogens, like , which can spread widely and rapidly via cooling system aerosols, causing Legionnaires' disease. Here, by studying hundreds of genomes collected over 21 years around a major Australian city, we uncovered limitations with the phylogenetic approach that could lead to a misidentification of outbreak sources. We implement instead a statistical learning technique that eliminates the ambiguity of inferring disease transmission from phylogenies. Our approach takes geolocation information and core genome variation from environmental isolates to build statistical models that predict with high confidence the environmental source of clinical during disease outbreaks. We show the versatility of the technique by applying it to unrelated Legionnaires' disease outbreaks in Australia and the UK.