Factors Associated with Legionella Detection in the Water Systems of National Lodging Organization Facilities with Water Management Programs in the United States

A better understanding of risk factors and the predictive capability of water management program (WMP) data in detecting Legionella are needed to inform the efforts aimed at reducing Legionella growth and preventing outbreaks of Legionnaires’ disease. Using WMPs and Legionella testing data from a national lodging organization in the United States, we aimed to (1) identify factors associated with Legionella detection and (2) assess the ability of WMP disinfectant and temperature metrics to predict Legionella detection. We conducted a logistic regression analysis to identify WMP metrics associated with Legionella serogroup 1 (SG1) detection. We also estimated the predictive values for each of the WMP metrics and SG1 detection. Of 5435 testing observations from 2018 to 2020, 411 (7.6%) had SG1 detection, and 1606 (29.5%) had either SG1 or non-SG1 detection. We found failures in commonly collected WMP metrics, particularly at the primary test point for total disinfectant levels in hot water, to be associated with SG1 detection. These findings highlight that establishing and regularly monitoring water quality parameters for WMPs may be important for preventing Legionella growth and subsequent disease. However, while unsuitable water quality parameter results are associated with Legionella detection, this study found that they had poor predictive value, due in part to the low prevalence of SG1 detection in this dataset. These findings suggest that Legionella testing provides critical information to validate if a WMP is working, which cannot be obtained through water quality parameter measurements alone.


Introduction
The bacterium Legionella can cause diseases, ranging from severe pneumonia (i.e., Legionnaires' disease (LD)) and the rarer but severe extrapulmonary legionellosis, to the milder respiratory illness of Pontiac fever.More than 95% of reported case-patients with LD in the United States are hospitalized, and approximately 10% result in death [1].Outbreaks of LD are often linked to poorly maintained water systems in settings such as hotels and healthcare facilities.The Centers for Disease Control and Prevention (CDC) and ASHRAE (formerly known as the American Society of Heating, Refrigerating and Air-Conditioning Engineers) recommend facilities with certain characteristics or devices develop and implement water management programs (WMPs) to help reduce the risk of Legionella growth/spread and to reduce outbreaks [2,3].WMPs include the establishment of controls, such as disinfectant residuals and temperatures, as well as measurements, to ensure that values are within control limits.The US Department of Veterans Affairs and the Centers for Medicare and Medicaid Services have WMP requirements.
Testing building water systems and devices for Legionella has several purposes, including establishing baseline measurements, assessing the impact of remedial measures, and investigating sources of exposure during public health investigations of the disease [4].In addition, testing is a way to confirm that a WMP is effectively working (validation) [2,3,[5][6][7].The CDC has outlined multifactorial indicators for routine Legionella test result interpretation for WMP performance, including the concentration of Legionella and associated trends over time, the extent of Legionella growth (i.e., one vs.multiple sources), the location of detection within the water system, and the Legionella species' association with LD [4,6].
A better understanding of risk factors and the predictive capability of WMP data in detecting Legionella are needed to inform future efforts to reduce Legionella growth and prevent outbreaks.In this analysis, by using WMP and Legionella testing data from a national lodging organization (NLO) with over 700 lodging facilities located in the United States, we aimed to (1) identify factors associated with Legionella detection and (2) assess the ability of WMP metrics to predict Legionella detection.

Data
The NLO shared 2018-2020 WMP and Legionella testing data, the details of which have been described elsewhere [8].The NLO requires each facility to have a WMP and to conduct at least annual environmental testing for Legionella by traditional spread-plate culture from the building's hot or cold water premise plumbing system.Positive culture results were categorized as Legionella pneumophila serogroup 1 (SG1) or non-SG1.The WMP data include free and total disinfectant levels, grouped into categorical levels, from primary test points in hot water (i.e., tap most distal from the water heater within the hot water distribution system), cold water (i.e., tap closest to the building's incoming water main), supply temperature data (i.e., water leaving the water heater), and return temperature data (i.e., hot water recirculation systems prior to reheating).

Outcome
Because SG1 is more strongly associated with disease than other serogroups and species [9], we define our primary outcome variable as binary SG1 detection (yes/no).When facilities conducted multiple tests per day, the results were aggregated by day for each facility (i.e., any positive test vs. none on that day), keeping tests that came from hot vs. cold water systems separate.In a sensitivity analysis, we also examined any Legionella detection (SG1 and/or non-SG1, i.e., Legionella pneumophila serogroups other than SG1 or other species of Legionella) as the outcome.

WMP Failure Metrics
For each day that Legionella testing occurred in a facility, we created six binary WMP failure metrics to identify if a failure had been detected at any point in the five weeks prior to the testing date.The failure metrics included any failure in the past five weeks in the following: (1) the return water temperature in guest rooms, (2) the supply water temperature in guest rooms, (3) the primary test point for total disinfectant levels in cold water, (4) free disinfectant levels in cold water at the primary test point, (5) the primary test point for total disinfectant levels in hot water, and (6) free disinfectant levels in hot water at the primary test point (Table 1).We used a five-week timeframe to balance temporality (i.e., addressing failures that occurred too far in the past) and sample size (because disinfectant and temperature testing were not conducted every week at every facility).We define failure based on control limits established in the NLO policy, which include a control limit for a return temperature of ≤118 • F and a supply temperature of <124 • F. For free disinfectant levels, the control limits in hot and cold water are less than 0.4 mg/L and 0.5 mg/L, respectively, and for total disinfectant, the control limits in hot and cold water are less than 1.0 mg/L each.We exclude any reported temperature values above 165 • F, as those are likely incorrect (<1% of total values; see [8] for more details on WMP data).

Statistical Analysis
We conducted a mixed-effects logistic regression analysis with random intercepts for the facility to identify WMP metrics associated with SG1 detection.We tested models with different combinations of the six WMP failure variables, using the Akaike information criterion (AIC), Bayesian information criterion (BIC), and the likelihood ratio test (for nested models) to determine the best fitting model.In all models considered, we also controlled for the hot vs. cold water system, the season-to account for the seasonality of LD (high: February-July vs. low: August-January), and year (2019 vs. 2018 and 2020 vs. 2018).Variables were considered statistically significant at alpha = 0.05 if their 95% confidence intervals did not include 1.We also assessed the interactions between temperature and disinfectant failures.

Predictive Measures
We calculated the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for each WMP failure metric and SG1 detection (as well as any Legionella detection, SG1 or non-SG1) as part of our sensitivity analyses.Sensitivity was calculated as the percentage of times the WMP metric indicated a failure when there was SG1 detection; specificity was calculated as the percentage of times the WMP metric did not indicate a failure when there was no SG1 detection; PPV was calculated as the proportion of times there was SG1 detection when the WMP metric indicated a failure, and NPV was calculated as the proportion of times there was no SG1 detection when the WMP metric did not indicate a failure.

Results
After aggregating daily data by the facility and water system, there were 5435 testing observations in the dataset from 2018 to 2020 (January 2018-November 2020; building operations consisted of low or no hotel occupancy during the COVID-19 pandemic).Of these observations, 411 (7.6%) had SG1 detection from 85 of the 725 facilities (12%), and 1606 (29.5%) had either SG1 or non-SG1 detection from 230 facilities (32%).Each WMP metric was available in the five weeks prior to each Legionella test for approximately 75% of Legionella tests (Appendix A).
In the best fitting model based on the AIC, which included three of the WMP failure metrics, the following variables were significantly associated with SG1 detection: failure regarding total disinfectant levels in hot water, samples collected from the hot water system (vs.cold), and samples collected in 2020 (vs.2018) (Table 2; see Appendices B and C for the other models considered).Failure at the primary test point for total disinfectant levels in hot water had the strongest association with SG1 detection: any failure in the past five weeks had 4.1 times the odds of SG1 detection (95% CI: 2.0, 8.5), compared to no failure in the past five weeks.Any failure in the past five weeks in the return water temperature in guest rooms had a positive but non-significant association with SG1 detection (OR = 1.6, CI = (0.8, 3.2)).The cooccurrence of failures at both primary test points for total disinfectant levels in hot and cold water was more common than the cooccurrence of failures in guest room return temperature with failures at either primary test point for total disinfectant levels (Appendix D).Interactions between temperature and disinfectant failures were not significant and were, therefore, not included in the final model.Despite the associations identified between WMP failures and SG1 detection, failures in WMP metrics were not strong predictors of SG1 detection (Table 3, Appendix A).Consistent with the strong association observed, the failure sensitivity in the past five weeks at the primary test point for total disinfectant levels in hot water for SG1 detection was 87.7%.However, the PPV was only 5.6%, meaning that most of the time when this failure occurred, there was no SG1 detection.We found qualitatively similar results when using any Legionella detection as the outcome, with slightly higher PPV; this finding was expected, given the higher percentage of positivity for SG1 or non-SG1 compared to SG1 only (Appendix E).

Discussion
With facility-level WMP metrics and Legionella test results, this dataset from an NLO provides detailed insights into factors associated with Legionella detection, which can help to inform the development of WMPs and Legionella testing practices.We found failures in commonly collected WMP metrics, particularly at the primary test point for total disinfectant levels in hot water, to be associated with SG1 detection.A similar analysis conducted in Veterans Health Administration healthcare buildings also found hot water samples and lower residual biocide concentrations to be associated with Legionella detection [10].These findings highlight that WMPs may be important for reducing the risk of Legionella growth and subsequent disease; moreover, it may be important to regularly monitor water quality parameters, such as disinfectant levels and temperature, to ensure they are within expected ranges.
However, while unsuitable water quality parameter results are associated with Legionella detection, this study found that they had poor predictive value for Legionella detection.This finding is not surprising due in part to the low prevalence of SG1 detection in this dataset, as positive predictive value is influenced by prevalence; this means that when prevalence is low, failures in WMP metrics have less ability to predict Legionella detection than when prevalence is higher.Similarly, as expected with low prevalence, the NPV is much higher than the PPV, meaning that suitable water quality parameters suggest a decreased likelihood of Legionella detection, but they do not guarantee that Legionella growth is well-controlled.
The odds of Legionella detection were significantly higher in 2020 than in 2018, potentially reflecting the impacts of reductions in building occupancy and water usage during the COVID-19 pandemic.The impact of COVID-19 on WMP performance was described in detail in our recent study [8].
Several limitations of the NLO data have been described in detail elsewhere [8].For these analyses, limitations to the data include a lack of information on the geography of where facilities are located (due to identifiability concerns), facility age, or occupancy fluctuations due to seasonal use, building closures, or renovations that occurred during the pre-pandemic months, all of which could potentially impact the odds of Legionella detection.Additionally, the disinfectant data did not specify the type of disinfectant used and were provided categorically, with zero disinfectants included in the category with the lowest levels.The number of data points by facility also varied greatly, although we attempted to account for this in our analytical approach.As described earlier, defining failures based on reported WMP data in the five weeks prior to a Legionella test was attempted to balance temporality and sample size.However, little prior research exists to determine the optimal window for the impact of WMP measures; in the present study, WMP measurements were not available for about 25% of the samples (Appendix A).Given the varying frequencies of testing for disinfectant and temperature levels, failures could have gone undetected.However, analyses of WMP failure metrics using data from two, three, and four weeks showed similar trends to the analysis using five weeks of data, strengthening our conclusions.Finally, due to the limitations in the Legionella data [8], such as variations in sampling and testing techniques that can impact the reliability of Legionella concentration data, we were not able to use the CDC's multifactorial approach (as described earlier) [4] to interpret routine testing results.Therefore, we could not assess whether each sampling event was consistent with a system in which Legionella growth was well-controlled, poorly controlled, or uncontrolled.

Conclusions
This analysis builds on previous analyses [8], which found increased odds of SG1 detection during the COVID-19 pandemic compared to 2018-2019, with increased positivity driven by facilities that also had Legionella-positive sample results before the pandemic.Those findings suggested the NLO's flushing protocols may have prevented some Legionella growth, but that additional control measures may be needed for some facilities.In both studies, we find that-while they remain an important part of any WMP-water quality parameter results that meet control limits do not guarantee that Legionella growth is wellcontrolled.These findings suggest that Legionella testing provides critical information to validate if a WMP is working, which cannot be obtained through water quality parameter measurements alone.

Appendix B. Random Intercept Model Choice
We chose the mixed effect modeling approach, with a random intercept for the facility to account for the correlation between repeat measures at the same facility.We also considered using a generalized estimating equation (GEE) approach.However, to properly model the correlation in a GEE model, the data need to be aggregated up to months or quarters due to the wide variations in the number of repeated measurements across facilities.This aggregation makes interpretation more challenging and temporality less clear.We also considered using Poisson instead of logistic regression models, with a percent positive rate, to account for repeated testing on the same day (or month, if aggregated).This approach was challenging due to the high proportion of zeros (i.e., negative tests), and there were some model fitting and convergence issues with zero-inflated models.

Appendix C. Random Intercept Modeling
We tested models with different combinations of the six WMP failure variables.Model 1 included all WMP failure variables, and the subsequent models removed variables based on the bivariate AICs or subject matter knowledge (e.g., Model 4 was selected based on AIC but Model 5 included both hot and cold water disinfectant failures).The final model was chosen based on the model metrics.In all models considered, we also controlled for the hot vs. cold water system, season (high vs. low), and year (2019 vs. 2018 and 2020 vs. 2018).

Table 1 .
Failure variables and definitions.

Table 2 .
Odds ratios comparing SG1 detection to failure to detect *.
Any failure in the past five weeks at the primary test point for total disinfectant in cold * significant.