Influence of Infection Origin, Type of Sampling, and Weather Factors on the Periodicity of Some Infectious Pathogens in Marseille University Hospitals, France

Abstract

This study aimed at systematically exploring the seasonalities of bacterial identifications from 1 February 2014 to 31 January 2020 in hospitalized patients, considering the infectious site and the community-acquired or hospital-associated origin. Bacterial identifications were extracted from the data warehouse of the Institut Hospitalo-Universitaire Mediterranée Infection surveillance system, along with their epidemiological characteristics. Each species’ series was processed using a scientific workflow based on the TBATS time series model. Possible co-seasonalities were researched using seasonal peak clustering and series cross-correlations. In this study, only the 15 most frequent species were described in detail. The three most frequent species were Escherichia coli, Staphylococcus aureus, and Staphylococcus epidermidis, with median weekly incidences of 145, 74, and 39 cases, respectively. Samplings of S. aureus and E. coli follow the same seasonal dynamics. S. aureus hospital-associated infections exhibited a significant association with temperature, humidity, and pressure change, whereas community-acquired infections were only associated with precipitations. More seasonal peaks were observed during the winter season. Among the 15 peaks of this seasonal maximum, 6.7% came from blood (Klebsiellia oxytoca) and 13.3% from respiratory specimens (E. coli and S aureus). Our results showed significant associations of periodicity between pathogens, origin of infection, type of sampling, and weather drivers.

1. Introduction

Seasonality is a constantly encountered notion in medicine, globally and throughout history. Many infectious diseases are known to be periodical or seasonal, essentially viral infections, but also common bacterial infections [1] (e.g., ‘influenza’ disease takes its name origin from the fact that it is influenced by the cold). However, the introduction of statistical methods adapted to periodicity analyses is more recent, with the systematic use in 1918 of periodograms for investigating measles outbreaks reported in the Bills of Mortality [2]. As Fisman argued, we consider seasonal processes to have “an incidence associated with a particular calendar period, and which have periodicity, although this is not limited to annual periodicity” [3]. Regarding viral infections, seasonality may be driven by vector seasonality, climatic conditions, abiotic or biotic environment, co-infections, viral antigenic drifts, seasonal human immune variation, human behaviors, or seasonality in domestic or wildlife animal hosts [4,5]. Seasonality studies of infectious diseases have long been based on meteorological and climatic factors, which were the most evident associated determinants. For example, Greer et al. reported that norovirus winter outbreaks in Toronto are associated with the seasonal fluctuation of the Lake Ontario temperature, favoring virus survival in winter [6]. Seasonal viral outbreaks due to seasonal climatic-dependent proliferation of vectors belong to the most well-known mechanisms, as for the Rift Valley fever [7] or for arboviruses in general [8]. Seasonality can be influenced by humans’ mobility behavior [9,10] or by periodical socio-cultural or religious gathering, as during the annual pilgrimage organized in Mecca for the Hajj [11]. However, the possibility of confounding factors and then of fallacious associations cannot be eliminated, as the subjacent mechanism is rarely understood [3].

It is only recently that seasonal periodicities of bacterial infections have been systematically studied [12]. Moreover, these systematic studies rarely take into consideration the sites of infection, differentiating, among others, bloodstream, urinary tract, skin and soft tissues, respiratory tract infections, and rarely differentiate infections from colonization. The reason for this is that large scale systematic studies are mainly carried out on surveillance data, which do not integrate this kind of information.

The association of various climatic factors such as temperature, precipitation, sunshine, atmospheric pressure, frost, and snow with a number of bacterial infectious diseases, has been reported in numerous studies [13]. As examples, a raised incidence of salmonellosis has been associated with increased temperature and precipitation [14,15]. Campylobacter spp. infections show a peak in spring [16]. Studies in the UK and Wales [17,18] reported that Campylobacter spp. and Cryptosporidium spp. cases were significantly associated with temperature and rainfall. Gram-negative bloodstream infections, including Escherichia coli, Acinetobacter spp., and Klebsiella spp., also exhibit seasonal variations, being frequently associated with temperature and rainfall [19,20,21]. It has also been suggested that seasonal viral infections may drive seasonal bacterial infection variations, as observed for influenza and invasive pneumococcal [22] and meningococcal diseases [23]. However, these interactions are especially difficult to decipher, as they may result from a spurious correlation or an unmeasured seasonal factor [24].

To date, although systematic studies investigating periodicities were performed for a broad number of pathogens e.g., [25], none took into account both the community-acquired or hospital-associated origin and the type of sampling. In contrast, the present study aimed at systematically exploring bacterial identifications in hospitalized patients, in search of confirmation of the seasonal variations of bacterial infections, taking into account the infection site and community-acquired or hospital-associated origin, and trying to identify climatic drivers.

2. Materials and Methods

2.1. Material

2.1.1. Bacterial Identifications and Related Data

The Institut Hospitalo-Universitaire Méditerranée Infection (IHUMI) is the infectious disease-dedicated hospital of Marseille Public hospitals (Assistance Publique—Hôpitaux de Marseille, APHM). It performs all microbiological analyses for all APHM hospitals, representing about 190,000 bacterial cultures per year. Since 1 February 2014, our microbiological surveillance system (Méditerranée Infection Datawarehouse and Surveillance—MIDaS) has allowed the weekly monitoring of the routine clinical microbiology activity and further in silico statistical investigations [26]. We included in this study all routine bacterial identifications at species level present in the MIDaS Datawarehouse, from 1 February 2014 until 31 January 2020 (6 years, ending just before the COVID-19 outbreak in France). The identifications were deduplicated, keeping the first occurrence of each combination of the hospital stay, the pathogen, and the sample type for analyses at the sample level, and the first occurrence of each combination of the hospital stay and the pathogen for the other ones. Taking into account studies on the effect of data aggregation granularity on time series seasonal analysis [27] and the influence of this granularity on case numbering, we aggregated the data to a weekly level.

All routine bacterial identifications were obtained using a Microflex MALDI-TOF mass spectrometer (company Bruker Daltonics, Bremen, Germany) and the Biotyper software (Bruker Daltonics, Bremen, Germany). As recommended by the manufacturer, bacterial identifications at species level were validated when the Biotyper matching log score was ≥2. Fastidious bacteria were excluded from this study, due to the choice of culture-based routine identifications.

We associated with each identification the epidemiological characteristics available in the data warehouse and required for our study, including the sampling date and origin, a community-acquired (comm)/hospital-associated (hosp) flag (community-acquired if the identification was conducted within the first 48 h following hospital admission), and the patient’s age.

2.1.2. Weather Data

Weather data for the study period were downloaded from the Météo France SYNOP (surface synoptic observations) open data service (https://donneespubliques.meteofrance.fr/?fond=produit&id_produit=90&id_rubrique=32) accessed on 3 February 2020. These data gather 57 weather description variables (Supplementary Table S1) collected from 62 stations in France, with a timestep of 3 h. For this study, we selected the following variables: temperature, rain, humidity, wind, and pressure differentials.

In order to detect possible associations between seasonal variations and weather conditions, we selected from the database the dataset collected by the Marignane station (latitude: 43.44° North, longitude: 5.22° East), which is 20 km away from Marseille. We aggregated the data using the same weekly time step as the epidemiological time series, calculating for each week the minimal and maximal values, the mean, and the sum.

2.1.3. Legal Statement

The surveillance system is in accordance with Regulation (EU) 2016/679 of the European Parliament and the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and the repealing Directive 95/46/EC (General Data Protection Regulation). It is registered by the Data Protection Officer under id 2019-73. All identifications are anonymized in the surveillance database and based on patient consent.

2.2. Methods

For each bacterial species, we made the distinction between hospital-infected infections, with further analyses on urines and blood samples, and community-acquired infections, with further analyses on urines, blood, respiratory, and skin samples. For both infections, specific analyses were made on ages (0–20, 21–40, 41–60, 61–80, and over 80 years-old). A total of 19 times series were then selected for each species (Supplementary Figure S1). Only time series with at least 100 deduplicated observations were considered for statistical analysis, and only the 15 most frequent species will be described in this paper.

For a systematic big data analysis of our database, we created an analysis workflow of each time series seasonality, as follows:

-

A Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test for testing trend stationarity, with stationarity rejection if p > 0.05 [28]. The KPSS test is a type of unit root test that tests for the stationarity of a given series around a deterministic trend. The p-value reported by the test is the probability score, based on which you can decide whether to reject the null hypothesis or not.

-

A ‘Seasonal and Trend decomposition using LOESS’ (STL), which is a versatile and robust method for decomposing time series. The LOESS method used for this decomposition is a method for estimating nonlinear relationships [29]. STL uses LOESS (locally estimated scatterplot smoothing) to extract smooth estimates of the three components. The advantage of this method is that the seasonal component is allowed to change over time, and it is robust to outliers. For the purpose of this study, we have forced the seasonal component to be identical across years.

-

A ‘Trigonometric seasonality, Box-Cox transformation, ARMA errors, Trend and Seasonal components’ (TBATS) analysis for allowing a search for multiple seasonalities [30].

-

An extraction of the detrended time series using the STL results.

This workflow was applied to each of the 19 species’ time series with at least 100 observations.

A co-seasonality analysis was performed using ascending hierarchical classification, based on the Euclidean distance between seasonality peaks, with the UPGMA aggregation method. For further in-depth co-seasonality analyses of selected time series, we performed cross-correlation analyses in order to confirm the statistical significance and the time lag [31].

Analyses of species periodicities in relation to possible meteorological drivers were performed using Poisson mixed-effect time series regressions on detrended series. Meteorological characteristics with a participation in the regression model significative with p ≤ 0.05 were considered as possible drivers.

The overall statistical analysis workflow is reported in Figure 1. All statistical analyses were performed using R [32] version 4 with packages ‘forecast’, ‘tseries’, ‘season’, and ‘urca’. When it was applicable, we used 0.05 as the threshold of statistical significance.

3. Results

The surveillance dataset used for this study included 314,884 bacterial identifications, deduplicated in 228,365 new sample-related identifications, which were counted on a weekly basis.

A total of 575 different bacterial species were retrieved in this dataset. Of these, only 86 species (15%) met our inclusion criteria (at least 100 new cases over the study period), and only the first 15 most frequent species are described in this paper. The full results are reported in the Supplementary Materials.

3.1. Seasonality Study

The three most frequent species were E. coli, S. aureus, and S. epidermidis, with median weekly incidences of 145, 74, and 39 cases, respectively. The full set of statistics for the weekly incidence series is summarized in

Table 1. Descriptive statistics of the weekly incidence series for the 15 most frequent bacterial species.

N°	Pathogens	Min	1st Qu	Median	Mean	3rd Qu	Max
1	Enterobacter cloacae	3	13	16	16.60	20	34
2	Enterococcus faecalis	9	23	27	27.75	32	48
3	Escherichia coli	109	135	145	145.4	155	203
4	Gardnerella vaginalis	3	16	22	23.46	30	51
5	Haemophilus influenzae	1	7	9	10.03	13	28
6	Klebsiella oxytoca	0	4	6	6.02	8	16
7	Klebsiella pneumoniae	18	31	35	36.06	41	66
8	Propionibacterium acnes	0	4	6	6.35	9	22
9	Proteus mirabilis	5	11	14	14.03	16	33
10	Pseudomonas aeruginosa	18	29	34	33.96	39	55
11	Staphylococcus aureus	45	67	74	74.04	80	111
12	Staphylococcus epidermidis	22	35	39	39.38	44	66
13	Staphylococcus haemolyticus	1	4	6	6.54	8	19
14	Staphylococcus hominis	0	6	8	8.04	10	23
15	Streptococcus agalactiae	7	18	22	21.52	26	39

Min = minimum; Max = maximum; 1st Qu = first quartile; 3rd Qu = third quartile.

.

The systematic periodicity analysis taking into account the sample origin showed that 60% of species (9/15) exhibited at least one seasonality. The full result set for the 15 most frequent species is summarized by the heat map of Figure 2, which reports in green the series with a statistically significant seasonality, and the full set for all 86 bacterial species is reported in the Supplementary Figure S2. Samples coming from hospital-associated infection and from community-acquired infections presented the sample frequency of seasonality (10/15, 66.7%). Seasonality of hospital-acquired infections was specific to three species: G. vaginalis, K. oxytoca, and S. haemolyticus, while seasonality of community-acquired infections was specific to four species: S. epidermidis, S. agalactiae, S. hominis, and P. acnes. P. aeruginosa, and S. aureus exhibited a seasonality for most of the samples, with the exception of urine samples coming from hospital-associated infections of patients younger than 20 years old, and urine samples of the 61–80-year-old population with community-acquired infections.

Some bacterial species, apparently without seasonality at the whole series, the global hospital-associated, or the global community-acquired levels, may exhibit a seasonality at a narrower level. This is the case for K. oxytoca, which presents a seasonality for its blood and respiratory tract levels for its community-acquired infections, or for P. mirabilis, which presents a seasonality only at the respiratory tract level for its community-acquired infections.

A further analysis of seasonal peaks was specifically performed for samples from community-acquired infections (Figure 3), without obscuring seasonal peaks from hospital-associated infections (Figure 4 and Supplementary Figure S4). More seasonal peaks were observed during the winter season, particularly during week 2, with 15 peaks (Figure 5) for 86 species, and during weeks 4 and 6, both with 9 peaks. Among the 15 peaks of this seasonal maximum, 1/15 (6.7%) came from blood (K. oxytoca), and 2/15 (13.3%) from respiratory specimens (E. coli, S. aureus). Most of the seasonal peaks belonging to respiratory samples are observed during winter (10/20, 50%) and spring (7/20, 35%). Blood samples exhibit a bi-modal seasonality with maxima during winter (6/15, 40%) and summer (5/15, 33%). Skin samples show only a seasonal minimum during summer (2/16, 13%), with S. aureus belonging to this group. The apparent seasonality of urinary tract infections is spring, with peaks 8/17 (47%).

3.2. Co-Seasonality and Cross-Correlation Analysis

The result of the hierarchical clustering of the seasonal peak distances for community-acquired infections is presented in Figure 6. Kind of sampling with seasonality are grouped according to the proximity of their seasonal peak. The dendrogram shows that samplings of S. aureus and E. coli follow the same seasonal dynamics, and that the seasonality of community-acquired infections of S. aureus is the same as its respiratory samples, but opposite to its skin samples. Infections of the respiratory tract of S. aureus and E. coli are synchronous, with a lag between them and H. influenzae. Several seasonal complexes may be found: a complex associating S. hominis and S. epidermidis with most of their locations, concerning S. aureus, E. coli, and P. aeruginosa, a first complex associating their respiratory locations and a second associating their blood infections.

The analysis showed that, for S. aureus, samples coming from all community-acquired infections and respiratory samples of these infections were synchronous with a strong cross-correlation (r = 0.56), both with their peaks during week 2 (Figure 7). This shows that the apparent overall seasonality of S. aureus community-acquired infections is mostly driven by respiratory infections. In contrast, cutaneous and respiratory S. aureus community-acquired infections exhibited a shifted cross-correlation (r = 0.24) (Figure 8), with peaks, respectively, during weeks 38 and 2. These results show that the apparent winter seasonality of S. aureus community-acquired infections masks the cutaneous infection seasonality of this species.

3.3. Species Periodicities and Meteorological Drivers

Temperature (°C), rain (mm), humidity (%), wind (m/s), and pressure change (Pa) were tested for significant positive or negative associations with seasonal infection incidences. Significant associations of incidences with possible meteorological drivers for respiratory, urine, blood, and cutaneous samples coming from community-acquired and hospital-associated infections, are presented in

Table 2. Meteorological drivers associated with seasonal times series for the 15 most frequent species.

Pathogens	Sample Size	Hospital-Associated Infections		p-Value	Community-Acquired Infections		p-Value
		Samples	Drivers		Samples	Drivers
S. aureus	23,173	all samples	temperature	0.009	all samples	rain	0.004
			humidity	0.011	urines	rain	0.037
			pressure change	0.013		pressure change	0.039
		urines	rain	0.021
S. epidermidis	12,325				resp	rain	0.018
						humidity	0.030
K. pneumoniae	11,287				urines	humidity	0.016
P. aeruginosa	10,631	blood	rain	0.020	blood	humidity	0.019
					skin	rain	0.043
						pressure change	0.015
G. vaginalis	7344	all samples	rain	0.013
		urines	temperature	0.011
S. agalactiae	6735	urines	temperature	0.005	resp	rain	0.010
			rain	0.013		humidity	0.029
E. cloacae	5195	all samples	wind	0.045	all samples	humidity	0.002
			pressure change	0.019
H. influenzae	3139	all samples	humidity	0.013	all samples	humidity	0.022
		blood	humidity	0.008		pressure change	0.004
					blood	rain	0.023
						pressure change	0.011
					urines	wind	0.049
S. hominis	2516	urines	pressure change	0.043
S. haemolyticus	2048	all samples	humidity	0.024
		urines	rain	0.023
			humidity	0.033
			wind	0.046
P. acnes	1986				all samples	humidity	0.016
					skin	rain	0.001
K. oxytoca	1884	all samples	rain	0.041

for the 15 most frequent bacterial species. Concerning S. aureus, hospital-associated infections exhibited a significant association with temperature, humidity, and pressure change, whereas community-acquired infections were only associated with precipitations. Community-acquired H. influenzae infections were globally associated with humidity and pressure changes (precipitations and pressure change for blood samples, and wind for urine samples).

E. cloacae exhibited significant associations with pressure changes and wind for hospital-associated infections, and with humidity for community-acquired infections.

K. pneumoniae infections were associated with humidity for urine samples in community-acquired infections.

Among a total of 156 combinations of bacterial species, origin, and sample types with a seasonality (Supplementary Table S2), 59/156 (37.8%) were significantly associated with precipitations, 43/156 (27.6%) with pressure change, 38/156 (24.4%) with humidity, 36/156 (23.1%) with wind, and 23/156 (14.4%) with temperature. In contrast, 19/156 (12.2%) combinations had no significant association with any meteorological parameter.

3.4. Seasonal Weekly Indexes per Pathogen

We constructed this table using the ratio between the weekly incidence and the 52 week rolling average (Supplementary Table S3) [33]. We added an equivalent index built from the seasonal decomposition of the time series by LOESS: the ratio for each week between the seasonal overhead (which may be negative) and the trend (Supplementary Table S4). This last index made the periodicity clearer, as it suppressed the error (random variations). For example, for E. cloacae, in the week of 4 August 2014its seasonal index was 1.09, indicating that this week recorded 9% more than the seasonal average. However, for the same species, its seasonal index was 0.90 in the week of 11 August 2014, which means that this week recorded 10% less than the seasonal average (Supplementary Table S3).

4. Discussion

This seasonality study was based on the university hospital microbiological laboratory activities from an area of the East Mediterranean coast of France, characterized by abundant sunshine, warm and dry summers, windy episodes, and associated with rainfalls from October to April. This big data systematic cross-analysis, taking into account the infection origin, types of samples, and age category, and not only bacterial species identifications [13,17,34,35], was processed using a statistical workflow in search of seasonal periodicities of the incidence time series. These analyses were supplemented by the search for co-seasonalities and association with meteorological drivers. From a methodological point of view, the originality of this study resides in the distinction made on the sample origin in the systematic analysis, which allows for the decomposition of the bacterial species’ overall time series into infection-related time series, which may have a different seasonality. A workflow was established for controlling series stationarity and for handling varying periodicities. A hierarchical clustering of the time series seasonal peaks allowed for a global exploration of co-seasonalities, with possible deeper analyses using cross-correlations. The weekly granularity of the time series brought us more precision in the location of the seasonal peak, and also enabled a more precise search for meteorological drivers. However, this last study must be taken with caution, as the linear model used does not take into specific consideration the possible lag between the periodicities of the meteorological drivers and the infections. In the same way, it must be considered that an apparent driver may in fact be the result of confusion with other determinants, as in the case of the seasonalities of patients’ metabolic conditions (e.g., the seasonality in the skin production of vitamin D), or even more complex factors in the case of the seasonality of hospital-acquired infections. We believe that this study is the first that has systematically examined the association between climatic drivers and bacterial species while taking into account sample kinds.

From 228,365 unique bacterial identifications during 6 years of data collection, 86 bacterial species had at least one time series with at least 100 observations, and were able to be analyzed by the workflow. Among them, this paper focused specifically on the 15 most frequent bacterial species. The same identification method (MALDI-TOF) with the same protocol has been used throughout the collection, enforcing the homogeneity of the identification capability.

The first result of this work is that seasonality is more frequent than non-seasonality. An incidence seasonality was retrieved for 55.9% of the time series belonging to the 15 most frequent species. Hospital-associated infection without sample distinction showed a seasonality for 2/3 of these species, and the same frequency of seasonalities in community-acquired infections without sample distinction. In the study by Al-Hasan et al. [19], over half the infections were community acquired (59%). It is important to underline here that the seasonality of a specific infection may be hidden in the non-seasonality of the global species incidence, as for K. oxytoca, and that an overall apparent seasonality of a species may be the result of the seasonalities of its various forms of infection with different phases. For example, S. aureus exhibited a community-acquired respiratory infection peak during winter, a community-acquired skin infection peak during summer, and an overall seasonal peak of community-acquired infection during winter, corresponding to the peak of its most abundant sub-population. Epidemiologically speaking, the microbiological point of view at species level may be different than the clinical infection point of view. In our study, winter is the season with the most seasonal peaks, especially the second week.

The association analysis of species with weather variables identified several possible drivers. H. influenzae incidence was associated with humidity and pressure change, and Staphylococcus capitis incidence with temperature and precipitations. Several previous studies have reported peaks in infections caused by Gram-negative bacteria during warmer months. For example, a multicenter study of Acinetobacter infections in patients in the United States between 1987 and 1996 found that infection rates were more than 50% higher in the warmer months than in the colder months [36]. Our study showed that hospital-associated S. aureus blood infections were significantly associated with temperature, as shown by Eber et al. [20], who estimated that an increase in temperature of 5.6 °C was associated with a 2.2% increase (95% CI 1.3–3.2) in S. aureus frequency. Our results are also consistent with those obtained by H Richet, who observed an increase in the correlation between temperature and blood infection rates for S. aureus, P. aeruginosa, E. coli, and K. pneumonia [37]. However, there are contradictions in the results of studies on bloodstream infections in Klebsiella spp. In addition, other studies have shown the presence of a S. aureus seasonal peak during summer [38] or autumn [39], or both [40], depending on the kind of clinical picture. In fact, with only a statistical co-occurrence analysis, it is difficult to determine if a species’ seasonality is driven by a weather condition, or if the weather condition is only a characteristic of the season corresponding to the species seasonal peak.

Similar studies of Campylobacter and Salmonella seasonality in the UK [16,25] demonstrated that the prevalence of Campylobacter spp. was associated with temperature, while our study showed an association of Campylobacter with precipitation and humidity (Supplementary Materials), and other studies found limited association with temperature [14]. Hospital-associated H. influenzae infections were associated with pressure change for all age groups, and humidity for the 0–20-year-old age group (Table 2). It was reported by Shaman and Kohn [41] that atmospheric pressure changes could have strong associations with several infectious agents, such as the influenza virus, creating optimal conditions for the manifestation of H. influenzae. In our study, week two gathered the most important number of seasonal peaks, which may be due to the low temperature and high humidity during this winter period. Most Gram-negative bacteria such as E. cloacae, E. coli, K. pneumoniae, and P. aeruginosa did not exhibit any significant association with temperature except Acinetobacter baumannii, in contrast with other studies [1,21,37] reporting a higher incidence of positive blood cultures for these species during the summer months.

This systematic big data study on bacterial identifications of hospitalized patients shows that seasonality is a frequent, but not systematic characteristic, and that the seasonalities of bacterial species are different from the seasonalities of the different infections of the species. Our study of meteorological drivers of seasonality must be expanded with further works that take into account the possible variable lags between disease and meteorological time series.

The limitations of this work lie in the fact that the correlations between temperature and the incidence of these pathogens can be distorted by a large number of non-measurable variables [21], because the possibility of confounding factors and fallacious associations cannot be eliminated, as the subjacent mechanism is rarely understood.

This study, once applied in hospitals, will enable us to understand periodicity in nosocomial pathogens, and will therefore strengthen vigilance on these pathogens through the surveillance system and establish preventive measures aimed at minimizing the impact of these seasonal diseases during high-risk seasons.