Data collection

Preferred reporting items for systematic reviews and meta-analyses (PRISMA) guidelines were employed to collect reports and cases of Bp from different sources since its inception from October 17, 2022, until April 28, 2023 (Fig 1). The melioidosis database (www.melioidosis.info) was used to collect records. Search for reports in PubMed database was done using combinations of keywords and MESH terms, namely, “Burkholderia pseudomallei”, “melioidosis”, “Southeast Asia”, and “sampling”. One thousand forty-two records were initially identified from these databases. Before screening, two hundred thirteen duplicate records were removed. The remaining eight hundred twenty-nine records were then assessed for eligibility by 1) selecting records obtained from SEA, 2) selecting only records obtained from environmental soil sampling, and 3) selecting records with precise geographic coordinates. The accuracy of all the one hundred forty-four geographic coordinates used for this modeling study was verified by conducting an initial data check. In accordance with the approach used by Limbo-Dizon and co-workers [18], the spThin package in R software using the spatial filtering approach was used to reduce spatial autocorrelation, which is the tendency of closer locations to be more similar than those further apart. In this procedure, the presence records were randomly selected according to a minimum nearest neighbor distance ≥10 km between each locality [19]. This approach resulted in ninety-nine geographical coordinates from the initial one hundred forty-four coordinates. The procedure was repeated ten times to produce a spatially filtered data set. These filtered geographical coordinates were saved as a comma-delimited (CSV) file type. These files are now publicly available at Zenodo (https://doi.org/10.5281/zenodo.11507319).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. PRISMA flow diagram.

Depiction of core procedures used in this review process such as the identification of records from different sources (databases and registers), removal of records, screening, and the inclusion of the remaining reports.

https://doi.org/10.1371/journal.pntd.0012684.g001

Acquisition of environmental layers

Factors identified in the study were hypothesized and considered to contribute to the presence of Bp in an area. Variables included are (1) soil characteristics from the Harmonized World Soil Database (HWSD) (http://www.iiasa.ac.at) and (2) nineteen bioclimatic variables from the WorldClim database (www.worldclim.org). Map layers come from the USGS public domain. All the downloaded environmental layers from HWSD and WorldClim databases were in GeoTIFF format. The ArcGIS software converted all the environmental layers into an ASCII extension. Using ENMtools in R software, highly correlated variables >0.7 were excluded from the study resulting in seven environmental variables (Table 1) from the initial nineteen variables [20]. This was done to 1) avoid overfitting, 2) obtain accurate response curves, and 3) prevent the model from being adjusted excessively to the data [21].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Percent Contribution and Permutation Importance of Selected Environmental Variables for the Suitability of Bp in SEA Countries.

https://doi.org/10.1371/journal.pntd.0012684.t001

Species distribution modeling

Maximum Entropy Model (MaxEnt Version 3·4·4) from https://biodiversityinformatics.amnh.org/opensource/maxent/ was used to predict the potential environmental niche of Bp. MaxEnt predicts the distribution (geographic range) of a given species by using presence-only data and finding the distribution that has maximum entropy (i.e. is the most geographically uniform) while being restricted to environmental constraints [22]. The software can handle continuous and categorical predictor variables while also automatically incorporating variable interactions. Specifically, we used various characteristics using environmental variables such as topographic, edaphic, and bioclimatic to restrict the geographic distribution. In this open-source software, an evaluation of the species’ habitat suitability ranges from 0 (least suitable) to 1 (highest suitability) [23,24].

Prior to the final model run, the dataset was subjected to an analysis using ENMeval R-based evaluation package [25]. This data quality filtering was included to optimize model complexity in order to balance the goodness of fit and predictive ability avoiding any possibility of overfitting [26]. The fine-tuned setting produced from the ENMeval analysis (method = randomkfold, kfold = 10) recommended the adjustment of a regularization multiplier (RM) of 3 and a feature type (FT) of a linear feature. To determine the significance of each environmental variable and its relative contribution to the model prediction “Create a response curve” and “Do jackknife test to measure variable importance” were ticked respectively. Model performance was assessed using the generated Area under the ROC (receiver-operator characteristic) curve (AUC) which ranges from 0 to 1 and is a metric for model performance [27]. The MaxEnt ASCII output file format was imported into the ArcMap 10·8 software for better visualization. For the specific classification of classes, the map was divided using an equal five-part interval. Each species’ habitat suitability was then translated from the five-part interval into five categories namely, very low suitability, low suitability, moderate suitability, high suitability, and very high suitability.

Model limitations

The usage of robust machine learning predictions such as the MaxEnt algorithm has been increasing rapidly in the span of the last decade. In fact, for the Southeast Asian region alone, this algorithm has been using correlation mechanisms to predict suitable environmental habitats for elusive animals [30], plants [31], soil protist [32], phytopathogens [33], and even cases of neglected parasitic diseases [34]. Despite such promising potentialities that the predictive power of a model can generate, it is important to note that such models are heavily reliant on excellent types of data being employed for the model. Hence, besides the number of recorded occurrences for the Bp, we are limited on which environmental predictors to use. Therefore, for this study, data that are easily available and transformed into meaningful spatial analysis are relatively considered. As such, despite the knowledge of complex mechanisms that can transpire in melioidosis, like land use, anthropogenic activities, and climate change, we have first limited the models to data that are much correlative in nature (bioclimatic, edaphic, topographic), like what has been done in the aforementioned studies. This is because abiotic factors such as those that were selected for this study better support environmental assumptions about melioidosis [35,36]. For a more mechanistic approach, it is highly suggested that we consider factors that are highly attributed for melioidosis and such theme warrants further investigations once reliant data becomes available for the region.

Important Predictors for Bp Precipitation

Our results showed both bioclimatic and edaphic predictors are equally significant in our model demonstrating the hydroclimatic influences on Bp-receptive soils where its distribution is directly implicated with the rise of melioidosis. Seasonal foci of disease outbreaks have long been associated with heavy rainfall events [11]. Subsequent bacterial output from groundwater seeps after rain or flood, exacerbated by high humidity and windy conditions, were postulated to bring about the shift towards airborne mode of transmission. This leads to a more serious form of illness and contribute to clinical clustering [11,14]. The effects of Bio12 were less significant to the model performance than the other variables (Bio8, Biocateg1) because of low values of percent contribution and permutation importance. This can possibly be attributed to changes in precipitation pattern and monsoonal weather conditions [13,9]. As shown in Fig 4A, an annual precipitation of <2300 mm contributes to high to very high suitability for the bacterial pathogen. It was observed that Bp was more common in Myanmar, Thailand, Laos, Cambodia, and Vietnam during the Southwest monsoon with an annual precipitation of ~1000 to 2800 mm [37] and similarly depicted by our model in Figs 3 and 6. This can also be associated with melioidosis cases peaking annually during the rainy season (May–October) in both Laos and Cambodia whereby the microorganism accumulating on surface soils occurs as a result of increasing water tables brought only by incremental rainfall [38,14]. This follows the trend observed in Fig 4A where there is less possibility for the bacterial presence in places with annual precipitation of over 2300 mm.

Soil type

The contribution of Biocateg1 in our model cannot be ignored with its potential habitat since the bacterium is intrinsically a soil dweller. Acrisol type contributed either high or very high suitability for Bp (see Fig 5). Our model has identified acidic, clay rich Acrisol consistent with the study of Goodrick [16] and Pongmala [39] and their co-workers. Other preceding studies accounted for not only Acrisol but also Anthrosol [7] or Luvisol type [15] while only Anthrosol type was associated by Birnie and co-workers [10]. The discrepancies in predicting soil type in these different reports may be due to various detection techniques and different modeling schemes used. Nevertheless, the implication of primarily associating Acrisol type or with other soil types is with the infection risk for melioidosis since many individuals are working for their livelihood in the tropical rural areas. Most often, they are involved in agriculture or domesticated animal industry. This is exemplified by workers tilling soil in rice paddy fields barefoot and are constantly exposed to the soil environment [15,4]. These rural areas in SEA are the ones lacking adequate medical services that missed out on the definitive diagnosis of melioidosis. Thus, we presently do not have the true picture of disease burden in this region. This is exacerbated by various anthropogenic activities on a shift in rural land use.

Temperature

Our predictive model suggests temperature (≥26°C) is the most influential of all the environmental predictors identified (Bio8) and has the greatest impact of more than sixty-seven (67·9%) percent contribution and forty-six (46·4%) in terms of permutation importance. Referring to the term used in this software, the mean temperature of the wettest quarter is the index that approximates the mean temperatures that prevail during the wettest season [40]. In an earlier study by Liu and co-workers [38], the increasing cases of melioidosis were noted from July to October, and January having a mean temperature of between 27.7°C ± 0.7°C. Similarly, SEA countries experiencing greater amounts of rainfall during the Southwest monsoon season and increasing ambient temperature were observed to have a higher risk for melioidosis [12,7]. Investigation of the physiologic characteristics of Bp at a wider range of temperatures demonstrated the bacterium can sustain growth, motility, biofilm formation, and resistance to oxidative stress at 37°C [41]. This was further supported by findings of the optimal temperature for the growth of Bp in soil media were between 37°C and 42°C [42]. All these previous studies coincided with our findings shown in Fig 4B that areas with higher temperature (≥26°C) have an increased likelihood for the bacterial presence. With the present onset of global warming, the ecology of this microorganism is likely to change the disease dynamics and epidemic potential amid sea surface and ambient temperature increase. This has been clearly anticipated by many studies [11,12,7,13,10].

Implications for Bp epidemiology

Overall, the predictive model suggested that the potential highest suitability of Bp presence peaks during the Southwest monsoon, where there is usually heavy rainfall with the sea surface and ambient temperatures at more than 26°C in Acrisol-laden areas are the conditions most suitable for this bacterial pathogen plotted in our MaxEnt distribution model (Fig 6). Most areas, that exhibited high to very high suitability, are clustered around mainland SEA (Cambodia, Laos, Malaysia, Myanmar, Thailand, and Vietnam). Some areas in East Timor, Indonesia, and Philippines also exhibited high to very high suitability. Brunei and Singapore did not show high suitability for the bacterium. Such information can now be targeted to document the seasonal burden of melioidosis in terms of morbidity and mortality and draw public health guidelines and safeguards, including those with additional health risk factors in acquiring melioidosis [8]. This will then reduce the disparity of true number of cases in these areas with the numbers predicted by many models. Our study suggests rising temperature, annual precipitation and geographic presence of Acrisol soil are important environmental-determining drivers for the distribution of Bp in SEA under the current climate scenario. The best-fitting model revealed its predicted occurrence is more influenced by bioclimatic variables than edaphic factors that will change melioidosis epidemiology due to the ongoing climate change. As such, there is a need to predict further future Bp geographical distribution within the areas of high endemicity and beyond. Hence, this regional modeling that scopes only the tropical to subtropical climates of the Southeast Asian region is already a big epidemiological step in utilizing innovative technologies such as machine learning to establish early public health warning systems before the disease becomes another global public health problem.