There has been a growing body of literature showing the effects of temperature on many aspects of the biology of triatomines, particularly from the physiological point of view (molting process, diuresis, protein synthesis, hormone secretion, respiration rate, orientation to radiant heat, shelter selection, performance of flight muscles and flight initiation, electrophysiology of the maxillary stylets, preferred temperatures, rate of weight loss, dispersal, feeding behavior, metabolism, water loss, biting rate, gas exchange patterns and spiracular control, spermiogenesis, heart rate, effects on insecticidal activity, reproductive efficiency, and many others). So below I only review some of the most relevant effects of temperature exclusively related to the triatomines’ life cycle and demographic parameters.

The development time and survival of triatomines eggs as function of temperature began to be studied as early as 1935/36: Clark (1935) showed that the hatching rate of the eggs of Rhodnius prolixus (Stål, 1859) increased with constant temperatures from 15 to 27°C, but that this increase was strongly dependent on relative humidity (hatching rate was higher at higher humidity but the slope of the increase of hatching rate with temperature decreased with humidity); Galliard (1936) determined that exposing triatomines eggs to an increasing and constant temperature from 10 to 35°C, the egg development time (time to hatch) increased by a factor of ≈ 3 in Triatoma vitticeps (Stål, 1859) and Triatoma dimidiata (Latreille, 1811), and by a factor of ≈ 1.4 in R. prolixus; Galliard (1936), applying the Réaumur law, also determined the thermal constant for each species (necessary energy to hatch) as a linear relationship (degree-days). However, the relationship between temperature and development time (a non-linear relationship depending upon complex enzymatic processes), was estimated for the first time for the triatomine eggs of Triatoma guasayana (Wygodzinsky and Abalos, 1941) by Rabinovich et al. (2006) (see below the effects of fluctuating temperatures). In Rhodnius neivai (Lent, 1953) Cabello (1999) found that egg to adult development times were 118.0, 86.7, and 68.9 days at 22, 27, and 32°C of constant temperatures. Villegas García and Santillán Alarcón (2004) obtained a very small difference in the hatching time of Meccus pallidipennis (Stål, 1872) at two constant temperatures: 17 days a 25°C and 15 days at 30°C.

Studies on the effects of temperature on the survival and development times of nymphal stages and longevity of adults of triatomines, are more abundant. There is always a decrease in development time with an increase of a constant temperature: the developmental cycle (egg to adult) lasted 602 days at 25°C and 186 days at 30°C in Triatoma jurbergi (Carcavallo, Galvão & Lent, 1998) (Gomes and da Silva, 2002); the developmental cycle of D. maximus (Uhler, 1894) was 271 and 177 days at constant temperatures of 25 and 30°C, respectively (Da Silva, 1990), a difference of 53.1%; however, for Triatoma infestans (Klug, 1834) the difference was smaller (35.9%), also between 25 (271 days) and 30°C (177 days) (Da Silva and Da Silva, 1988). Cabello (1999) rearing R. neivai at three constant temperatures (22, 27, and 32°C), found a non-linear relationship of total adult mean longevity (24.6, 36.6, and 16.7 days) and life expectancy when entering adult stage (21, 34, and 20.7 days) as a function of those constant temperatures of 22, 27, and 32°C, respectively. In the triatomine M. pallidipennis Villegas García and Santillán Alarcón (2004) found that the average development time from first instar nymph to adult was 171 days at 25°C and 154 days at 30°C (11% reduction), and an adult female longevity of 301 and 379 days (26% reduction) at 25 and 30°C, respectively. Martínez-Ibarra et al. (2008) evaluated the life-cycle of Triatoma mexicana (Del Ponte, 1930) with cohorts under constant temperatures of 25°C/50% RH, 25°C/75% RH, and 30°C/75% RH. Development time from egg to adult was 286, 283, and 256 days, respectively, while no difference was found in nymphal mortality (57, 60, and 64%), respectively.

Reproduction is also affected by temperature in triatomines, although Da Silva (1990) found only a small difference (11.4%) in the total number of eggs laid by females of D. maximus in their first 30 days of adult life: 29.1 and 26.1 eggs at 25 and 30°C, respectively; however, under the same conditions, Da Silva and Da Silva (1988) found that in T. infestans this difference was twice as large (22.5%). In R. neivai Cabello (1999) found that the average number eggs/female/week were 11.1, 14.5, and 17.7, and the average number of eggs per female per lifetime were 264, 724, and 603, for three constant temperatures (22, 27, and 32°C), respectively. In the triatomines M. pallidipennis Villegas García and Santillán Alarcón (2004) found that the average number of eggs laid in a 6-month period were 826 at 25°C and 934 at 30°C, both as constant temperatures.

It is well known that there are important physiological effects between regimes of constant and fluctuating temperatures, even with the same mean values; these effects are represented by the so-called Jensen’s inequality, that states that for non-lineal processes, such as most biological phenomena dependent upon temperature, the effects of fluctuations in environmental temperature cannot be predicted from the mean of an equivalent constant temperature (Ruel and Ayres, 1999). This has been established in the triatomines by experiments carried out by Luz et al. (1999); Damborsky et al. (2005), Rabinovich et al. (2006), and Rolandi and Schilman (2018).

Luz et al. (1999) compared the development time of R. prolixus from hatching to the adult stage at constant (15, 20, 25, 28, 35°C) and alternating (16/8 h: 15/28, 20/25, 25/28, and 25/35°C) temperatures. Nymphs of this species reached the adult stage most rapidly (86.7 days) at a constant temperature of 28°C; under alternating temperatures development time was completed within 100 days or less at 25/28 and 25/35°C, while development time was slowest at fluctuating 15/28 and 20/25°C (>185 days), and at constant 20°C (>300 days). Damborsky et al. (2005) compared the life cycle of Triatoma rubrovaria (Blanchard, 1843) in a climatic chamber under constant temperature, and under the natural fluctuating temperature of the laboratory; under controlled temperature the average hatching time was 15.6 days and the average egg to adult development time was 10 months, while under the natural fluctuating temperature hatching time was 19.1 days and egg to adult development time was 14 months; no difference was found in the egg hatching rate nor in the total nymphal mortality, but a small difference was found in reproduction, with a mean number of eggs/female per life of 817.6 eggs under controlled conditions and of 837.1 eggs under fluctuating conditions (2.4%). Rabinovich et al. (2006) analyzed the effects of fluctuating temperatures on the development time of the eggs of the triatomine T. guasayana using three nonlinear models (Devar, Lactin, and Rueda); these authors found a lower egg development threshold between 15 and 17.5°C, and a thermal death point of eggs at 34.4°C; using a generalized development rate response of T. guasayana eggs to increasing temperature was proposed, with three ranges: between 10 and 14.8°C development rate increases in an accelerating way; between 14.8 and 29.2°C development rate increases more or less linearly, and between 29.2 and 34°C development rate decreases. Rolandi and Schilman (2018) studied the effect of daily temperature fluctuation (DTF) of 17–32°C (mean = 24°C) on R. prolixus, and found that development time and fertility were not affected by DTF, but that fecundity was lower in females reared at DTF than at constant temperature; considering Jensen’s inequality and the species’ tropical distribution these authors predicted that living in a variable thermal environment would have associated higher energetic costs.

Cabello (1999) is one of the few authors to construct a complete life table of a triatomine (in R. neivai), and to evaluate the demographic parameters at three constant temperatures (22, 27, and 32°C). The estimation of the demographic parameters showed a non-linear trend: the net reproduction rate (R₀) was 114.6, 374.4, and 142.4, while the intrinsic rate of natural increase (r₀) was 0.14, 0.25, and 0.22, for constant temperatures of 22, 27, and 32°C, respectively.

It has been known in general that the feeding frequency in triatomines affects some life cycle and demographic parameters, although there is not much quantitative data available on these effects. Braga and Lima (2001) manipulated in the laboratory the feeding frequency of Panstrongylus megistus (Burmeister, 1835) by creating six groups (GI to GVI) of 15 couples each, with the following feeding schedule: GI, on days 5 and 25; GII, on days 5 and 35; GIII, on days 5 and 45; GIV, on day 20; GV, on day 30; and GVI, on day 40 after the imaginal ecdysis (an additional group of 15 couples, fed on day 5 after the imaginal ecdysis and subsequently fortnightly, were used as a control group). The results indicated a clear reduction in longevity and fecundity: GI produced more eggs, mating and fertile eggs; GII had longer life spans, higher fecundity and hatching rate; GIII had a shorter life span, low fecundity, fertility, and hatching rate; GVI presented the least favorable results for all parameters.

Cabello et al. (1988) got contradictory results in the longevity and life expectancy of adults of the triatomine R. neivai, when reared at two feeding frequencies (every 7 and 15 days) depending upon the animal species used as blood source: longevity and life expectancy became longer when fed every 7 days (as compared to being fed every to 15 days) when fed on rabbit, but became shorter from 7 to 15 days when fed on chicken. However, several demographic parameters (the intrinsic rate of natural increase, r₀, among them) were consistently higher at 7 days feeding interval than at 15 days feeding interval, independently of the animal species used as blood source.

Sant’Anna et al. (2001) claimed that the triatomine R. prolixus achieves high values of total blood ingestion rate (depending on the host used) due to its high mechanical and salivary efficiency, and that this leads to higher nutritional status, allowing (through shorter development times and higher fecundity) to have a higher population growth rate and reach higher densities.

Being triatomines almost exclusively hematophagous, they are affected not only by the frequency of feeding as shown above, but also by the source of the animal blood ingested. Nattero et al. (2013b) tested the effects of the blood of guinea pigs and pigeons on the fecundity of Triatoma patagonica (del Ponte, 1929); although these authors found that females fed on pigeons ingested more blood and needed a higher amount of blood to produce an egg than females fed on guinea pigs, there were no differences in the number of eggs laid and hatched between insects fed on each feeding source. The higher amount of blood ingested and consumed by T. patagonica fed on pigeons did not translate into higher fecundity or fertility, and Nattero et al. (2013b) concluded that the lower amount of guinea pig blood ingested was offset by its high nutritional quality.

Even morphological traits are being affected by blood source. Although Nattero et al. (2013a) found that in T. infestans females exhibited significantly greater head size than males for both food sources (mammals and birds), they also found that the greater head size of T. infestans adults fed on guinea pigs is not consistent with variation of body size; these authors claim that although body size seems to respond to food source, phenotypic plasticity in head size was not a simple consequence of allometric plasticity between body and head size, and hypothesized that the diet-induced phenotypic plasticity in head size for bugs fed on guinea pigs might be due to changes in developmental allocation to tissue growth that maintain growth of head size while reducing growth of body size, and that this, in turn, was probably due to the reduced amount of blood ingested when fed on guinea pigs relative to those fed on pigeon. There are other several indications that we should expect a relationship between body measures and feeding in triatomines. Zeledón et al. (1970) found that body length in adult T. dimidiata were larger when collected in the field than those reared in the laboratory, and suggested that this was the result of a more frequent feeding rate of those insects caught in the field. Gürtler et al. (2017) investigated the links among individual bug bloodmeal contents in T. infestans and body length as well as host blood sources and habitats; body length correlated positively with other body-size surrogates and was modified by habitat type, bug stage, and recent feeding (e.g., bugs from chicken coops were significantly larger than pig-corral and kitchen bugs).

Lunardi et al. (2017) did not find variations in the head centroid size in Triatoma williami (Galvão & Col., 1965) related to the blood source, while in T. infestans Nattero et al. (2013a) found that the head was greater when fed on mammals, but the body was greater when fed on birds. Furthermore, Nattero et al. (2013a) demonstrated experimentally that adult T. infestans showed evidences of phenotypic plasticity in head size.

Guarneri et al. (2000) showed in four triatomines species of the genus Triatoma [T. infestans, Triatoma brasiliensis (Neiva, 1911), Triatoma sordida (Stål, 1859), and Triatoma pseudomaculata (Erichson, 1848)] that the development time was on average almost twice as long when fed on pigeons (279 days) than when fed on mice (159 days). Similarly, total nymphal mortality (as %) was 1.6 times larger when fed on pigeons (24.2%) than when fed on mice (14.6%). However, the average total body length, for both sexes together, did not differ when fed on pigeons (2.25 cm) than when fed on mice (2.16 cm) for the pool of all four species.

Martínez-Ibarra et al. (2003) tested the response of development time of the triatomine Triatoma picturata (Usinger, 1939) to being fed on hens and rabbits and, assuming normality, I carried out a sample t-test of their results (mean = 196.8 days; SD = 15.8; n1 = 69 for hens, and mean = 189.6 days; SD = 22.88; n2 = 68, for rabbits) showing that there was a statistically significant effect of the blood source used for feeding (t = 2.1459, df = 135, p-value = 0.03367). These differences, coupled with survival differences between feeding types led to Martínez-Ibarra et al. (2003) to show that there was also an effect on demographic parameters: the net reproductive rate (R0, estimated as the number of females produced by cohort × mean eggs laid per female divided by the number of eggs that started the cohort) was larger for rabbits (R0 = 148.5) than for hens (R0 = 114.7).

Cortico Correa Rodrigues et al. (2007) developed a life table of the triatomines species Triatoma tibiamaculata (Pinto, 1926), and estimated the longevity of adults and oviposition of females, and found that although there was no difference in the survival from the egg to the adult stage for insects fed on pigeon or on cocks, there was a large difference in the life-time fecundity: 7,832 eggs were laid by the cohort fed on cocks, and about half that number (3,633 eggs) was laid when the cohort was fed on pigeons.

In summary, considering the richness and variety of information on triatomines, I set as a goal for this article the compilation of the quantitative data available to investigate the possible effects of morphology, life cycle, and environment factors on the population fitness in triatomines, through a Machine Learning approach.

I carried out an exhaustive review of the literature on triatomines using the BibTri bibliographic database¹, hosted by the Centre for the Study of Parasitology and Vectors of the National University of La Plata, Argentina. This access-free database has 14,164 bibliographic records on Chagas disease, of which 7,207 refer exclusively to all possible features of the triatomines. Those records were then narrowed down by filtering them with the following 15 keywords available as a special field in the database: life cycle, population growth, demography, density effects, diet and foraging, physiology, age, fecundity, longevity, mortality, population parameters, survival, life tables, temperature, and development time. By this procedure I obtained 2,968 records complying with one or more of those 15 fields of interest (see Supplementary Material 1 for the complete list of those 2,968 publications); I analyzed them individually, and the publications that did not have the data in quantitative form as needed for the objective of this article were discarded, and 171 papers remained as adequate for the present analysis (see Supplementary Material 2 for a list of those 171 publications). I would like to highlight that all these publications provided results of laboratory experiments (see section “Discussion” about the importance of this feature).

From the final selection of 171 papers a total of 71 triatomine species were found, representing 539 different case studies (many papers had more than one triatomine species under study, and frequently for any given species more than one experiment under different laboratory conditions were carried out, resulting in many more cases than papers). The full information of those 539 study cases were transcribed to an Excel sheet that had 60 headers (see Supplementary Material 3 for the complete list of those headers), that represented the following seven types of information (N refers to “nymph”):

A relatively important variable effecting life history traits and fitness is infection of triatomines by T. cruzi (Cordero-Montoya et al., 2016), who found that insect hosts infection affects size, number of eggs laid, egg hatching rate, hatching success, and survival. However, despite this study is one of the most relevant to fitness estimation, some life history traits were not in the form adequate to estimate demographic parameters, and additionally development time was not estimated (all experiments were carried out with adult individuals). So, the T. cruzi infection factor could not be included in this analysis.

As I was interested in finding a possible association between life cycle, morphological and environmental variables with some fitness parameter among the triatomine species, I selected the population growth rate as demographic parameter (r₀ and/or R0, the population intrinsic rate of natural increase, and the net reproductive value, respectively) that are usually considered good indicators of population fitness (Engen and Saether, 2016), and that have been used as such in other triatomine studies (e.g., Pelosse et al., 2013). So, I had to filter again the dataset for those cases that provided some of the demographic parameters directly, or that at least provided enough life cycle information to calculate them. Whenever the demographic parameters were not available and I had to calculate r₀ and/or R0 from the original data in a publication, I used the Lotka equation (∑i=1ωlx⁢mx⁢e-r0*x) (Southwood, 1978), which requires the survival (l_x) and fecundity (m_x) schedules (a quantitative information rarely published); here x is the age (or stage) of the insect, and w is the last reproductive age. Lotka’s equation cannot be solved explicitly so the iterative method used by Rabinovich (1980) was applied, implemented in an Excel spreadsheet using the Solver tool. Both growth rate parameters are related by the formula r0=l⁢o⁢ge⁢RoT (Bengstron, 1969), where T is the cohort generation time (mean age of the females in the cohort at the birth of the female offspring). As both demographic parameters are directly related, I decided to use only one of them: the intrinsic rate of natural increase, r₀. This new requirement of the r₀ parameter (or l_x and m_x for its calculation), reduced the original 539 case studies and 71 triatomine species to 90 cases and 36 species.

Several of the 60 variables of the original dataset (see Supplementary Material 3) were redundant or unnecessary because they were included indirectly in other variables (e.g., total of egg to adult development time included the development time of each of the egg, NI, NII, NIII, NIV, and NV stages). Using a parsimonious criterium I selected (in addition to the name of each triatomine species) a subset of the variables available; e.g., the head width and pronotum length were the morphological variables that best differentiated species (Nattero et al., 2017); as a result, the following 11 variables remained in the final dataset to be used in the present analysis:

Although it is well known that population growth rate is affected by fecundity, survival, and development time (Southwood, 1978), both fecundity and development time were included in the analysis (unfortunately not enough data on survival was available) because it would be very informative to learn about the relative importance of those two variables under the variety of species and environmental conditions of the dataset here used. The complete dataset used in the present analysis, plus the triatomine species names, is given in the Appendix.

First an exploratory analysis was carried out, using classical box-plot graphics and normal fitting procedures, using the ggplot2 package of R software (Wickham, 2016). Then I selected the Random Forest analysis, using the randomForest (Breiman and Cutler, 2018) and the randomForestExplainer (Paluszynska et al., 2020) packages of the R library (R Core Team, 2019); these packages work by aggregating the predictions made by multiple decision trees of varying depth, with every decision tree in the forest being trained on a subset of the dataset (the bootstrapped dataset). Finally, to verify if the insights gained from the machine learning approach were robust, I compared those results with an ordinary least square (OLS) regression. All analyses were carried out with different packages of the R program (R Core Team, 2019). The R code used for the analyses is given in Supplementary Material 4.

Supplementary Figure 1 shows the distribution of the dependent variable used in this analysis (r₀, the intrinsic rate of natural increase). The distribution fits well a normal distribution (the Shapiro test resulted in a p-value = 0.659, so there is no evidence on which to reject the null hypothesis that the variable r₀ is normally distributed). As some species had several estimates of r₀ (sometimes under different environmental conditions), it was also of interest to visualize the distribution of this parameter among species (Supplementary Figure 2).

Similar box-plot graphs were prepared with the two numerical environmental variables: mean temperature and feeding frequency (Supplementary Figure 3); both variables show an important and useful degree of variation: the range of the mean temperature was 19.5–32°C, and that of the feeding frequency was 2–25 days (see black dots in Supplementary Figure 3, that represent the extreme values).

The two life cycle variables used in this analysis (egg to adult development time in days, and fecundity, expressed as average female eggs produced per female per day) show similar variability to the numerical environmental variables (Supplementary Figure 4); the range of the egg to adult development time was 69–570 days, and that of the fecundity was 0.17–2.42 female eggs/female/day (the black dots in Supplementary Figure 4, represent the extreme values).

Finally, the four morphological variables were also graphed (Supplementary Figure 5) in the form of box-plots in a single figure (with morphological traits in the x-axis), because each species has only one measurement for each morphological trait; so, in Supplementary Figure 5 the values of the morphological traits represent the variability among and not within species.

Why did I select Random Forest for analysis? Regression and classification trees have a very different flavor from the more classical approaches for regression and classification but, unfortunately, trees generally do not have the same level of predictive accuracy as some of the other regression and classification approaches (Gareth et al., 2014). However, methods such as Bagging, Random Forests, and Boosting use trees as building blocks to construct more powerful prediction models. Random forests provide some improvements over bagged trees by a random “de-correlation” of trees: when building the decision trees, each time a split in a tree is considered, a random sample of m predictors is chosen as split candidates from the full set of p predictors; then that split is allowed to use only one of those m predictors, and a fresh sample of m predictors is taken at each split (technically, the usual procedure is to choose m⁢p, that is, the number of predictors considered at each split is approximately equal to the square root of the total number of predictors; Gareth et al., 2014). In other words, in building a Random Forest, at each split in the tree, the algorithm is not even allowed to consider a majority of the available predictors.

Before running the Random Forest analysis, I checked for possible collinearity among some of the independent variables, to avoid any degree of strong redundancy. Using the classical criterium of r > | 0.7|, the only significantly correlated variables were Total length and Pronotum length (r = 0.846); I decided to delete Total length from the dataset because the Pronotum length, being a completely chitinized structure, is more reliable.

To run the standard Random Forest method with the dataset (see Appendix for the full dataset), I used 1,000 regression trees (the number of bootstrap samples), and r₀ (population intrinsic rate of natural increase) as the dependent variable; I first tried with several values of the parameter mtry (the number of predictors that are randomly sampled as candidates for each split), and selected the mtry value that maximized the percent of the explained variance; as a result I obtained a mean of squared residual of 1.6378e^–05, and a variance explained of 67.6% with mtry = 6. I believe that a variance explained of 67.6% is a sensible prediction accuracy.

Figure 1 shows the minimal depth among the trees of the Random Forest applied to the triatomine dataset; the visualization of minimal depth among trees provides an estimate of the important of each of the variables in the model. There are three “dominant” variables: Species, mean egg to adult development time, and mean female eggs per female per day; these three variables are the significant predictors given by the “importance measure” function of the Random Forest package. It is illustrative to look not only at the mean but also at the whole distribution of the minimal depths as given in Figure 1, that suggests more insights into the role that a predictor plays in a forest.

Two additional common measures of variable importance are given by the indices IncNodePurity (the total decrease in node impurities from splitting on the variable, averaged over all trees; i.e., it indicate that a node contains predominantly observations from a single class), and %IncMSE (the increase in the mean square error of predictions as a result of the variables being permuted, i.e., it indicates the extent to which the predicted response value for a given observation is close to the true response value for that observation). mse_increase is measured by the well-kwon Gini index; as mse_increase is the most robust and informative measure of importance in Random Forest (Gareth et al., 2014), while IncNodePurity is a biased measure, I preferred to present the results of variable importance as mse_increase (Figure 2); the higher the value of mse_increase associated with a predictor, the more important the role of the predictor in its effect on r₀.

Table 1 provides the statistical information resulting from the Random Forest analysis, and it is presented with the p-values in ascending order, confirming that only Species, mean egg to adult development time, and mean female eggs per female per day are the only significant predictors determining the population intrinsic rate of natural increase (r₀). It was relatively surprising that none of the environmental variables seem to play any role in determining r₀ (see section “Discussion”).

I have also produced the so-called a multi-way importance plot (see Supplementary Figure 6) using the mse_increase (in the x-axis), and the IncNodePurity (in the y-axis), and their location on the graph confirms that Species, mean egg to adult development time, and mean female eggs per female per day are the only significant variables.

The Random Forest package also allows for an estimation of interaction among the predictor variables, by graphing the mean minimal depth (MMD) of the main interactions (Supplementary Figure 7); from the MMD among the main six possible interactions the three interactions between Species and the two life cycle variables (mean egg to adult development time, and mean female eggs per female per day) are the dominant ones, although the interaction of those variables with the mean rearing temperature was also important; however, as mean rearing temperature was not statistically significant in its effect on r₀, that means that it is only important as a conditional effect on other variables.

As, in addition to species, the two life cycle variables keep showing to be the ones with the most significant influence on r₀, it was of interest to look at the empirical response of r₀ as a function of the simultaneous change of the mean egg to adult development time, and the mean female eggs per female per day; this is shown as a contour graph in Figure 3. It is clear that the former (mean egg to adult development time) has a stronger effect on r₀ than the mean female eggs per female per day.

As a final step, it was of interest to compare the results produced by the Random Forest method with the ones produced by a regular parametric analysis. I resorted to a simple OLS regression, and Table 2 shows the results of the regression analysis with the same dataset used for the Random Forest analysis, for all species pooled; it was confirmed that only the mean egg to adult development time, and fecundity (mean female eggs per female per day), are the only significant independent variables. However, when the linear regression was carried out with interactions among the independent variables, no effect was statistically significant (results not shown), an important difference with the Random Forest analysis (Supplementary Figure 7) and the observed data (Figure 3).