key: cord-0872965-wchxkdtj
authors: Lofstedt, Jeanne; Dohoo, Ian R.; Duizer, Glen
title: Model to Predict Septicemia in Diarrheic Calves
date: 2008-06-28
journal: J Vet Intern Med
DOI: 10.1111/j.1939-1676.1999.tb01134.x
sha: c2d753091e830bacade4f77ff796ff5962d2f383
doc_id: 872965
cord_uid: wchxkdtj

The difficulty in distinguishing between septicemic and nonsepticemic diarrheic calves prompted a study of variables to predict septicemia in diarrheic calves, 28 days old that were presented to a referral institution. The prevalence of septicemia in the study population was 31%. Variables whose values were significantly different (P < .10) between septicemic and nonsepticemic diarrheic calves were selected using stepwise, forward, and backward logistic regression. Variables identified as potentially useful predictors were used in the final model‐building process. Two final models were selected: 1 based on all possible types of predictors (laboratory model) and 1 based only on demographic data and physical examination results (clinical model). In the laboratory model, 5 variables retained significance: serum creatinine > 5.66 mg/dL (> 500 μmol/L) (odds ratio [OR] = 8.63, P = .021), toxic changes in neutrophils ≥ 21 (OR = 2.88, P = .026), failure of passive transfer (OR = 2.72, P = .023), presence of focal infection (OR = 2.68, P = .024), and poor suckle reflex (OR = 4.10, P = .019). Four variables retained significance in the clinical model: age ≤ 5 days (OR = 2.58, P = .006), presence of focal infection (OR = 2.45, P = .006), recumbency (OR = 2.98, P = .011), and absence of a suckling reflex (OR = 3.03, P = .031). The Hosmer—Lemeshow goodness‐of‐fit chi‐square statistics for the laboratory and clinical models had P‐values of .72 and .37, respectively, indicating that the models fit the observed data reasonably well. The laboratory model outperformed the clinical model by a small margin at a predictabilty cutoff of 0.5, however, the predictive abilities of the 2 models were quite similar. The low sensitivities (39% and 40%) of both models at a predicted probability cutoff of 0.5 meant many septicemic calves were not being detected by the models. The specificity of both models at a predicted probability cutoff of 0.5 was .90%, indicating that .90% of nonsepticemic calves would be predicted to be nonsepticemic by the 2 models. The positive and negative predictive values of the models were 66–82%, which indicated the proportion of cases for which a predictive result would be correct in a population with a prevalence of septicemia of 31%.

T he mortality risk of live-born neonatal calves Ͻ1 month of age has been reported to range from 15 to 30%. [1] [2] [3] [4] The majority of deaths are attributable to infectious diseases; diarrhea, pneumonia, and septicemia are the most common. [5] [6] [7] [8] The predominant pathogen cultured from calves with septicemia is Eschericia coli, but other gramnegative, gram-positive, and mixed bacterial infections have been documented. [9] [10] [11] [12] Important risk factors for the development of septicemia in calves include decreased passive transfer of colostral immunoglobulins and exposure to invasive bacterial serotypes. 13 Neonatal diarrhea may also predispose calves to septicemia. In 2 separate studies conducted on a California veal operation, blood cultures revealed bacteremia in 31% and 19% of calves with signs of diarrhea, depression, and/ or weakness. 10, 14 Septicemia in these calves was attributed to intestinal mucosal damage caused by bacterial, viral, or parasitic gastrointestinal infections, which allowed opportunistic gut pathogens to enter the systemic circulation. 10 The early signs of septicemia in neonatal foals and calves are vague and nonspecific and are often indistinguishable from signs of noninfectious diseases or those of focal in-fections such as diarrhea. 14, 15 Positive blood cultures are required for a definitive antemortem diagnosis of septicemia, but results are not reported for 48-72 hours, and false-negative culture findings are common. 12, 16 No single laboratory test has emerged as being completely reliable for the early diagnosis of septicemia in farm animal neonates, 12, 17 therefore, various scoring systems and predictive models using easily obtainable historical, clinical, and clinicopathologic data have been developed for this purpose. 14, [17] [18] [19] The goal of these mathematical models is to identify septicemic neonates early in the course of disease when appropriate therapeutic intervention would most likely result in a favorable outcome.

For a period of time, routine blood cultures were performed on all diarrheic calves presented to the Atlantic Veterinary College Teaching Hospital regardless of whether the clinical or clinicopathologic findings indicated a diagnosis of septicemia. Results of this exercise indicated that septicemia was more common in diarrheic neonatal calves than we had anticipated and that it increased the cost of treatment while decreasing the prognosis for survival. A model capable of predicting sepsis in diarrheic neonatal calves presented to the hospital for treatment would assist the clinicians in distinguishing between calves with undifferentiated diarrhea and calves with diarrhea complicated by sepsis. In the case of a relatively low-value calf with diarrhea and predicted sepsis, the farmer may be advised not to initiate treatment because of the expense involved and the poor prognosis. In contrast, the owner of a valuable diarrheic calf that was predicted to be septic could be informed that early initiation of costly but appropriate antimicrobial and supportive therapy may result in an improved outcome.

Models to predict septicemia in calves have been published. 14, 19 Study populations in these reports consisted of veal calves, 1 day to 4 months of age, with signs of serious illness (diarrhea, depression, and/or weakness) 14 and calves Ͻ30 days of age that were presented to a veterinary teach- 

Medical records of calves Ͻ28 days of age presented to the Atlantic Veterinary College Teaching Hospital between the years 1989 and 1993 with a primary complaint of diarrhea were retrieved. Pretreatment data were extracted from the medical records of these calves. Independent variables included demographic information and physical examination findings and clinicopathologic values for hematology, venous blood gases, serum chemistry, and immunoglobulins (Table 1) . Immunoglobulin concentration in serum was determined quantitatively or qualitatively using a variety of procedures, including the quantitative zinc sulfate turbidity test (69 calves), 20 sodium sulfite precipitation test (17 calves), gluteraldehyde coagulation test (62 calves), and radial immunodiffusion test (5 calves). Failure of passive transfer of immunoglobulins was defined as an IgG concentration of Յ800 mg/dL by any of the above mentioned tests, a globulin concentration of Յ2 g/ dL (Յ20 g/L), or total serum protein of Յ5 g/dL (Յ50 g/L) ( Table  1 ). Other information extracted from the medical records of the calves in the study population were fecal culture, fecal electron microscopic examination, and fecal floatation results.

Logistic regression assumes that the relationship between any independent variable recorded on a continuous scale (eg, respiratory rate) has a linear relationship with the logit of the probability of sepsis. For each such independent variable, this assumption was assessed by evaluating the distribution of the independent variable, categorizing the independent variable in 2 or more groups and cross-tabulating the new variable with sepsis, or running unconditional logistic regressions with the variable in question and evaluating the Hosmer-Lemeshow goodness-of-fit statistics. The most appropriate form for each independent variable was then chosen (Table 1) . Categorical variables were all entered into the predictive models as a series of dummy variables.

The dependent variable of interest was a diagnosis of septicemia. A diagnosis of septicemia in diarrheic calves was based on the following antemortem criteria: (1) positive blood culture, (2) culture of the same bacterial agent from Ն2 body fluids, or (3) culture of a bacterial agent from a single joint in a calf with joint effusion involving multiple joints. Blood for culture was collected aseptically from the jugular vein and inoculated into the Oxoid Signal blood culturing system (Oxoid Canada, Inc, Nepean, ON, Canada). A number of criteria contributed to a postmortem diagnosis of septicemia, including (1) morphologic changes such as multiple disseminated abscesses of similar size, purulent vasculitis and intravascular identification of bacteria, or fibrin in multiple body cavities; (2) bacterial isolation from heart blood; or (3) recovery of the same bacterial organism from Ն2 body tissues (excluding intestine). Other outcome variables recorded were days hospitalized and calf survival.

Means for continuous variables were compared between septicemic and nonsepticemic calves using Student's t-test. 21 Categorical variables were analyzed using contingency table chi-square analysis. 21 Variables that were significantly different (P Ͻ .10) between septicemic and nonsepticemic calves were grouped into 1 of 3 categories: physical variables, hematological variables, and chemistry variables, which included measures of passive transfer. Variables in each group were selected for logistic regression models using stepwise, forward, and backward selection procedures. 22 Variables identified as possibly useful predictors by the various selection procedures were used in the final model building process. This process involved comparing a number of possible models and selecting those that maximized the area under the receiver operating characteristic (ROC) curve 23 but that did not include any variables whose coefficients were not significant (P Ͻ .05). Two final models were selected: one based on all possible types of predictors (the laboratory model) and one based only on demographic data and physical examination results (the clinical model).

Odds ratios (OR) derived from the models were interpreted as measures of increased risk of disease. Because septicemia was not a rare condition, the OR were not precise measures of increased risk, but the approximation was reasonable and was utilized to clarify the presentation of results.

Once final models were selected, their sensitivity and specificity were determined at predictive probability cutoff points of 0.5 and 0.3 (the points that roughly balanced the sensitivity and the specificity). The goodness of fit of the models was evaluated using the Hosmer-Lemeshow goodness-of-fit chi-square statistic 22 with the data divided into 5 groups. For each model, the Pearson residuals, the standardized Pearson residuals, and the ⌬␤ values were computed for all covariate patterns to determine if any specific covariate pattern had an undue influence on the model.

Because evaluating the fit of a model using the same data that were used to build the model is likely to overestimate the predictive ability of the model, a split sample goodness-of-fit evaluation was also carried out. The data were randomly (using a computer based random number generator) divided into 2 separate databases with 60% and 40% of the observations in each. The 2 models were refit using the database containing 60% of the observations by forcing in all of the independent variables included in the 2 final models described above. The predictive ability of the resulting models was assessed using the database containing 40% of the observations.

Analyses were performed on a personal computer using the statistical software package STATA (Stata Corp, College Station, TX).

Two hundred fifty-four calves met the criteria for inclusion in the study. However, 2 calves were excluded from analysis because they could not be classified as either septicemic or nonsepticemic based on criteria established for the study; Campylobacter-like organisms were seen on Gram stains of their blood, but the organisms were not recovered from blood culture.

The mean age of diarrheic calves at presentation was 9.4 days. Crossbred beef calves (32.9%) and Holstein calves 

Seventy-eight (31%) of the calves met the study criteria for a diagnosis of septicemia. Positive blood cultures were used to diagnose 50 cases (40.7% of all blood cultures were positive). A single organism was cultured from most calves but 3 of the blood cultures yielded 2 isolates each. E coli accounted for 29 (55%) of the isolates from blood. An additional 10 bacterial agents were cultured from the blood, including Campylobacter spp. (8) (1), and Clostridium spp. (1) . Culture of the same pathogen from Ն2 body fluids antemortem were used to diagnose 6 cases, and a positive antemortem joint culture from a calf with multiple enlarged joints was used to diagnose 1 case. The same bacterial agent was cultured from Ն2 tissues at necropsy in 33 calves, postmortem heart blood culture was positive in 1 calf, and morphologic lesions such as multiple abcessation, purulent vasculitis, intravascular identification of bacteria, and fibrin exudation in multiple body cavities, were present in 18 calves. E coli was also the most frequent isolate in antemortem body fluid and postmortem tissue samples, accounting for 7/11 (64%) and 93/154 (60%) isolates, respectively. Because individual calves often satisfied multiple criteria for a diagnosis of septicemia, the totals for all the above criteria are greater than the total number of septicemic calves.

One hundred forty-one of the 174 nonsepticemic calves (81%) were discharged from the hospital (survived) compared with only 23 of 78 septicemic calves (29.5%). The mean hospital stay for nonsepticemic calves that lived was 4.1 days compared with 5.0 days for septicemic calves that survived.

Variables identified as potentially useful predictors (P Ͻ .1) are described in Table 2 . Septicemic calves had significantly higher (P Ͻ .05) values for respiratory rate, packed cell volume, band neutrophil count, and venous PCO 2 than did nonsepticemic calves. Mean values for rectal temperature and total plasma and serum protein, globulin, calcium, and glucose concentrations were significantly lower (P Ͻ .05) in septicemic than in nonsepticemic calves. A significantly larger (P Ͻ .05) proportion of septicemic calves was Յ5 days of age, unresponsive or comatose, and in sternal or lateral recumbency at presentation when compared with nonsepticemic calves. The proportion of septicemic calves exhibiting a weak or absent suckling reflex, scleral injection, and hyperemic or cyanotic oral mucous membranes was also significantly greater (P Ͻ .05). Elevated serum creatinine concentration, toxic changes in neutrophils Ն2ϩ, and failure of passive transfer of colostral immunoglobulins were present more frequently in septicemic than in nonsepticemic calves (P Ͻ .05).

Two final models, the laboratory model and the clinical model, were selected based on the maximal area under the ROC curve (Table 3) . Variables included in the laboratory model were serum creatinine concentration, toxic changes in neutrophils Ն2ϩ, failure of passive transfer, focal infection, and a poor suckle reflex. The clinical model identified age of Յ5 days, focal infection, recumbency, and a poor suckle reflex as predictors of septicemia. The laboratory model indicated that calves with moderately increased serum creatinine concentration were twice as likely to be septicemic (OR ϭ 2.07; 95% CI ϭ 0.836-5.119), whereas those with creatinine concentration of Ͼ5.66 mg/dL (Ͼ500 mol/L) were 8 times more likely to be septicemic (OR ϭ 8.63; 95% CI ϭ 1.383-53.827). Similarly, evidence of toxic changes in their neutrophils (Ն2ϩ), failure of passive transfer, or evidence of focal infections all increased the risk of septicemia by 2.68-2.88 times. Calves with a poor suckle reflex were 4 times as likely to be septic. The clinical model (ie, model containing only data from the physical examination and history) showed that each of the following was a factor that increased the risk of being septicemic by approximately 2.5-3-fold: Ͻ5 days of age, signs of focal infection, recumbent on admission, and poor suckle reflex.

The Hosmer-Lemeshow goodness-of-fit chi-square statistics for the laboratory and clinical models had P-values of 0.72 and 0.37, respectively, indicating that the models fit the observed data reasonably well. Evaluation of residual patterns did not identify any outlining observations, and no specific group of calves (ie, calves with a common covariate pattern) exerted a large influence on the model. We concluded that the models fit the data reasonably well.

The predictive abilities of the models are presented in Table 4 . When evaluated at a cutoff value of 0.5 (ie, a calf was predicted to be septicemic if the predicted probability was Ն0.5), all models had quite low sensitivity and very high specificity. A cutoff of 0.3 roughly balanced the sensitivity and specificity and has been presented for comparison purposes. The laboratory models outperformed the clinical models, but only by a small margin. Similarly, the models based on the full data set outperformed those in which the model was constructed on 60% of the data and its predictive ability was evaluated in the other 40%. However, the differences were small.

Overall, using a probability cutoff of 0.5 to define predicted sepsis, the model sensitivities ranged from 25 to 40% and the specificities ranged from 90 to 95%. The low sensitivities mean that many septicemic calves would not be detected by the model. However, the positive and negative predictive values ranged from 66 to 82%, which indicates the proportion of cases in which the predicted result was correct (for a population with a prevalence of septicemia of 31%).

At a cutoff of 0.3, the models had sensitivities and specificities in the 70-75% range, which means that a much higher proportion of septicemic calves would have been detected by the model. However, a positive prediction was only associated with a 52-68% probability (positive predictive value) that the calf was truly septic. However, a negative prediction was associated with an 83-89% probability (negative predictive value) that the calf was not septic.

Thirty-one percent of the diarrheic calves in this study were diagnosed with septicemia based on results of blood cultures, antemortem or postmortem tissue or fluid cultures, and morphologic changes at postmortem. The prevalence of septicemia in this study was identical to that reported for calves with diarrhea, depression, and/or weakness on a veal raising facility, 14 which suggests that the predictive values of the models developed herein may be relevant to other calf populations. The risk of contamination of the blood sample for culturing probably was low, so it is unlikely that many truly nonsepticemic calves were misclassified as septicemic. However, because of the low sensitivity of blood cultures 24, 25 and the fact that blood was not cultured for some surviving calves, some truly septicemic calves may have been misclassified as nonsepticemic if they survived and were not subjected to postmortem examination. However, this misclassification probably was not common because (1) the calves survived, and the probability of survival was low if the calf was truly septicemic; and (2) the effect of misclassification of truly septicemic calves on the model would likely been a reduction in specificity, but low specificity was not a problem with either model. E coli was the bacterial agent cultured with greatest frequency from septicemic calves in this study, a finding that agrees with results of previous investigations 10, 12 and may be attributed to the fact that certain strains of E coli possess virulence factors that promote systemic invasion. 26 In addition to E coli, a variety of noncoliform bacteria were isolated. Of particular interest was the recovery of Campylobacter fetus subsp. fetus from blood of 8 diarrheic calves (15% of the isolates from blood); there is only 1 previous report of this agent being cultured from a septicemic calf. 10 Possible explanations for the frequent isolation of Campylobacter spp. in this study were that the Oxoid Signal blood culture system was more effective than other blood culture systems for growing this agent or that there were unique predisposing factors leading to Campylobacter bacteremia in calves in our geographic location.

Two models, the laboratory model and the clinical model, were ultimately selected to predict septicemia in diarrheic neonatal calves presented for treatment. The laboratory model, which included all possible variables, was intended for use in a hospital setting where the clinician would have access to laboratory facilities. The clinical model, which contained only demographic and physical examination variables, was developed for use in the field.

Clinicopathologic parameters identified by the laboratory model as being associated with an increased risk of septicemia were moderate (1.99-5.66 mg/dL) (176-500 mol/ L) and marked (Ͼ5.66 mg/dL) (Ͼ500 mol/L) increases in serum creatinine concentration, moderate to marked toxic changes in neutrophils (Ն2ϩ), and failure of passive transfer (IgG concentration Յ 800 mg/dL, globulin Յ 2 g/dL [Յ20 g/L]), and total serum protein Յ 5 g/dL (Յ50 g/L). Because Azotemia secondary to shock and dehydration was anticipated in both septicemic and nonsepticemic diarrheic calves presented for treatment, it was interesting to note that moderate and marked increases in serum creatinine concentration in this study increased the risk of a calf being septicemic by 2-and 8-fold, respectively. Moderate to severe azotemia in septicemic calves was attributed to endotoxin induced tubular necrosis and embolic nephritis, in addition to decreased renal perfusion associated with dehydration. Toxic changes in neutrophils, which are induced by inflammatory mediators and endotoxins, are frequently encountered in septicemic calves 12 ; therefore, inclusion of this variable in the laboratory model was not unexpected. Because the relationship between septicemia and failure of colostral antibody transfer has been documented in previous studies, 12,14 the increased risk of septicemia in calves with failure of passive transfer was also anticipated. Both the laboratory model and the clinical model indicated that the risk of septicemia in calves with identified sites of focal infection (omphalitis, arthritis, meningitis, uveitis) was more than double that of calves without evidence of focal infection. This association was biologically plausible because one of the initial sites of bacterial invasion in septicemic calves is the umbilicus, 13 and hematogenous spread of infection to the meninges, synovial lining, and uveal tract are known to occur. 12 Inclusion of poor suckling reflex by both models and recumbency by the clinical model as predictors of septicemia in a population of diarrheic calves presented for treatment was unforseen. The fluid, blood gas, and electrolyte derrangements associated with the diarrhea could have caused recumbency and disinterest in suckling in the majority of calves in the study. However, multiple organ invasion by bacterial pathogens and the cascade of inflammatory reactions initiated by these infecting agents were more likely responsible for the increased frequency of recumbency and poor suckling reflex in septicemic calves. The clinical model indicated that being Ͻ5 days of age increased the risk of septicemia by 2.5fold. This contrasts with results of a study conducted in veal calves where age Ͼ7 days significantly increased the risk of being blood culture positive. 14 The reason for this discrepancy between the 2 studies is unclear. Because most calves in the present study were presented for treatment in the winter months when all calves presented to the hospital would be born inside and there is increased stocking density in maternity pens, they could have been heavily exposed to pathogens early in life. The theory that enteritis predisposes calves Ͼ1 week of age to septicemia 10 was not supported by results of this study.

Although the laboratory model outperformed the clinical model by a small margin at a probability cutoff of 0.5, the predictive abilities of the 2 models were remarkably similar (Table 4) . Thus, the practitioner in the field, without immediate access to a laboratory, could use the clinical model to predict septicemia in a diarrheic calf with almost the same accuracy as that of the clinician at a referral institution, who would be using the laboratory model for the same purpose. The low sensitivities (39% and 40%) of both mod- els at a probability cutoff of 0.5 meant that many septicemic calves were not being detected by the models. One explanation is that other useful predictors of septicemia were not included in the models. However, with the exception of historical data pertaining to calving and management practices, most variables previously identified as risk factors in calves and foals were evaluated. 14, 17, 18 There may be no group of variables unique to the septicemic calf, partly because septicemia is a dynamic process and laboratory and clinical parameters can vary widely depending on when in the disease course the calf is evaluated. The specificity of both models at a cutoff of 0.5 was Ͼ90%, indicating that Ͼ90% of nonsepticemic calves would be predicted to be nonsepticemic by the 2 models.

When the predictive probability cutoff value was set to 0.5, the laboratory model had a positive predictive value of 76%. This cutoff could be used when the clinician wants to be relatively certain that a calf predicted to be septicemic would, in fact, be septicemic. For example, when dealing with a relatively low-value calf, a decision to euthanize the calf may be made given the poor prognosis for septicemic calves. In this situation, a high positive predictive value would be desirable.

When the predictive probability cutoff was set to 0.3, the positive predictive value decreased but the negative predictive value rose to 89%. Consequently, a clinician could be relatively certain that a calf that tested negative was, in fact, nonsepticemic. This would be a desirable situation when treating a valuable calf with diarrhea in which ancillary treatments for septicemia (broad spectrum antimicrobials and plasma) would only be omitted from the treatment plan if the clinician was relatively confident that the calf was nonsepticemic.

The predictive values for any test depend on the sensitivity and the specificity of the test and the prevalence of the disease. All of the predictive values discussed above are based on the assumption that the prevalence of septicemia was 31%. If the prevalence of septicemia in a population of calves presented for treatment were Ͻ31%, the negative predictive value of the test would rise but the positive predictive value would drop off dramatically (Fig 1) .

The relatively small reduction in the predictive ability of the models when they were evaluated using a split sample approach was encouraging and suggests that the models may perform reasonably well in predicting future observations. However, there may still be a slight upward bias in the assessment of the performance of these models because the variables chosen for inclusion in the split sample assessment were those obtained from the analysis of the whole data set.

In conclusion, the predictive models fit the observed data reasonably well and had moderate predictive ability. The main limitation in the models seemed to be a low sensitivity. Many of the septicemic calves may have been early in the course of the disease and may not have had any distinguishing features that could be used to identify them as septicemic. Reassessment of these calves as the disease progresses may clarify their septicemia status but would not help in making initial decisions about therapy.

Calf mortality

Management factors associated with calf mortality in South Carolina dairy herds

Dairy calf mortality rate: Influence of management and housing factors on calf mortality rate in Tulare County, California

Factors influencing dairy calf mortality in Michigan

Descriptive epidemiology of calfhood morbidity and mortality in New York Holstein herds

Factors affecting susceptibility of calves to disease

An epidemiologic study of calf health and performance in Norwegian dairy herds. III. Morbidity and performance: Literature review

Dairy calf management, morbidity and mortality in Ontario Holstein herds. III. Association of management with morbidity

Blood culture from calves and foals

Bacteriological culture of ill neonatal calves

Pasteurella multocida septicaemia in two calves

Neonatal septicemia in calves: 25 cases (1985-1990)

Septicaemic colibacillosis and failure of passive transfer of colostral immunoglobulin in calves

Use of a clinical sepsis score for predicting bacteremia in neonatal dairy calves on a calf rearing farm

Brewer BD, Koterba AM. Bacterial isolates and susceptibility patterns in foals in a neonatal intensive care unit

Development of a scoring system for the early diagnosis of equine neonatal sepsis

Comparison of empirically developed sepsis score with a computer generated and weighted scoring system for the identification of sepsis in the equine neonate

Bovine neonatal sepsis score

Quantitation of bovine immunoglobulins: Comparison of single radial immunodiffusion, zinc sulfate turbidity, serum electrophoresis, and refractometer methods

Primer of Biostatistics

Applied Logistic Regression

Basic principles of ROC analysis

Bacteremia: Pathogenesis and diagnosis

Comparison of bacteriological culture of blood and necropsy specimens for determination of the cause of foal septicemia: 47 cases (1978-1987)

Escherichia coli from calves with bacteremia