Prediction of Mortality in Acute Renal Failure in Tropics

Abstract

Despite significant improvements in medical care, acute renal failure (ARF) remains a high risk for mortality. It is important to be able to predict the outcome in these patients in view of the emotional and ethical needs of the patients and to address questions of efficiency and quality of care. We analyzed the risk factors predicting mortality prospectively in a group of 265 patients using univariate and multiple logistic regression analysis. A prognostic model was evolved that included 10 variables. The model showed good discrimination [(receiver operating characteristic (ROC) area = 0.91) and correctly classified 88.30% of patients. The variables significantly associated with mortality were coma odds ratio (OR) = 9.8], oliguria (OR = 4.9), jaundice (OR = 3.7), hypotension (OR = 3.1), assisted ventilation (OR = 2.3), hospital acquired ARF (OR = 2.3), sepsis (OR = 2.2), and hypoalbuminemia (OR = 1.7). Age and male gender were included in the model as they are clinically important. The score was validated in the same sample by boot strapping. It was also validated in a prospective sample of 194 patients. The model was calibrated by the Hosmer-Lemeshow goodness-of-fit test. It was compared with two generic illness scores and one specific ARF score and was found to be superior to them. The model was verified in different subgroups of ARF like hospital acquired, community acquired, intensive care settings, nonintensive care settings, due to sepsis, due to nonsepsis etiologies, and showed good predictability and discrimination.

• acute renal failure
• risk factors
• prediction
• mortality
• India

INTRODUCTION

Acute renal failure (ARF) is a syndrome rather than a disease, which may have numerous causes and can be observed in a wide variety of clinical settings. The mortality associated with ARF is as high as 50% to 80% and has remained so for the last 50 years, despite the advances in critical care and dialytic technology. Increasingly, older patients are now developing ARF and patients with severe acute illnesses now survive long enough to develop ARF.

Early and individualized prediction of outcome of ARF is a crucial priority, in view of the devastating nature of the illness as well as emotional and ethical needs of the patients, relatives, and society. Prognostic scoring systems aid physicians in calculating outcomes in patient groups and individual patients, advising patients' families regarding continued life support, comparing institutions, and addressing questions of effectiveness, efficiency, and quality of care. A good probability model should have good discrimination, good calibration, and good performance in different subsets of patients. The study sample should be representative of the center; data collection should be complete. The model should show adaptation for the periodic variation in case mixand limited variability in prognostic risk within the population.[1]

Various scoring systems have been developed for making predictions of outcome especially in the intensive care unit. Generic illness severity systems (APACHE III,[2] SAPS,[3] SOFA.[4] were initially used to predict mortality in ARF, but they have not discriminated well in most published studies [5] A generic index will typically under perform in a population uniformly affected with ARF, as indicators of renal function are major factors in these scores. Hence, as ARF creates an additional risk of mortality, disease-specific scoring systems, namely Lohr,[6] Liano,[7] SHARF,[8] Schaefer,[9] and Mehta,[10] were developed.

All these models were developed in the setting of highly advanced critical care and dialysis facilities, which may not apply to a developing country, where etiologies of ARF are different, resources are scanty, and management is decided mostly by economic considerations. This study, from a tertiary referral center in southern India, has tried to evolve a prognostic model for predicting mortality in ARF by analyzing the risk factors significant to the etiologies and clinical scenario prevalent in the area.

AIMS

1. To analyze the risk factors in ARF and to evolve a prognostic model which uses mortality as endpoint, possesses a good fit, and can be easily used as a bedside tool.

2. To validate the model internally in the same sample and in a prospective population with ARF and to compare it with other existing models.

MATERIALS AND METHODS

Patients

This prospective study was done during the period of July 2002 to June 2003 at Christian Medical College, Vellore. The developmental sample for generation of the model was collected from July 2002 to December 2002, and the validation sample was collected from January 2003 to June 2003. Using a computer-generated search of serum creatinine from the hospital laboratory, patients with serum creatinine greater than 1.5 mg/dL were identified. From this group, patients with acute renal failure were identified using the definitions given later. Demographic, clinical profiles, and diagnosis were recorded in patients diagnosed with ARF. Risk factors considered important for the development of the model were recorded at 24 hours after admission or at 24 hours of developing renal failure in hospital-acquired ARF. All patients were followed up daily from the date of admission (for community-acquired ARF) or date of diagnosis of renal failure (for hospital-acquired ARF) until discharge or death. The consulting nephrologist made the decisions regarding management as well as dialysis. Chronic kidney disease patients with baseline serum creatinine more than 5 mg/dL and renal transplant recipients were excluded from the study.

Definitions

Acute renal failure was defined as an increase in serum creatinine level of 1) ≥ 0.5 mg/dL for patients with normal renal function or who had a baseline serum creatinine level of 1.9 mg/dL or less and 2) ≥ 1.0 mg/dL for patients with a baseline level between 2.0 to 4.9 mg/dL. In patients who were diagnosed with hospital-acquired renal failure, the lowest serum creatinine level before development of renal failure was used as the baseline value. The same criteria were applied for community-acquired renal failure[11] also, but baseline value was defined as the lowest creatinine value 2 months prior to admission if available. If this value was not available, the diagnosis was based on the clinical features and serum creatinine at recovery. Community-acquired ARF was defined as renal failure developing outside the hospital and diagnosed at hospital admission, whereas hospital-acquired ARF was defined as renal failure that developed during hospitalization in patients with normal renal function or preexisting stable chronic renal failure (baseline serum creatinine < 5 mg/dL). All the patients who developed ARF during their hospital stay were included in the hospital-acquired ARF group, irrespective of the primary illness, whether of renal or nonrenal etiology. Oliguria was defined as urine output less than 400 mL/24 hours. Sepsis[12] was defined as the systemic response to infection, manifested by two or more of the following conditions as a result of infection: 1) temperature greater than 38°C or less than 36°C, 2) heart rate greater than 90 beats/min, and 3) respiratory rate greater than 20 breaths/min or arterial CO₂ pressure less than 32 mm Hg, and 4) white blood cell count greater than 12,000/µL, less than 4000/µL, or greater than 10% immature (band) forms. Hypotension was defined as systolic blood pressure ≤ 90 mm Hg or the use of inotropic support. Hypoalbuminemia was defined as serum albumin < 3.5 mg/dL. Jaundice was defined as a serum bilirubinmore than 2 mg/dL. Coma was defined as a Glasgow coma scale < 5 in the absence of sedation.

STATISTICAL ANALYSIS

The initial developmental sample had 265 patients. Conventional parametric statistics were used for primary analysis of data. Mortality was considered as the dependent variable. Initially a diverse list of risk factors was tested for significance in predicting mortality by univariate analysis, using Student's t-test and chi-square. Those found significant along with clinically important risk factors were included in a multiple logistic regression analysis. Odds ratios were estimated for each factor found in the regression, and 95% confidence intervals were derived. With this model, the effect of any independent variable could be assessed while controlling the effect of all other independent variables and then the relative weight of each specific risk factor can be calculated. The model calculates the probability of death for each individual patient, through an equation, adding to a constant value the relative weight of each risk factor multiplied by each value.

After deriving the final model (Vellore model) from the developmental sample, it was validated first in the same sample by the technique of bootstrapping. Subsequently it was validated prospectively in a group of 194 patients. Discrimination of the model was assessed using the area under the receiver operating characteristic (ROC) curve.[13] The ROC curves show the relationship between sensitivity (correct identification of those who die) and 1.0 minus specificity (incorrect identification of those who survive). If the area under the ROC curve is 0.5 the model has no discriminatory power, and if the area is 1.0, the model discriminates perfectly. The Hosmer-Lemeshow goodness-of-fit test[14] was used to assess calibration. A nonsignificant value for Hosmer-Lemeshow chi-square suggests an absence of bias at extremes of risk. Likelihood ratio was calculated to assess overall model performance. The discrimination and model performance were compared with two representative generic models, APACHE III, SAPS2, and one specific ARF score, the Liano score. Statistical analysis was carried out by STATA 8.0 and SPSS 11.5 software.

RESULTS

A total of 60,568 patients were admitted to the hospital during the 1-year study period (July 2002 to June 2003). Four hundred and fifty-nine patients were diagnosed with ARF by criteria mentioned earlier of whom 265 consecutive patients (from July 2002 to December 2002) formed the developmental sample while 194 patients (from January 2003 to June 2003) formed the validation sample.

Characteristics of both the groups are given in Table 1. Both the samples were similar in characteristics like age, ARF in elderly (> 65 years age), preexisting chronic renal failure, cardiovascular disease, neoplastic disease, hypertension, community-acquired and hospital-acquired ARF, and requirement of dialysis. There were statistically significant differences in the proportion of patients with diabetes, oliguria, and sepsis, which was by chance alone.

Table 1. Comparison of basic characteristics of patients in developmental and validation samples.

Characteristic	Developmental	Validation	p value
	n = 265	n = 194
	n (%)	n (%)
Mean age (years)	47.9 ± 15.9	46.6 ± 16.1	0.4
Males	188 (71)	132(68)	0.5
Hospital-acquired ARF	109(41)	67(35)	0.2
Community-acquired ARF	156(59)	128(66)	0.1
Requiring dialysis	68(26)	63(32)	0.2
Preexisting renal failure	17(6)	19(10)	0.2
ICU patients	100(38)	84(43)	0.3
Hypertension	49(19)	30(16)	0.4
Elderly (Age > 65yrs)	76(28)	48(25)	0.5
Mortality	100(38)	68(35)	0.5
Surgical ARF	64(25)	44(23)	0.6
Hepatobiliary disease	25(9)	21(11)	0.6
Neoplastic disease	24(9)	18(9)	0.9
Oliguric ARF	77(29)	76(39)	0.02
Diabetes	71(27)	34(18)	0.02
Sepsis	5(29)	75(39)	0.02

The etiology of ARF in the sample studied is shown in Table 2. Compared to western studies there is a higher incidence of causes specific to the tropics such as malaria, leptospirosis, and snakebite. A significant proportion of patients had sepsis as etiology, which is in contrast with one previous study from northern India[15] in which diarrheal diseases were the predominant cause of ARF. This probably reflects the changing pattern of referral to tertiary care centers.

Table 2. Etiology of ARF.

Characteristic	Developmental	Validation
	n = 265	n = 194
	n (%)	n (%)
Medical
Sepsis	56(21.1)	50(25.8)
Diarrheal diseases	32(12.1)	25(12.8)
Malaria	20(7.5)	13(6.7)
Leptospirosis	17(6.4)	12(6.1)
Hepatorenal syndrome	11(4.2)	6(3.1)
Acute glomerulonephritis	10(3.7)	10(5.6)
Snakebite	11(4.2)	8(4.1)
Congestive cardiac failure	8(3.0)	5(2.6)
Rhabdomyolysis	2(0.7)	1(0.5)
Drug-induced	6(2.1)	3 (1.5)
Contrast-induced	4(1.4)	3(1.5)
Others	5(1.9)	12(1.0)
Surgical
Sepsis	25(7.9)	20(10.3)
Obstructive uropathy	16(6.3)	11(5.7)
Trauma	10(3.7)	7(3.6)
Drugs	8(3.0)	4(2.1)
Others	5(1.9)	2(1.0)
Obstetric
Sepsis	5(1.9)	4(2.1)
Preeclampsia	6(2.2)	5(2.6)
Hemorrhage	3(1.3)	3(1.5)
Others	2(0.7)	1(0.5)

Seventeen patients in the developmental sample and 19 patients in the validation sample had preexistingchronic renal failure. Of these, 10 patients in each group were admitted with community-acquired renal failure, causes of which were diarrheal diseases, infections, and congestive cardiac failure.

Univariate Analysis

Initial univariate analysis showed that oliguria, hypotension, jaundice, hypoalbuminemia, heart failure, serum creatinine, assisted ventilation, and coma were significant risk factors predicting mortality. Old age (defined as age greater than 65 years), male gender, preexisting renal failure, preexisting liver disease, diabetes, hypertension, and peripheral vascular disease were found to be not significant. The odds ratio and confidence intervals of the significant and nonsignificant risk factors are given in Table 3. Among the variables, serum creatinine and age in years were continuous variableswhile others were categorical in nature. Age in years as a continuous variable was found to be not significant. The variables were obtained at 24 hours after admission in community-acquired ARF and at 24 hours after diagnosis of ARF in hospital-acquired ARF.

Table 3. Univariate analysis of mortality.

Risk factor	OR	95% C.I.	p
Coma	68.5	23.5 to 200.2	< 0.001
Assisted ventilation	28.4	13.8 to 58.6	< 0.001
Hypotension	21.8	11.0 to 43.6	< 0.001
Jaundice	8.5	4.2 to 17.0	< 0.001
Hypoalbuminemia	7.8	4.4 to 13.7	< 0.001
Oliguria	6.1	3.4 to 10.8	< 0.001
Hospital-acquired ARF	4.4	2.6 to 7.5	< 0.001
Sepsis	3.4	1.9 to 5.9	< 0.001
Hepatobiliary disease	1.6	1.1 to 2.3	0.06
Age	1.0	0.9 to 1.0	0.94
Neoplastic disease	1.0	0.57 to 1.7	0.58
S. creatinine	0.9	0.8 to 1.0	0.01
Hypertension	0.8	0.4 to 1.5	0.51
Male gender	0.7	0.4 to 1.2	0.22
Preexisting renal failure	0.6	0.3 to 1.5	0.16

OR—Odds ratio; C.I.—confidence interval.

Multiple Logistic Regression Analysis

The primary multivariate analytic method chosen was logistic regression. The factors found to be significant were coma, oliguria, jaundice, hypotension, assisted ventilation, hospital-acquired ARF, sepsis, and hypoalbuminemia. Age and male gender were included in the model as both are clinically important as they influence the risk for and manifestation of organ failures and complications. Clinically important risk factors of potential importance could be incorporated whether significant or not, depending on adequacy of the sample.[16&17] Age in years was included as a continuous variable. Serum creatinine was excluded as it was not significant on multivariate analysis. The Odds ratio (corrected to one decimal), confidence intervals, and P value of the risk factors evaluated are given in Table 4. The regression coefficients were used to develop a model (Vellore model). The regression equation thus derived is

Table 4. Multivariate logistic regression model for predicting mortality in ARF.

Risk factor	OR	95% C.I.	p
Coma	10.4	2.58 to 42.2	< 0.001
Oliguria	4.9	1.85 to 12.93	< 0.001
Jaundice	3.7	1.09 to 12.58	0.04
Hypotension	3.1	1.12 to 8.81	0.03
Assisted ventilation	2.3	0.72 to 7.73	0.04
Hospital-acquired ARF	2.3	0.96 to 5.53	0.02
Sepsis	2.2	0.92 to 5.32	0.04
Hypoalbuminemia	1.7	0.35 to 2.81	0.05
Age	1.0	0.99 to 1.04	0.16
Male gender	0.6	0.23 to 1.37	0.20

OR—Odds ratio; C.I.—confidence interval.

Probability of death = e^y/1 + e^y where

Age is expressed in years. Male gender, oliguria, hypotension, jaundice, coma, assisted ventilation, sepsis, hospital acquired ARF, and hypoalbuminemia arepresented as absent (0) or present (1). When a discriminant score of 0.5 was used as the cut-off point, a positive predictive value of 90% and sensitivity of 77% were achieved.

Example

A 40-year-old man developed acute renal failure following an attack of malaria. When evaluated 24 hours after admission, he was oliguric, jaundiced, hypotensive, serum albumin was 2.9 g/dL, and serum creatinine was 7.2 mg/dL. He was conscious and was maintained on 50% oxygen by venturi mask with which oxygen saturation was maintained. The probability of death as calculated by the equation is 0.79.

Model Validation, Comparison with Other Models and Model Performance in Subgroups

One hundred (265 patients) bootstrap samples were run on the data. Table 5 shows the calculated OR and 95% CI for each of these variables in the original equation, validated in the developmental sample by bootstrapping. The model was also prospectively validated in a set of 194 patients and also compared with other models (APACHE III, SAPS2, and Liano) in the same set of patients. Likelihood Ratio, ROC area, Confidence intervals, and Hosmer-Lemeshow goodness-of-fit p value[18] are shown in Table 6. The model was superior to two generic models, (APACHE III, SAPS2) and one ARF specific model (Liano). Figure 1 shows the ROC curves for the new model, two generic models, and the Liano model.

Figure 1. Receiver operating curves comparing Vellore model with other generic models and Liano score.

Table 5. Bootstrap Estimation (Replicated 100 Times).

Variable	OR	95% C.I. (bias corrected)
Coma	10.45	2.38 to 42.86
Oliguria	4.90	1.90 to 24.1
Jaundice	3.70	1.13 to 13.04
Hypotension	3.13	0.97 to 9.73
Assisted ventilation	2.37	0.79 to 7.98
Hospital-acquired ARF	2.37	0.69 to 7.39
Sepsis	2.20	0.40 to 6.67
Hypoalbuminemia	1.68	0.48 to 3.42
Age	1.02	0.98 to 1.05
Male gender	0.86	0.18 to 1.94

OR—Odds ratio; C.I.—confidence interval.

Table 6. Comparison of Vellore model with other models.

Model	Likelihood ratio	Area under ROC curve	Standard error	95% C.I.	Hosmer-Lemeshow Goodness of fit p
Vellore	119.5	0.92	0.02	0.87 to 0.96	0.67
APACHEIII	89.6	0.86	0.03	0.81 to 0.92	0.36
SAPS2	91.5	0.87	0.03	0.81 to 0.92	0.36
Liano et al.	114.0	0.90	0.02	0.86 to 0.96	0.46

ROC—Receiver operating characteristic C.I.—Confidence interval.

The model was also validated in different subgroups including hospital-acquired ARF, community-acquired ARF, ARF in critical care units, ARF in general medical wards, ARF with sepsis as etiology, and ARF with other etiologies and was found to have excellent discriminatory capacity by ROC curve area and good calibration as evidenced by nonsignificant Hosmer Lemeshow goodness-of-fit p value (Table 7).

Table 7. Model performance in different subgroups of ARF.

Model	Likelihood ratio	Area under ROC curve	95% CI	Hosmer-Lemeshow Goodness of fit p
Whole group (n = 459)	297.5	0.91	0.88 to 0.94	0.46
Hospital-acquired	117.6	0.93	0.89 to 0.96	0.67
Community-acquired	129.7	0.88	0.82 to 0.94	0.35
ARF in ICU setting	102.5	0.91	0.86 to 0.96	0.37
ARF in general wards	39	0.77	0.68 to 0.87	0.26
Sepsis	89.9	0.91	0.85 to 0.96	0.52
Nonsepsis etiology	180.1	0.91	0.86 to 0.95	0.45

ROC—Receiver operating characteristic; C.I.—Confidence interval.

DISCUSSION

Numerous predictive models providing both generic and specific scores have been developed to predict the outcome in acute renal failure, most of them in the critical care setting. There is a need for an accurate specific model predicting mortality, which is applicable for different patient populations. This can identify a fraction of those patients who will eventually die with extremely high specificity. We evolved a new prognostic model by logistic regression analysis, which pertains to a tertiary care hospital in a tropical environment.

Bootstrapping in the original developmental sample as well as in a prospective validation sample validated it. It was also compared with two generic models (APACHE III and SAPS2) and one ARF specific model. These models were selected for comparison as they are representative of each type of model and are widely used for prognostication. Many of the prognostic factors identified by this study correspond to those reported by other investigators. Assisted ventilation, jaundice, and coma have been implicated as predictive factors by different studies and indicate organ system failure. Although not statistically significant, male gender was associated with a favorable outcome, the reason for this is unclear. This is in contradiction with some of the other studies.[19] As statistical techniques permit inclusion of clinically important variables in a model,[16&17] age and gender were included as they influence the complications and severity of organ system involvement. Oliguria was a significantrisk factor predicting mortality, with a four-fold risk. Earlier studies have shown that once dialysis is initiated, oliguria predating initiation of dialysis is no longer a risk factor.[18] This underlines the importance of the decision for dialysis influenced by oliguria. Hypoalbuminemia portends high mortality in ARF.[20] This was found to hold true in our study also, although the odds ratio (OR) was low in the multivariate analysis. Sepsis as etiology carries a high risk for mortality. This is probably due to the high incidence of organ system failure and complications observed in sepsis-induced ARF. Hospital-acquired ARF also portended a grave prognosis in view of the multiple comorbid and debilitating factors in these patients.

This study is robust, as it has been validated internally (by bootstrapping) as well as prospectively in a comparable sample. The receiver operating characteristic curve obtained in both the developmental sample and validation sample showed that the model had good discriminatory capacity. The model showed good calibration; it also showed good performance in different subsets of patients. Although the study has been carried out in a general ARF population, the fact that acute tubular necrosis (ATN) was the major etiology might have influenced the model. But the model was performing equally well in both sepsis-induced ARFs and patients with etiologies other than sepsis.

There are a few potential limitations for this study. The inclusion of patients irrespective of whether they were dialyzed or not might have biased the study sample to a less critically ill population. Intermittent peritoneal dialysis was not employed during the study period, so the results may not apply to a center where it is an important modality of therapy.

In summary, we developed a regression model, the Vellore model that predicts mortality in the ARF in tropics using prospective data from a cohort including both hospital-acquired and community-acquired ARF. The model parameters identified in this study were age, male gender, coma, oliguria, jaundice, hypotension, assisted ventilation, hospital-acquired ARF, sepsis, and hypoalbuminemia. The Vellore model was superior to two generic prediction models and one specific ARF model. This model needs to be cross-validated in other centers and other populations.

ACKNOWLEDGMENTS

The authors thank Mr. Ebnezer Sunderraj and the Computerized Hospital Information Processing Service for providing the daily serum creatinine reports of all the inpatients of the hospital for the study period.