Article Text
Abstract
Objective It is difficult to predict the outcome in patients with acute liver failure (ALF) using existing prognostic models. This study investigated whether early changes in the levels of dynamic variables can predict outcome better than models based on static baseline variables.
Design 380 patients with ALF (derivation cohort n=244, validation cohort n=136) participated in a prospective observational study. The derivation cohort was used to identify predictors of mortality. The ALF early dynamic (ALFED) model was constructed based on whether the levels of predictive variables remained persistently high or increased over 3 days above the discriminatory cut-off values identified in this study. The model had four variables: arterial ammonia, serum bilirubin, international normalised ratio and hepatic encephalopathy >grade II. The model was validated in a cohort of 136 patients with ALF.
Results The ALFED model demonstrated excellent discrimination with an area under the receiver operator characteristic curve of 0.91 in the derivation cohort and of 0.92 in the validation cohort. The model was well calibrated in both cohorts and showed a similar increase in mortality with increasing risk scores from 0 to 6. The performance of the ALFED model was superior to the King's College Hospital criteria and the Model for End stage Liver Disease score, even when their 3-day serial values were taken into consideration. An ALFED score of ≥4 had a high positive predictive value (85%) and negative predictive value (87%) in the validation cohort.
Conclusion The ALFED model accurately predicted outcome in patients with ALF, which may be useful in clinical decision-making.
- ALF
- prognosis
- India
- acute fatty liver
- acute liver failure
- gastrointestinal pathology
- hepatic encephalopathy
- hepatitis
- liver failure
- liver
- liver cirrhosis
- cancer
- hepatitis E
- Helicobacter pylori
- acid-related diseases
- non-ulcer dyspepsia
- genetic polymorphisms
- gastric neoplasia
Statistics from Altmetric.com
- ALF
- prognosis
- India
- acute fatty liver
- acute liver failure
- gastrointestinal pathology
- hepatic encephalopathy
- hepatitis
- liver failure
- liver
- liver cirrhosis
- cancer
- hepatitis E
- Helicobacter pylori
- acid-related diseases
- non-ulcer dyspepsia
- genetic polymorphisms
- gastric neoplasia
Significance of this study
What is already known about this subject?
The existing prognostic models of acute liver failure (ALF) are based on admission parameters and have poor accuracy in predicting outcome.
ALF is a dynamic process where variables predicting outcome at admission change over time, so the clinical course and outcome vary accordingly.
Performance of a model based on early dynamicity of predictive variables has never been evaluated.
What are the new findings?
In a prospective study which included a large number of patients with ALF (derivation cohort), a prognostic model was developed using changes in the independent prognostic variables associated with mortality over 3 days, the ALF early dynamic (ALFED) model, and this model was validated in a subsequent validation cohort of patients with ALF.
The ALFED model was based on early dynamicity of four variables: arterial ammonia, international normalised ratio, serum bilirubin and hepatic encephalopathy.
This model had excellent discrimination and calibration in both cohorts.
The performance of the ALFED model was superior to the King's College criteria and the Model for End stage Liver Disease score, even when their early dynamicity was taken into consideration in a similar way.
How might it impact on clinical practice in the foreseeable future?
Because of its simplicity and accuracy, the ALFED model might be very useful in risk stratification and clinical decision-making in patients with ALF. This may assist clinicians to select appropriate candidates for liver transplantation and also to avoid unnecessary transplantation which is expensive, has perioperative mortality and needs lifelong immunosuppression.
Introduction
Acute liver failure (ALF) has a high mortality without liver transplantation (LT).1 2 However, a proportion of patients survive with supportive care alone. The prognostic models used to predict mortality in ALF should therefore be robust in order to select appropriate candidates for LT and to prevent avoidable LT. Furthermore, the results of urgent LT in ALF continue to be inferior to non-urgent LT.3 LT is an expensive treatment, is rarely available in developing countries and requires lifelong immunosuppression. A number of prognostic models have been proposed for patients with ALF. These include the King's College Hospital (KCH) criteria,4 Clichy criteria,5 serum Group-specific component protein levels,6 liver volume on CT scanning,7 blood lactate levels,8 hyperphosphataemia,9 Acute Physiology And Chronic Health Evaluation II score,10 serum alfa-fetoprotein levels (AFP)11 and the Model for End stage Liver Disease (MELD) score.12 13 Unfortunately, none of these models has consistently demonstrated a reliable accuracy in predicting outcome. Many studies have considered transplanted cases as ‘non-survivors’ which may be erroneous. In general, these prognostic markers have high specificity but unacceptably low sensitivity.14 15
ALF is a dynamic process where variables determining prognosis at admission change over time, and thus the clinical course varies accordingly. Serial measurement of predictive variables may therefore be more informative in following the clinical course in such patients. A study on acetaminophen-induced ALF showed that patients with a continuing deterioration in prothrombin time between days 3 and 4 after overdose had a higher mortality than patients in whom the prothrombin time improved (93% vs 22%).16 A study of serum AFP in patients with ALF showed that patients with increasing AFP levels during 3 days after admission had better survival than those with declining levels.17 Another study has shown a good correlation between serial arterial ammonia levels and evolving cerebral edema (CE) in patients with ALF.18 It therefore seems logical that evaluating serial changes in the predictive variables can predict outcome better than using baseline variables.
The aim of the present prospective study was to develop a prognostic model based on early changes in levels of variables which predicted outcome independently at admission in a cohort of patients with ALF; to determine whether this model was better than the established prognostic models like KCH-criteria and MELD score, and finally to validate the model in a separate cohort of Patients with ALF, in order to determine the potential value of this model for clinical decision making. Our situation provided us opportunity to assess above hypothesis prospectively because LT was not available at our centre, and each patient was managed medically till recovery or death.
Patients and methods
Between January 2004 and June 2011 we prospectively evaluated 405 consecutive patients with ALF who presented at the Department of Gastroenterology, All India Institute of Medical Sciences. Patients who died on or before day 2 after admission (n=25) were excluded. Of the remaining 380 patients, the first 244 consecutive patients were selected for the derivation cohort and 136 patients were assigned to the validation cohort. The study was approved by the ethics committee of our institute and consent was obtained from the nearest relative of the patient at the time of enrolment.
Definitions of variables
ALF was defined by the occurrence of encephalopathy within 4 weeks of symptoms in the absence of pre-existing liver disease.19 Cerebral edema (CE) was defined by the presence of spontaneous or inducible decerebrate posturing or by the presence of any two of the following: hypertension (BP >150/90 mm Hg), bradycardia (heart rate <60/min), pupillary changes or neurogenic hyperventilation.20
Grading of hepatic encephalopathy (HE)21 was done as follows:
I. Loss of sleep rhythm, drowsiness, confusion and flapping tremors.
II. Features of grade 1 encephalopathy with loss of sphincter control.
III. Unconsciousness with no response to oral commands, but responding to painful stimuli.
IV. Deep unconscious state with no response to pain.
Management protocol
All patients were managed in the intensive care unit. A uniform management protocol was followed which included stress ulcer prophylaxis, monitoring and correction of blood sugar levels, maintenance of mean arterial pressure >60 mm Hg and elective ventilation for patients with grade IV encephalopathy or grade III encephalopathy with CE. Intravenous mannitol was used to control CE. Prophylactic antibiotics were given in all cases. Each patient underwent daily microbiological surveillance to detect infection. Antibiotic therapy was modified as indicated, based on the results of positive cultures. Renal replacement therapy (haemodialysis) was used when required.
Estimation of arterial ammonia
Arterial ammonia levels were estimated by an enzymatic method (Randox Lab Ltd, Crumlin, UK) in heparinised plasma22 at admission and every 24 h for the next 5 days. All samples were transported in ice and processed within 15 min. Samples for ammonia were obtained in fasting state before the dose of mannitol and before dialysis, if needed. None of patients received lactulose, probiotics, prebiotics or non-absorbable antibiotics during hospitalisation.
Statistical analyses
Normally distributed continuous variables were expressed as mean (SD) and continuous variables with skewed distribution were expressed as median (range). Data were analysed by using SPSS software V.15.0.
In the derivation cohort, univariate analyses were performed using appropriate tests to determine the variables which were significantly different between patients who survived or died. Significant continuous variables were dichotomised using a cut-off identified on receiver operator characteristic (ROC) curves plotted between variables levels and mortality. Multivariable logistic regression model was applied by taking those predictor variables which were significant with the outcome (death vs survived) using a stepwise forward selection procedure. Each independent predictor was assigned a value of 1 corresponding to the rounded integer value of its β coefficient in the regression model. An overall risk score was calculated for each patient by adding together the points corresponding to their risk factors. Because all variables which predicted mortality at admission revealed changes in the levels following hospitalisation, patients were categorised into groups depending on changes in the levels across the discriminate cut-off values by day 3 (high persistent or rising levels vs low persistent or declining levels). After incorporating these groups into a multivariate logistic regression analysis, another model was constructed—the ALF early dynamic (ALFED) model. The discrimination of the models was described using the area under the ROC (AUROC) curve, while their calibration was assessed using an observed versus predicted plot and the Hosmer–Lemeshow statistic. A p value >0.05 implied no significant difference between the observed and expected values, and the goodness of fit was considered acceptable.
The performance of the model was then externally validated in an independent cohort. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), likelihood ratio (LR) and diagnostic accuracy of the model at relevant cut-off values were calculated.
Results
Patient characteristics
The characteristics of the patients in the derivation and validation cohorts are compared in table 1. The two populations were not exactly similar. The mean international normalised ratio (INR) (p=0.01), median arterial ammonia (p=0.002), hepatic E virus (HEV)-induced ALF (p<0.001) and mortality rates (55% vs 43%, p=0.02) were significantly higher among patients in the validation cohort than those in the derivation cohort. The median time of death from admission was 6 (3–24) days in the derivation cohort and 5 (3–18) days in the validation cohort.
Predictors of mortality in the derivation cohort
In a univariate analysis, the baseline variables which differed significantly between those who died and those who survived were age (p=0.05), serum bilirubin (p=0.003), arterial ammonia (p<0.001), advanced (>grade II) HE (p<0.001), CE (p<0.001) and INR (p<0.001). However, in multivariate analysis, only advanced HE, INR, serum bilirubin and arterial ammonia independently predicted mortality (table 2). The best cut-off values for INR, arterial ammonia and serum bilirubin discriminating between those who died and those who survived were 5, 123 μmol/l and 15 mg/dl, respectively. The OR, 95% CI and β coefficient for each independent predictor are shown in table 2.
Prognostic model development
Prognostic models were generated from the derivation cohort (table 2). The initial model was based on independent predictors of mortality at admission. Corresponding to the β coefficient in the regression model, each baseline predictor was assigned 1 point and a 4-variable model with risk score varying from 0 to 4 was generated. The model passed the Hosmer–Lemeshow goodness-of-fit test (p=0.10); however, its discriminative ability was only modest (AUROC 0.67, 95% CI 0.60 to 0.74).
It was noted that patients with persistently high levels or continuing rising levels of variables above the corresponding discriminatory cut-off value for >2 days had higher mortality than those in whom the levels declined or remained below the cut-off value (figure 1). A dynamic model was therefore constructed on the basis of early changes in the levels of each significant variable. We dichotomised patients into two categories, one with persistently high or rising levels and the other with persistently low or declining levels after 3 days of hospitalisation. After incorporating these groups into a multivariate regression analysis, the ALFED model was developed (table 2, lower panel). The four-variable ALFED model had risk scores varying from 0 to 6. This model had excellent discrimination (AUROC 0.91, 95% CI 0.87 to 0.94, figure 2A) and good calibration based on the observed versus predicted plot and the Hosmer–Lemeshow statistic (p=0.80, figure 2C).
Risk stratification with the ALFED model
Based on the risk score and associated risk of mortality, patients were stratified into three risk categories as follows: 0–1: low risk, 2–3: moderate risk, 4–6: high risk (table 3). A total of 77 of 87 patients (88.5%) in the high-risk category died compared with only 2 of 78 patients (2.6%) in the low-risk category and 26 of 79 patients (33%) in the moderate-risk category. For predicting mortality we identified score cut-off values which had high sensitivity (≥3), high specificity (≥5) and the highest combined sensitivity and specificity (≥4).
The performance of the model at each of these cut-off values is shown in table 4.
Validation of the predictive model
When applied to the validation cohort, the ALFED model retained good discrimination as well as calibration (AUROC 0.92 (95% CI 0.87 to 0.96), Hosmer–Lemeshow statistic p=0.19; figure 2B,D). The model successfully stratified the 136 patients according to their risk of mortality and showed a similar stepwise increase in mortality with increasing score. For each score the mortality rates were similar between the derivation and validation cohorts (table 3). Because the aetiologies of ALF differed between the two cohorts (table 1), we tested the performance of the model separately in patients with HEV- and non-HEV-induced ALF. The model showed good discrimination and calibration in these subgroups (HEV-ALF: AUROC 0.90 (95% CI 0.79 to 0.99), goodness of fit p=0.21; non-HEV-ALF: AUROC 0.92 (95% CI 0.86 to 0.97), p=0.44). Likewise, the prognostic accuracy for all relevant cut-off scores was also calculated for validation cohort (table 4).
The combined sensitivity and specificity of the ALFED score for predicting mortality was highest at a score of ≥4 in both cohorts. At each cut-off the LR was lower in the validation cohort. Because the LR represents the probability of death relative to the mean risk in the cohort, a lower positive LR in patients in the validation cohort was because of a higher pretest probability (prevalence) of death in this cohort (55% vs 35%, p=0.02). However, the post-test odds, which provide the post-test probability of death, calculated after adjusting LR for pre-test odds, showed good agreement between the derivation and validation cohorts (table 4).
Comparison with alternative predictive models
The MELD score was significantly higher among non-survivors than survivors (36.4±3.1 vs 32.5±64.3, p<0.001 in the derivation cohort and 37.4±2.8 vs 34.1±4.0, p<0.001 in the validation cohort). The best discriminatory MELD cut-off for predicting mortality was 35. In the derivation cohort, while 64% of patients (73/104) with admission MELD >35 died, death also occurred in 24% of patients (32/130) with MELD <35. Remarkably, 29% of patients (n=30) who died never developed a MELD score ≥35 even by day 3 of admission. However, MELD displayed a fair discrimination when its trend over 3 days (ie, high persistence or increasing MELD vs low persistence or decreasing across a cut-off of 35) was used to predict mortality, with AUROC 0.72 (95% CI 0.65 to 0.78) in the derivation cohort and 0.76 (95% CI 0.66 to 0.86) in the validation cohort. However, these values were inferior to those obtained using the ALFED criteria (>0.90 in both cohorts).
The KCH criteria for mortality were documented in 104 (42.6%) and 61 (45%) patients with ALF in the derivation and validation cohorts, respectively. The proportion of patients with KCH criteria among patients with ALF who died was significantly higher than in similar patients who survived (66% (n=69) vs 25% (n=35) in the derivation cohort (p<0.001) and 57% (n=43) vs 33% (n=20) in the validation cohort (p<0.001)). The performance of the KCH criteria was also suboptimal; for example, in the derivation cohort 36 patients (25.7%) died without meeting the KCH criteria and 44 (42%) of those who died never fulfilled the KCH criteria even by day 3. In contrast, 35 patients (34%) survived even though they fulfilled the KCH criteria. Also, compared with the ALFED model, the KCH criteria over the first 3 days had inferior discrimination with AUROC 0.70 (95% CI 0.63 to 0.77) in the derivation cohort and 0.62 (95% CI 0.52 to 0.72) in the validation cohort. All characteristics determining the prognostic accuracy of the MELD or KCH criteria were inferior to those obtained by the ALFED model in the two cohorts (table 4).
Discordance between predictive models
Significant discordance was defined as categorisation of patients into a low-risk category by KCH criteria or MELD and into a high-risk category by the ALFED model and vice versa.
Derivation cohort
Among 140 patients who did not meet the KCH criteria for mortality, 31 (22%) fulfilled the ALFED high-risk criteria for mortality, of whom 25 died. Similarly, 26 of 130 patients (20%) whose MELD score was <35 satisfied the high-risk criteria of the ALFED model and 21 of them eventually died. The number of patients falling into the low-risk category by the ALFED model among patients with positive KCH criteria (n=104) or high MELD (n=114) were 18 and 20, respectively. Among these, 17 of 18 patients and 18 of 20 patients, respectively, survived.
Validation cohort
Among 73 patients who did not meet the KCH criteria, 33 (45%) fulfilled the ALFED high-risk criteria and 27 (82%) of these died. Similarly, 17 of 48 patients (35%) whose MELD score was <35 satisfied the high-risk ALFED criteria and 12 (71%) of them died. All eight patients in the low-risk ALFED category (score <2) among the 85 patients with a high MELD score survived. Among patients meeting the KCH criteria (n=63), none satisfied the low-risk ALFED criteria; however, 8 of 9 patients with an AFLED score of 2 survived.
Discussion
The results of our study suggest that early dynamicity of prognostic markers predicts outcome better than the static baseline levels. The ALFED model was based on dynamicity of variables during 3 days of hospitalisation. This model showed excellent discrimination and nearly identical performance in both cohorts (AUROC 0.91 in the derivation cohort and 0.92 in the validation cohort). The discrimination of this model was better than the MELD score and KCH criteria, even when their evolution over 3 days was considered. The model also showed good calibration in both cohorts.
The ALFED model effectively stratified all patients according to their mortality rates (table 3). The mortality rates among low-risk patients were only 2.6% (2/78) in the derivation cohort and 3.8% (1/26) in the validation cohort. Among high-risk patients the mortality rates were 88.5% (80/84) in the derivation cohort and 85% (68/80) in the validation cohort. The ALFED scores were more accurate at the borders of the scale in both cohorts. The mortality rate in the moderate-risk category (score 3) seemed to be different between the two cohorts (36.8% vs 19%), but this difference was not statistically significant (p=0.19, table 3). Even this difference could be attributed to a higher pretest probability (prevalence) of death in the validation cohort resulting in a higher proportion of patients who died being grouped into the high-risk category. High-risk patients identified by this model may be considered as candidates for LT while low-risk patients can be managed medically. The strategies have to be individualised in patients in the moderate-risk category because the spontaneous survival obtained in the present study for this category (63–81%) is similar to the reported survival after emergency LT in ALF.3 23 Whether patients with ALF with an ALFED score of 3 should be transplanted or not should be at the discretion of the individual centre based on the transplant infrastructure and organ availability. The ALFED model accurately identified several patients in the low-risk category who were in the high-risk category using the MELD or KCH criteria and vice versa. Thus, the ALFED model can effectively select patients who are likely to survive without LT. However, it needs to be tested in a transplant centre whether patients with a high ALFED score (≥4) would benefit from LT because the outcome of LT may be affected by other variables such as graft availability, graft quality, the extent of neurological status and other organ failure at the time of transplantation.
The KCH criteria are the most widely used worldwide.15 23 Their specificity is acceptable but sensitivity is limited.4 15 24–26 For non-acetaminophen-induced ALF the diagnostic accuracy of the KCH criteria varies from 55% to 92%.15 25 In our derivation cohort, 25.7% of patients (n=36) died without having KCH mortality criteria and 42% (n=44) of those who died never fulfilled KCH mortality criteria even by day 3. The suboptimal performance of the KCH criteria is in agreement with results of previous studies from India,27 28 which may be attributed to the predominant hyperacute presentation and viral aetiology of ALF in these cohorts. MELD is another model which has been used for predicting mortality in patients with ALF.5 12 13 29 There are reports indicating superiority of MELD over the KCH criteria as well as similar efficacy of both models in predicting mortality.5 13 27 29 The US ALF study group reported that a MELD score >30 had a high PPV (81%) for death or transplantation but a low NPV (41%).30 In the ALF cohort in the present study the performance of MELD was inferior to the ALFED model. Even serial evaluation of MELD and KCH criteria until day 3 did not result in significant improvement in their performance (table 4).
The PPV and NPV of a model are important parameters for selecting a candidate for LT. Preferences for PPV ensure that all patients who need a transplant get it while a preference for NPV minimises unnecessary LT. In the ALFED model a score of ≥4 had high PPV (85%) and high NPV (87%). Furthermore, a score of 5 had a PPV of 94% while a score of 3 had a NPV of 90% (table 4). The improved performance of our model was due to serial evaluation of predictors identified at admission, indicating that the dynamicity of the model is indeed important in ALF. Furthermore, inclusion of arterial ammonia and advanced HE added strength to the model. At admission, INR had a higher OR (3.1) for mortality than arterial ammonia (OR 2.3). However, evaluation of the serial levels of these factors showed that a persistently high or increasing level of arterial ammonia for 3 days was more ominous (OR 10.8) than a persistently high or rising INR (OR 3.4) (table 2). This may be because ammonia has a direct pathogenic role and has been implicated in the development of HE, CE, brain herniation and neutrophil dysfunction in patients with ALF.18 31–33 Our study identified that severity of HE at admission (OR 2.3) and its early trend (OR 13.5) is a powerful predictor of mortality. Various studies have shown that the severity of HE is a strong predictor of outcome in patients with ALF.13 20 28 A recent meta-analysis in patients with non-acetaminophen-induced ALF showed that the KCH criteria had the best performance in patients with advanced HE.34 Notably, the severity of HE is not taken into account in the MELD or KCH criteria. In our study, serum bilirubin had the lowest OR and β coefficient in the regression model and hence had the lowest impact on the ALFED score, which is in line with the recently published paper by Bechmann et al.35 However, bilirubin was still retained in the model because its impact was larger in association with other variable(s).
A valid criticism of this study is that the identification of patients is too late for a successful LT. However, the existing prognostic models are inaccurate in determining both the need and timing for LT. In our study the median time from admission to death was 6 days in the derivation cohort and 5 days in the validation cohort and only 26 of 406 patients died before 3 days, suggesting the applicability of this model in nearly 95% of patients with ALF. We also found that the discrimination of outcome by the baseline model was only modest (AUROC 0.67). On day 2 of admission the changes in the levels of predictive variables across the discriminatory cut-off values were seen in only a minority of patients, hence the performance of the model was not expected to be improved much on day 2. Only by day 3 were the changes in the predictive variables more noticeable and more accurately associated with outcome, and thus the ALFED model gave excellent discrimination. Notably, the day 3 parameters were assessed during the early hours of day 3 (just after 48 h of admission) so the actual waiting time was not exactly a complete 3 days. Modern intensive care has improved outcomes of patients with ALF.36 With better control of haemodynamic parameters, CE and infection, the window of transplant-free survival can be increased. The ALFED model can avoid unnecessary LT in a significant proportion of patients. Also, the waiting time from listing to LT remains 2–3 days in most transplant centres. It may therefore be reasonable to wait until day 3 so that patients are selected more accurately for either LT or medical treatment. In order to minimise the delay in obtaining a graft once the decision for LT is made, all patients with ALF may be considered as potential candidates for LT. However, a definitive decision for LT may be based on the dynamicity of predictive variables. The ALFED model may be more suitable for living donor LT because it may reduce the waiting time for obtaining a graft. The aetiology of ALF in our patients was mainly viral hepatitis (especially HEV), which is different from that reported from the USA, UK and some parts of Europe. However, in many other parts of the Western world 20–50% of ALF has been reported to be due to viral hepatitis.37 38 Furthermore, HEV is hyperendemic in India, China and other South Asian countries and African countries such as Algeria, Sudan and Somalia. The present study population may therefore indeed be representative of the majority of patients with ALF globally. HEV hepatitis is also increasingly being recognised in several developed countries such as France, Italy, Greece, the Netherlands, UK, Japan, Hong Kong and Australia.39 A recent report from the UK showed that 13% of cases with drug-induced liver injury were due to HEV.40
The principal strength of our model is its simplicity and accuracy. It consists of clearly defined and routinely available predictors. This model is based on prospective data on a large number of patients with ALF from a single centre, which ensures a homogeneous cohort managed with a similar treatment protocol. A comprehensive natural course was studied in all patients without interruption by LT. Because our model was validated only at the original study site, it needs to be validated in regions with differing populations and where the pathologies of ALF differ. A minor drawback of this model is inclusion of the HE which is a subjective variable. However, because HE is a strong predictor of mortality, it cannot be ignored. Moreover, we followed Riegler and Lake's modification21 of the West Haven criteria which provides more objective parameters to differentiate between grades of HE.
In summary, in this study we have prospectively derived and validated a dynamic model for outcome in patients with ALF that is simple, reliable and accurate. This model might be very useful in risk stratification and clinical decision-making in such patients.
References
Footnotes
Funding Supported by the Indian Council of Medical Research (ICMR) through the project “Advanced Center for Liver Diseases” to the Department of Gastroenterology, All India Institute of Medical Sciences, New Delhi.
Competing interests None.
Ethics approval Ethics approval was approved by the ethics committee of All India Institute of Medical Sciences.
Provenance and peer review Not commissioned; externally peer reviewed.