The Pancreatitis Outcome Prediction (POP) Score: A newprognostic index for patients with severe acute pancreatitis*

David A. Harrison, PhD; Giovanna D’Amico, MD; Mervyn Singer, MD, FRCP

Acute pancreatitis is a commondisorder, affecting 10–20 permillion of the U.K. populationper annum (1). The clinical

picture ranges from relatively mild symp-toms to multiple organ failure with inter-current sepsis, particularly in the severenecrotizing or hemorrhagic forms of thedisease. The overall mortality is reportedas 10% to 15%, ranging from �5% forthose presenting with mild acute pancre-atitis to 20% to 25% with severe acutepancreatitis (1). Time to death is bipha-sic, occurring either within days of ad-

mission or after a protracted hospital staylasting weeks to months (2–5). Variousprognostic scoring systems have been de-veloped yet are generally based on eitherexpert consensus—such as the Atlantaclassification (4)— or small specialist-center populations—such as the Ranson(6), Glasgow (7), and modified Glasgow(8) criteria—usually collected several de-cades ago and only containing a smallproportion of patients with severe acutepancreatitis. Their applicability to the21st century patient with severe acutepancreatitis in a more technologically ad-vanced, albeit nonspecialist, intensivecare unit (ICU) is uncertain. Indeed, arecent international consensus confer-ence (9) redefined severe acute pancreati-tis from a critical care perspective to rep-resent pancreatitis in the context ofsystemic organ dysfunction, irrespectiveof local complications, such as peripan-creatic fluid collection, necrosis, abscess,or pseudocyst.

We thus decided to interrogate a largenational intensive care patient databasecollected from U.K. ICUs to examine out-comes in a large cohort of patients admit-

ted with severe acute pancreatitis and tosee whether an improved prognosticscoring system can be developed based onthis sizable data set using physiologic andbiochemical variables collected withinthe first 24 hrs of intensive care admis-sion.

MATERIALS AND METHODS

The Intensive Care National Audit and Re-search Centre Case Mix Programme Database(CMPD) contains information on case mix,outcome, and activity for all admissions toadult, general critical care units participatingin the Case Mix Programme, a national com-parative audit in England, Wales, and North-ern Ireland, active since 1995. The Case MixProgramme has received approval from the Pa-tient Information Advisory Group to hold patientidentifiable information without consent (ap-proval number: PIAG 2-10(f)/2005). Approval ofthe study by an institutional review board wasnot required. The data are collected by traineddata collectors according to precise rules anddefinitions and undergo extensive validation forcompleteness, logic, and consistency (10). Thedata available for analysis covered admissions to159 ICUs during the time period December 1995to June 2003, inclusive.

*See also p. 1787.From the Intensive Care National Audit & Research

Centre, Tavistock House, Tavistock Square, London,United Kingdom (DAH); and Bloomsbury Institute ofIntensive Care Medicine, University College London,Cruciform Building, Gower Street, London, UnitedKingdom (GD, MS).

The authors have not disclosed any potential con-flicts of interest.

For information regarding this article, E-mail:david.harrison@icnarc.org

Copyright © 2007 by the Society of Critical CareMedicine and Lippincott Williams & Wilkins

DOI: 10.1097/01.CCM.0000269031.13283.C8

Objective: Severe acute pancreatitis, defined as pancreatitiswith distant organ dysfunction, is a condition carrying a highmortality and morbidity. Current outcome prediction scores arebased on small populations, usually from single specialist cen-ters. Some scores cannot be applied until several days intohospital admission. We thus sought to develop a new and moresensitive outcome prediction score—the Pancreatitis OutcomePrediction (POP) Score—for these high-risk patients.

Design: Retrospective cohort study of a large multicenter in-tensive care database.

Setting: One hundred fifty-nine U.K. intensive care units.Patients: Participants were 2,462 patients admitted to inten-

sive care units with severe acute pancreatitis.Interventions: None.Measurements and Main Results: Demographic, physiologic,

and biochemical data collected within the first 24 hrs of intensivecare unit admission were used to develop a risk prediction scoreusing logistic regression. The six variables with the strongest

relationship to hospital outcome—arterial pH, age, serum urea,mean arterial pressure, PaO2/FIO2 ratio, and total serum calcium(in order of decreasing impact)—produced a model with a prog-nostic discrimination (area under the receiver operating charac-teristic curve � 0.838) superior to other models. These six factorswere used to develop an objectively weighted multivariate prog-nostic score ranging from 0 to 40 points.

Conclusions: Prognostic stratification of patients with severeacute pancreatitis requiring intensive care offers a useful audittool to gauge unit performance and improve delineation of sub-sets for prospective trials. Prospective validation of this newoutcome prediction score is required, preferably in different coun-tries. The validity of the POP Score for either hospital or intensivecare admission could also be tested and assessed for superiorityover existing scores. (Crit Care Med 2007; 35:1703–1708)

KEY WORDS: pancreatitis; critical care; hospital mortality; inten-sive care units; models; statistical; severity of illness index

1703Crit Care Med 2007 Vol. 35, No. 7

Reasons for admission to the ICU are re-corded in the CMPD using a specially derivedhierarchical coding system, the Intensive CareNational Audit and Research Centre codingmethod (11). Admissions were selected if theirprimary reason for admission was recorded asacute pancreatitis, chronic pancreatitis, or in-fective pancreatitis. Admissions coded aschronic pancreatitis were included in the anal-yses as chronic pancreatitis alone was not con-sidered sufficient reason for admission to ICUand so these patients were presumed to havesevere acute exacerbations. Likewise, we donot feel any distinction should be made be-tween acute and infective pancreatitis. Thefollowing exclusions were applied: readmis-sions of the same patient within the samehospital stay, transfers from another ICU, pa-tients aged �16 yrs, and patients with missingoperative status or outcome (mortality at ul-timate discharge from an acute hospital).

Variables for the model were selected apriori based on an anticipated relationshipwith mortality from existing models for pan-creatitis and expert clinical opinion. The vari-ables included in the initial full model wereage, gender, operative status, lowest mean ar-terial pressure, lowest PaO2/FIO2 ratio, lowestarterial pH, highest serum urea, highest se-rum creatinine, highest serum glucose, high-est total serum bilirubin, lowest total serumcalcium, lowest serum albumin, highest whiteblood cell count, and lowest platelet count. Allphysiologic measurements were assessed overthe first 24 hrs following admission to ICU.

Continuous variables were examined usinggeneralized additive models (12) to plot therelationship between each variable and risk ofdeath. These plots were used to select a suit-able categorization. To simplify the model fur-ther, the resulting categorical variables wereeach entered into a logistic regression modeland adjacent categories were combined if theresulting coefficients were not significantlydifferent at the 10% level. Finally, admissionswith missing physiologic values were assignedto the category with the most similar mortal-ity rate.

All variables were entered into a multiplelogistic regression model. This was fitted in adevelopment data set of two thirds of criticalcare units and assessed for calibration anddiscrimination in a validation data set of theremaining one third of units. Calibration wasassessed by the Hosmer-Lemeshow C statistic(13) and discrimination by the area under thereceiver operating characteristic curve (AUC)(14). The least significant variable was re-moved from the model, and the process wasrepeated until no variables remained in themodel. The performance of each model wasassessed in the validation data, and a bestmodel was selected to balance parsimonyagainst model performance.

The results of the best model were con-verted to an integer score by multiplying allcoefficients by a constant and rounding. Fi-nally, logistic regression was used to relate thescore to a predicted probability of mortality.

The discrimination of this score was com-pared with the Acute Physiology and ChronicHealth Evaluation (APACHE) II score (15) andto a score composed of seven of the eightmodified Glasgow criteria (8) (excluding se-rum lactate dehydrogenase levels) that couldbe evaluated in the CMPD.

All analyses were performed in Stata ver-sion 8.2 (StataCorp LP, College Station, TX).

RESULTS

Of a total of 219,468 intensive careadmissions in the database, 2,839 were

admitted with pancreatitis given as theprimary reason for ICU admission, repre-senting 1.3% of all admissions, 1.9% ofall ICU deaths, and 2.6% of all ICU beddays. Of the 2,839 admissions with pan-creatitis, 161 were excluded because theywere transferred from another ICU, a fur-ther 132 were excluded because they werereadmissions of the same patient withinthe same hospital stay, nine were ex-cluded because they were aged �16 yrs,two were excluded because they weremissing operative status, and 73 were ex-cluded because they were missing hospi-tal outcome. This left a total of 2,462valid admissions included in the model.

Demographic, activity, and outcomedata for the 2,462 admissions are shown

Table 1. Demographic, activity, and outcome data for admissions with severe acute pancreatitis

No.Median (IQR)

or %Hospital

Mortality, %

Age, yrs 2,462 63 (48–74)Gender

Male 1369 55.6 41.9Female 1093 44.4 41.9

Primary reason for admissionAcute pancreatitis 2191 89.0 42.5Infective pancreatitis 148 6.0 48.0Chronic pancreatitis 123 5.0 23.6

Surgical statusNonsurgical 2074 84.2 44.3Emergency surgery 258 10.5 34.5Elective surgery 130 5.3 18.5

Source of admissionWard, same hospital 1456 59.1 42.3Theater and recovery, same hospital 388 15.8 29.1HDU, same hospital 232 9.4 59.9Emergency department, same hospital 145 5.9 40.0Other hospital (not ICU/HDU) 104 4.2 41.3Other intermediate care area, same hospital 58 2.4 51.7Recovery only, same hospital 30 1.2 46.7Radiograph, CT scanner or similar, same

hospital25 1.0 44.0

HDU, other hospital 24 1.0 33.3APACHE II scorea 2367 17 (13–22)Vasoactive drug treatment during first 24 hrsb

No 983 56.3 26.1Yes 763 43.7 61.2

Mechanical ventilation during first 24 hrsNo 1326 54.0 26.2Yes 1131 46.0 60.3

Length of stay, daysPre-ICU 2457 1 (1–3)ICU 2453 3.7 (1.4–11.0)

ICU survivors 1679 3.7 (1.6–10.0)ICU nonsurvivors 774 3.8 (1.0–13.4)

Post-ICU 1637 15 (8–30)Hospital survivors 1419 15 (9–30)Hospital nonsurvivors 218 14 (4–37)

MortalityICU 778 31.6Post-ICU, in-hospital 254 10.3

IQR, interquartile range; HDU, high-dependency unit; ICU, intensive care unit; CT, computedtomography; APACHE, Acute Physiology and Chronic Health Evaluation.

aAdmissions staying �8 hrs in the ICU only; badmissions staying �24 hrs in the ICU only.

1704 Crit Care Med 2007 Vol. 35, No. 7

in Table 1. The majority were nonsurgicalpatients, admitted from general wards.Nearly 42% of patients died in the hospital,a quarter of whom died after discharge

from intensive care. The median length ofstay in the hospital before admission to ICUwas 1 day, with 22% of admissions admit-ted to the ICU on the same day as their

admission to hospital and 14% of admis-sions having been in hospital for �1 wk.

Figure 1 shows generalized additivemodel plots of the risk of ultimate hospi-tal mortality against each of 12 continu-ous variables measured within the first 24hrs of ICU admission. Vertical lines divideeach variable into the categories thatwere entered into the model, after com-bining adjacent categories that were notsignificantly different.

A total of 14 variables, including thecategorical variables gender and opera-tive status, were entered into a stepwisemodel selection (Table 2). The develop-ment sample consisted of 87 ICUs (n �1,494) and the validation sample 61ICUs (n � 968). The selected modelincluded the six variables with thestrongest relationship to hospital out-come—arterial pH, age, serum urea,mean arterial pressure, PaO2/FIO2 ratio,and total serum calcium (in order ofdecreasing impact). The discriminationof this model in the validation samplewas only slightly worse than that of thefull model (AUC � 0.838 compared with0.841), and there was no significantlack of calibration (Hosmer-Lemeshow�2

10 � 15.36, p � .119).These six factors were used to develop

an objectively weighted multivariateprognostic score, the Pancreatitis Out-come Prediction (POP) Score, rangingfrom 0 to 40 points (Table 3). Comparedwith the other prognostic models, theAUC (95% confidence interval) of the fi-nal score in all admissions was 0.853(0.838 – 0.866) compared with 0.670(0.651– 0.688) for the seven availablemodified Glasgow criteria and 0.804(0.787–0.820) for the APACHE II score.The ratio of observed to expected deathsbased on the U.K. APACHE II model was0.97 (0.92–1.02).

The predicted probability of mortal-ity, p, is described by the followingequations:

R � �4.908 � (0.248 � POP score)

[1]

p � exp(R)/(1 � exp[R]) [2]

Figure 2 depicts this relationship be-tween the POP Score and predicted prob-ability of mortality. Figure 3 shows thecalibration of the final model.

DISCUSSION

Patients admitted to ICUs with severeacute pancreatitis constitute a significant

Figure 1. Generalized additive model (GAM) plots of risk of ultimate hospital mortality against eachcontinuous variable. Predicted probability of mortality from a logistic GAM with five degrees offreedom. Vertical lines show categorical variables entered into the logistic regression model.

Table 2. Stepwise model selection

Variables Dropped Variables AUC HL �2a

None 14 0.8413 30.13White blood cell count 13 0.8406 29.66Serum creatinine 12 0.8428 25.21Serum glucose 11 0.8422 21.69Platelet count 10 0.8415 22.53Serum bilirubin 9 0.8411 22.28Surgical status 8 0.8406 20.51Gender 7 0.8399 20.06Serum albumin 6 0.8381 15.36Total serum calcium 5 0.8347 14.90PaO2/FIO2 ratio 4 0.8272 14.50Mean arterial pressure 3 0.8154 22.04Serum urea 2 0.7911 20.90Age 1 0.7305 —Arterial pH 0 0.5000 —

AUC, area under the receiver operating characteristic curve; HL, Hosmer-Lemeshow.aHosmer-Lemeshow chi-square goodness-of-fit statistic on 10 df; values �18.30 indicate a signif-

icant lack of fit at the 5% significance level. Models fitted in development sample of two thirds of units(n � 1494); AUC and HL �2 calculated in validation sample of one third of units (n � 968). Bold typeindicates the variables included in the final model and the fit of the final model in the validationsample.

1705Crit Care Med 2007 Vol. 35, No. 7

population carrying a high mortality rateand often a prolonged length of stay rel-ative to other ICU admissions. Given thehigh mortality and morbidity associated

with severe acute pancreatitis, such pa-tients are probably best treated in an in-tensive care/high-dependency environ-ment with close attention paid to volume

resuscitation, gas exchange, nutrition,and treatment of infection and othercomplications (16). Early prediction ofthe severity of an acute attack thus hasimportant implications for managementand timely intervention. Patients codedas having either infective or chronic pan-creatitis comprised only 11% of the totalpopulation, and their inclusion does notaffect the results. While mortality waslower for admissions coded as chronicpancreatitis, this difference in mortalitywas fully explained by the POP Score. Asdiscussed earlier, such patients were likelyto have an acute exacerbation requiringICU admission, and we thus feel it is rea-sonable to include them in the analysis.

A variety of assessment tools havebeen used to identify patients with acutepancreatitis at risk of developing compli-cations and/or death. These tools fall intofour broad categories. First, there areclinico-biochemical tools specific toacute pancreatitis, such as the Ranson (6,17) and Glasgow (or Imrie) (7) criteriaand modifications thereof (8, 18 –20).However, these were developed on out-come data from single- or two-center spe-cialist units rather than a general hospi-tal population and mainly consisted ofrelatively small numbers of general wardpatients with few risk factors and a goodeventual outcome. Indeed, the largest co-hort was based on 1,005 patients of whomonly 25% fulfilled the criteria for severeacute pancreatitis (20). The increased ex-pertise in dealing with the disease in aspecialist center may be offset by a highnumber of sick tertiary referral patients,thus giving a potentially biased reflectionof nationwide mortality rates. In addition,these outcome data are largely historical,dating back several decades, and do notnecessarily reflect contemporary patientmanagement and outcomes. Finally, the

Table 3. Pancreatitis Outcome Prediction Score

Score 0 1 2 3 4 5 6 7 8 10

Age, yrs 16–29 30–39 40–49 50–59 60–69 �70Mean arterial

pressure, mmHg

�90 80 to 89 60–79 50–59 40–49 �40

PaO2/FIO2 ratio, mmHg (kPa)

�225 75–224 �75(�30) (10–29.9) (�10)

Arterial pH �7.35 7.30–7.35 7.25–7.29 7.20–7.24 7.10–7.19 7.00–7.09 �7.00Serum urea, mg/dL

(mmol/L)�14 14–22.3 22.4–30.7 30.8–47.5 �47.6

(�5) (5–7.9) (8–10.9) (11–16.9) (�17)Total serum

calcium, mg/dL(mmol/L)

8.0–9.19 7.2–7.99 6.4–7.19 �6.4(2.0–2.29) (1.8–1.99) (1.6–1.79) (�1.6)

OR OR9.2–9.99 �10(2.3–2.49) (�2.5)

Figure 2. Predicted probability of ultimate hospital mortality from the Pancreatitis Outcome Predic-tion Score.

Figure 3. Calibration of the Pancreatitis Outcome Prediction Score. Observed vs. predicted mortalityin ten equal-sized groups formed by deciles of predicted mortality.

1706 Crit Care Med 2007 Vol. 35, No. 7

Ranson and Glasgow scores require datacollected at 48 hrs after admission. Therecently published BALI score (21), basedon blood urea nitrogen, age, lactate de-hydrogenase, and interleukin-6 levelstaken on samples at admission to an in-ternational phase III trial of the platelet-activating factor receptor-antagonistLexipafant (22), provided prognostic abil-ity similar to Ranson, Glasgow, andAPACHE II scores. However, the inabilityto offer routine interleukin-6 measure-ment precludes its introduction atpresent into clinical practice.

Second, prognostication has been pro-posed using computed tomographic ex-amination and based on a five-grade se-verity score (23) or the presence anddegree of pancreatic necrosis (24). A re-cent single-center study claimed superi-ority of the Balthazar computed tomog-raphy severity index over the Ranson andAPACHE II scores in predicting outcomefrom acute pancreatitis (25).

Third, generic intensive care scoringsystems, such as the APACHE II (15) andIII (26) scores and Simplified Acute Phys-iology Score (27), have been applied tocohorts of pancreatitis patients (28–32).Generic scores have also been modifiedfor the prediction of severe acute pancre-atitis, for example by incorporating obe-sity (33). However, these modified scoreshave many of the same drawbacks of thepancreatitis-specific scores, being basedon small data sets from single centers.

Fourth, the 1992 Atlanta classification(4), which was developed by expert panelconsensus and uses all-encompassing cri-teria (local complications, systemic com-plications, need for surgery, or death),has been proposed. Yet, as the senior au-thor recently acknowledged, no attemptshave been made at large-scale objectivevalidation (34). Our database does notenable this to be performed.

The presence of three or more criteriain either the Ranson or Glasgow scoringsystems or an APACHE II score �8 hasbeen used to indicate severe acute pan-creatitis (28, 35–37). Sensitivity and spec-ificity for predicting mortality have beenreported as 63% and 76% for Ranson and72% and 84% for the modified Glasgowcriteria, respectively (38). The APACHE IIsystem shows similar sensitivity andspecificity rates using data collected overthe first 24 hrs (29), but it is more cum-bersome to perform (30). The ideal cutoffscore also remains to be determined. Interms of comparability, Venkatesan et al.(39) recently reported that only one half

of their cases classified as severe pancre-atitis by the Atlanta system could be cor-rectly predicted by an APACHE II score�8, although a severe episode was un-likely with scores �8 (negative predictivevalue of 89%). Both the APACHE II andthe Simplified Acute Physiology Scoresystems were of limited clinical utility inthe early prognostic evaluation of acutepancreatitis (30), but a deterioratingscore at 48 hrs had greater prognosticvalue (29).

The major strength of our study is theavailability of a large data set from a largeand representative sample of generalICUs in the United Kingdom. However,the data set was not collected for thepurpose of identifying patients with pan-creatitis. It was therefore not possible tovalidate the diagnosis of pancreatitis(e.g., based on amylase levels or com-puted tomography scans) or to evaluateall potential prognostic variables (e.g., se-rum lactate dehydrogenase as was in-cluded in the modified Glasgow criteria).Other pertinent details, such as the etiol-ogy of the pancreatitis and comorbiditiessuch as diabetes and nonsevere heart fail-ure, are also missing. We also lack datacollected at ICU admission for these pa-tients to test the validity of the POP Scoreat this earlier time point.

The ICU and hospital mortality ratesof 31.6% and 41.9%, respectively, may,on first sight, appear high for a medianAPACHE II score of 17. However, pancre-atitis did not warrant a separate diagnos-tic coding in the original APACHE II sys-tem, and these patients were includedwithin a diagnostic category of “gastroin-testinal infection.” Despite this, therewere still not sufficient cases for this di-agnostic category to receive a coefficient,and so these patients were weighted inthe general “gastrointestinal” categorywith a nonsurgical coefficient of 0.501and a surgical coefficient of �0.613. Wehave applied the U.K. APACHE II model(40), which gives a nonsurgical coeffi-cient of 0.658 and a surgical coefficient of0.442 for gastrointestinal infection. Thesehigh values reflect the high mortality rateof this condition relative to many others,particularly as the intense inflammatorystimulus underlying severe acute pancre-atitis cannot, in general, be removed andpatients often deteriorate further despiteintensive care. By comparison, we previ-ously reported ICU and hospital mortalityrates for the CMPD of 20.3% and 28.6%,respectively, with a mean first 24-hrAPACHE II score of 16.5 (10). The median

stay for the pancreatitis patients of 3.7days (survivors) and 3.8 days (nonsurvi-vors) compares with 1.7 days (survivors)and 2.0 days (nonsurvivors) for all ICUpatients in the CMPD (10). There was abiphasic distribution in stay for the pan-creatitis patients. Nonsurvivors with highfirst 24-hr APACHE II scores (�23) diedsoon after admission (median 2 days),presumably from their multiple organfailure. However, those with low scores(�13) who eventually died did so after anaverage ICU stay of 9 days, suggestingcontinuing physiologic deterioration inthis initially lower severity subset (41) orlate-onset complications perhaps relatedto infected necrosis (42). The conversewas seen with survivors.

CONCLUSIONS

Our retrospective analysis of a largedatabase of patients admitted to U.K.ICUs with a primary diagnosis of severeacute pancreatitis enabled the develop-ment of a new scoring system based onsix readily available physiologic and bio-chemical variables collected within thefirst 24 hrs of ICU admission. This wasmore sensitive than existing scoring sys-tems for which most criteria were acces-sible in our data set. Clearly, this newscore needs to be validated prospectivelyand, ideally, in a non-U.K. population.The utility of data taken on ICU admis-sion and, perhaps, on hospital admissionalso needs to be compared against the24-hr data reported herein, to provide aneven earlier indication of severity andprognosis. A score with good prognosticability provides the possibility of identify-ing those patients who are more likely tobenefit from new therapeutic interven-tions and management regimens. It willalso considerably facilitate prospectivestudies and comparisons of performance.

ACKNOWLEDGMENTS

We thank all those intensive care unitscontributing data to the Intensive CareNational Audit and Research Centre CaseMix Programme Database.

REFERENCES

1. Beckingham IJ, Bornman PC: Acute pancre-atitis. BMJ 2001; 322:595–598

2. Kodesch R, DuPont HV: Infectious complica-tions of acute pancreatitis. Surg Gynecol Ob-stet 1973; 136:763–768

3. Jacobs ML, Daggett WM, Civetta JM, et al:

1707Crit Care Med 2007 Vol. 35, No. 7

Acute pancreatitis: Analysis of factors influ-encing survival. Ann Surg 1977; 185:43–51

4. Bradley EL III: A clinically based classifica-tion system for acute pancreatitis. Arch Surg1993; 128:586–590

5. McKay CJ, Evans S, Sinclair M, et al: Highearly mortality rate from acute pancreatitisin Scotland, 1984–1995. Br J Surg 1999;86:1302–1306

6. Ranson JHC, Rifkind KM, Roses DF, et al:Prognostic signs and the role of operativemanagement in acute pancreatitis. Surg Gy-necol Obstet 1974; 139:69–81

7. Imrie CW, Benjamin IS, Ferguson JC, et al: Asingle centre double blind trial of trasyloltherapy in primary acute pancreatitis.Br J Surg 1978; 65:337–341

8. Blamey SL, Imrie CW, O’Neill J, et al: Prog-nostic factors in acute pancreatitis. Gut1984; 25:1340–1346

9. Nathens AB, Curtis JR, Beale RJ, et al: Man-agement of the critically ill patient with se-vere acute pancreatitis. Crit Care Med 2004;32:2524–2536

10. Harrison DA, Brady AR, Rowan K: Case mix,outcome and length of stay for admissions toadult, general critical care units in England,Wales and Northern Ireland: The IntensiveCare National Audit & Research Centre CaseMix Programme Database. Crit Care 2004;8:R99–R111

11. Young JD, Goldfrad C, Rowan K: Develop-ment and testing of a hierarchical method tocode the reason for admission to intensivecare units: The ICNARC Coding Method. Br JAnaesth 2001; 87:543–548

12. Hastie TJ, Tibshirani R: Generalized AdditiveModels. New York, Chapman & Hall, 1990

13. Hosmer DW, Lemeshow S: Goodness of fittests for the multiple logistic regressionmodel. Commun Stat 1980; A9:1043–1069

14. Hanley JA, McNeil BJ: The meaning and useof the area under the receiver operatingcharacteristics (ROC) curve. Radiology 1982;143:29–36

15. Knaus WA, Draper EA, Wagner DP, et al:APACHE II: A severity of disease classifica-tion system. Crit Care Med 1985; 13:818–829

16. Soran A, Chelluri L, Lee KK, et al: Outcomeand quality of life of patients with acute pan-creatitis requiring intensive care. J Surg Res2000; 91:89–94

17. Ranson JH, Pasternack BS: Statisticalmethod for quantifying the severity of clini-cal acute pancreatitis. J Surg Res 1977; 22:79–91

18. Osborne DH, Imrie CW, Carter DC: Biliarysurgery in the same admission for gallstoneassociated acute pancreatitis. Br J Surg 1981;68:758–761

19. Corfield AP, Cooper MJ, Williamson RC, et al:Prediction of severity in acute pancreatitis:Prospective comparison of three prognosticindices. Lancet 1985; 2:403–407

20. Cavallini G, Frulloni L, Bassi C, et al: Pro-spective multicentre survey on acute pancre-atitis in Italy (ProInf-AISP): Results on 1005patients. Dig Liver Dis 2004; 36:205–211

21. Spitzer AL, Barcia AM, Schell MT, et al: Ap-plying Ockham’s razor to pancreatitis prog-nostication: A four-variable predictive model.Ann Surg 2006; 243:380–388

22. Johnson CD, Kingsnorth AN, Imrie CW, et al:Double blind, randomised, placebo con-trolled study of a platelet activating factorantagonist, lexipafant, in the treatment andprevention of organ failure in predicted se-vere acute pancreatitis. Gut 2001; 48:62–69

23. Balthazar EJ, Ranson JH, Naidich DP, et al:Acute pancreatitis: Prognostic value of CT.Radiology 1985; 156:767–772

24. Balthazar EJ, Robinson DL, Megibow AJ, etal: Acute pancreatitis: Value of CT in estab-lishing prognosis. Radiology 1990; 174:331–336

25. Leung TK, Lee CM, Lin SY, et al: Balthazarcomputed tomography severity index is su-perior to Ranson criteria and APACHE IIscoring system in predicting acute pancreati-tis outcome. World J Gastroenterol 2005;11:6049–6052

26. Knaus WA, Wagner DP, Draper EA, et al: TheAPACHE III prognostic system. Risk predic-tion of hospital mortality for critically illhospitalized adults. Chest 1991; 100:1619–1636

27. Le Gall JR, Loirat P, Alperovitch A, et al: Asimplified acute physiology score for ICU pa-tients. Crit Care Med 1984; 12:975–977

28. Larvin M, McMahon MJ: APACHE-II score forassessment and monitoring of acute pancre-atitis. Lancet 1989; 2:201–204

29. Khan AA, Parekh D, Cho Y, et al: Improvedprediction of outcome in patients with severeacute pancreatitis by the APACHE II score at

48 hours after hospital admission comparedwith the APACHE II score at admission. ArchSurg 2002; 137:1136–1140

30. Dominguez-Munoz JE, Carballo F, GarciaMJ, et al: Evaluation of the clinical useful-ness of APACHE II and SAPS systems in theinitial prognostic classification of acute pan-creatitis: A multicenter study. Pancreas1993; 8:682–686

31. Eachempati SR, Hydo LJ, Barie PS: Severityscoring for prognostication in patients withsevere acute pancreatitis: Comparative anal-ysis of the Ranson score and the APACHE IIIscore. Arch Surg 2002; 137:730–736

32. Chatzicostas C, Roussomoustakaki M,Vlachonikolis IG, et al: Comparison of Ran-son, APACHE II and APACHE III scoringsystems in acute pancreatitis. Pancreas 2002;25:331–335

33. Johnson CD, Toh SK, Campbell MJ: Combi-nation of APACHE-II score and an obesityscore (APACHE-O) for the prediction of se-vere acute pancreatitis. Pancreatology 2004;4:1–6

34. Bradley EL III: Atlanta redux. Pancreas 2003;26:105–106

35. Imrie CW: Observations on acute pancreati-tis. Br J Surg 1974; 61:539–544

36. Ranson JHC: Acute pancreatitis. Curr ProbSurg 1979; 16:1–84

37. Mergener K, Baillie J: Acute pancreatitis.BMJ 1998; 316:44–48

38. Steinberg WM: Predictors of severity of acutepancreatitis. Gastroenterol Clin North Am1990; 19:849–861

39. Venkatesan T, Moulton JS, Ulrich CD II, et al:Prevalence and predictors of severity as de-fined by Atlanta criteria among patients pre-senting with acute pancreatitis. Pancreas2003; 26:107–110

40. Rowan KM: Outcome comparisons of inten-sive care units in Great Britain and Irelandusing the APACHE II method [DPhil thesis].Oxford, UK, University of Oxford, 1992

41. Flint R, Windsor JA: Early physiological re-sponse to intensive care as a clinically rele-vant approach to predicting the outcome insevere acute pancreatitis. Arch Surg 2004;139:438–443

42. Carnovale A, Rabitti PG, Manes G, et al: Mor-tality in acute pancreatitis: Is it an early or alate event? JOP 2005; 6:438–444

1708 Crit Care Med 2007 Vol. 35, No. 7