Eur J Vasc Endovasc Surg 19, 184–189 (2000)

doi:10.1053/ejvs.1999.0974, available online at http://www.idealibrary.com on

Ruptured Abdominal Aortic Aneurysm: a Novel Method of OutcomePrediction Using Neural Network Technology∗

E. P. L. Turton†1, D. J. A. Scott1, M. Delbridge1, S. Snowden2 and R. C. Kester1

1Departments of Vascular and Endovascular Surgery, and 2Medical Physics, St James’s University Hospital,United Leeds Teaching Hospitals Trust, Leeds, LS9 7TF, U.K.

Background: reported survival following emergency surgery for ruptured abdominal aortic aneurysm (RAAA) varieswidely between institutions. This is largely attributable to differences in case mix. The aim of this study was to identifyand evaluate a set of prognostic variables that would accurately predict outcome for individual patients from perioperativeindices.Methods: perioperative factors associated with subsequent mortality at our institution were identified by retrospectivereview of 102 consecutive operations for RAAA over a 7-year period (January 1990 to January 1997). Logistic regressionanalysis was used to select the most significant variables associated with subsequent mortality. These were used toconstruct, train, and validate a neural network designed to predict survival from surgery in individual cases on aprospective basis.Results: the 30-day mortality rate was 53%. Multivariate analysis identified four highly significant independent predictorsof mortality; preoperative hypotension, intraperitoneal rupture, preoperative coagulopathy, and preoperative cardiac arrest.Using these inputs, the neural network correctly predicted outcome in 82.5% of individual cases.Conclusion: a neural network based on just four perioperative variables can accurately predict outcome of RAAA.Prognostic variables should be reported in studies as a measure of the effect of case mix on survival data. Neural networkshave potential to aid decision-making relating to outcome for individual cases.

Introduction systems used to predict mortality in this setting, such

as POSSUM and APACHE-II, have not been found to

The incidence of RAAA is increasing, predominantly be useful tools to predict outcome in individual cases.9

Although other studies have identified prognostic fac-attributable to an ageing population.1,2 Surgery or

endoluminal repair currently represents the only tors associated with survival after RAAA,10–21 it is still

not clear which are most useful.effective treatment. Elective aortic aneurysm repair is

associated with the best outcome and mortality has The aim of our study was to establish the most

important predictors of mortality following surgery forfallen from approximately 20% forty years ago to less

than 5% today.3 The operative mortality for patients ruptured AAA, and to use these variables to construct a

neural network to validate their prognostic importancewith rupture is much more variably reported at be-

tween 14 and 70%.2,4,5 Survival data may be biased in in individual cases.

some series by differential reporting of cases and a

process of selective surgery to avoid operating on

individual patients felt to have a poor chance of sur-

vival.4,6–8 Artificial Neural NetworksThe evaluation of the effect of case-mix on reported

survival data is evidently significant, but dependent An artificial neural network is a form of computeron the identification of the most important prognostic intelligence that is modelled after biologic neural sys-variables associated with mortality. Standard scoring tems, and may be implemented as a computer software

program. It is capable of identifying relations in input

data that are not easily apparent with current common∗ Presented in part at the meeting of the Association of Surgeons analytic techniques and therefore have certain ad-of Great Britain and Ireland, Edinburgh, May 1998.

vantages over standard regression analysis.22 They† Please address all correspondence to: E. P. L. Turton, Ash House,Main Street, Barkston Ash, North Yorkshire LS24 9PR, U.K. have been the subject of increasing research interest

1078–5884/00/020184+06 $35.00/0 2000 Harcourt Publishers Ltd.

AAA: a Novel Method of Outcome Prediction 185

Table 1. Variables tested as predictors of mortality within 30 daysof surgery.

Variable Predictor

Age7,10,14,33–36 >76 years

Male gender7,37 Yes

Systolic BP10,33,34,36–38 <91 mmHg

Haemoglobin14,38 <8 g/dl

Haematocrit7,17,37 <25%

Platelet count9,20 <100×109/l

APTT20 >60 seconds

PT20 >18 seconds

Amylase39 >340 units/lFig. 1. The inputs I1−x, are weighted by a factor W1−x, andsummated to give a specific value. A transfer function modifies this Creatinine10,14,38 >300 lmol/lto produce a single output. I1−x=input one to infinity W1−x= Preoperative cardiac arrest7,40,41 Yesweighting factor one to infinity.

Type of rupture21,34,42,43 Intraperitoneal

PT=prothrombin time; PCV=packed cell volume; BP=blood pres-sure; APTT=activated partial thromboplastin time.

in the last 10 years, most extensively for problems in

engineering, but applied more recently to medical

problems.23–28

the outcomes are known. These are presented to theA neural network consists of an arrangement of

network, record by record, over a (potentially veryprocessing elements (PEs) each of which com-

large) number of iterations. The weights in themunicates with one or more others by means of

network are updated by the learning algorithm soweighted connections. A typical processing element

as to minimise the overall error. Training ceaseshas several inputs that are ‘‘weighted’’ by a process

when the overall error has declined maximally. Theof summing and thresholding to produce a single

accuracy of the network to predict outcome correctlyoutput (Fig. 1). Neural networks are capable of being

is then assessed using a different dataset and com-trained to discern patterns and non-linear associations

paring the output of the trained network with thein data. Once trained, a network is ‘‘generalisable’’

known result.and can be used to classify or predict outcomes

from similar data.29 Fundamentally, the behaviour

(i.e. the output) of a biological neurone in response

to a set of inputs depends upon the strengths of Patients and Methodsthe connections (neurotransmitters) between the cell

body and the inputs. In an analogous manner the Data were obtained from a retrospective review of

102 consecutive patients who underwent emergencyoutput of a PE depends on the individual weighting

factor on each input. The behaviour of the network surgery for ruptured abdominal aortic aneurysm at

our institution over a seven-year period. The medicalas a whole can therefore be determined by ap-

propriate fixing of every weight in the network. It records of 13 patients were not available for review.

Data were also excluded on three patients withis necessary to have a means of setting the weights

(“training the network”). In a back-propagation net- inflammatory aneurysms, one patient with aorto-

enteric fistula, and one with an iliac aneurysm. Awork the weights are set by repeatedly presenting

a set of data records to the input layer for which transperitoneal approach was used for all patients.

Rupture was classified as intraperitoneal if there wasthe desired output is known. The error between the

output produced by the network and the desired free blood in the peritoneal cavity, and retroperitoneal

as indicated by the presence of retroperitoneal haem-(known) output for the corresponding record in the

dataset is used to adjust the weights in the network atoma. Twelve perioperative indices of mortality were

chosen for analysis, based on previously identifiedaccording to some learning algorithm, so as to reduce

the error. The basis of this learning method was significant variables in the literature and the data

available from our review (Table 1). The systolicestablished by Rumelhart et al.30 After many cycles

through the training data the weights will converge blood pressure and haematological variables recorded

on arrival to the hospital were used in the analysis.to a set which represents a minimum error condition

over the entire dataset. Records where values of a variable were unrecorded

were excluded from the analysis of that variable.The network is trained using a dataset for which

Eur J Vasc Endovasc Surg Vol 19, February 2000

E. P. L. Turton et al.186

Table 2. Neural network: coding of input and output variables. size from 2 to 5 processing elements (PEs), were as-

sessed. The network with 3 PEs in the hidden layerVariable Codegave us best results. The network was trained from

Input five different initial weight randomisations. The resultsBP (<91=1, else 0) of the best network are presented.Pre-op coagulopathy (Platelets <100×109=1, else 0)

Pre-op cardiac arrest (True=1, else 0)

Intraperitoneal rupture (True=1, else 0)

Output ResultsDeath 1

Survival 0One hundred and two patients were included for

Pre-op=preoperative; BP=systolic blood pressure. analysis in the study (70 men, 32 women; median

age 78.8, range 61 to 102 years). Forty-eight patientsAssociations between preoperative and perioperative survived to 30 days post surgery (overall mortalityrisk factors and subsequent mortality within 30 days rate 53%).were analysed by univariate and multivariate binary Univariate analysis identified 10 perioperative vari-logistic regression using a commercial statistics pack- ables that were significantly associated with mortality,age (MINITABTM for Windows, version 11.2; MINITAB comparing favourably with the current surgical lit-Inc, 3081 Enterprise Drive, State College, PA 16801- erature (Table 3). Of these, a low packed cell volume,3008, U.S.A.). hypotension, anaemia, thrombocytopenia and intra-

Regression coefficients were used to develop a multi- peritoneal rupture, each had a highly significant as-variate model based on the most significant risk factors. sociation (p<0.001) with subsequent mortality. MaleThe neural network was constructed using Autonet gender and a prolonged prothrombin time (>18 sec-WindowsTM software on a PentiumTM 100 personal onds) showed no significant association.computer with 32 megabytes of random access mem- Three variables were excluded from multivariateory (RAM). The database of 102 patients’ records analysis due to incomplete data for that variable:(for which the outcomes were known) were divided preoperative amylase not recorded in 51 patients; pro-randomly into two sets of 51 records, a training set thrombin time or activated partial thromboplastin time(29 deaths, 22 survivors) and a test set (25 deaths, 26 not recorded in 41 patients. Multivariate analysis con-survivors). firmed a significant independent association with mor-

It is normal practice to use three datasets. The first tality for just 4 of the perioperative variables significantis used to train the network (training set). The second on univariate analysis (Table 4). A preoperative cardiacis used to validate the network (validation set). Lastly, arrest was the most significant independent variablethe performance of the network was assessed by pre- associated with the highest odds ratio (OR) of dyingsenting it with the third dataset, and comparing the (OR=15.06, p=0.017). The other most significant in-output of the trained network with the known result dependent variables predictive of subsequent mor-(DTI Best Practice Guidelines). Training is repeated tality were the occurrence of an intraperitoneal bleedusing a variety of network layouts until the best per- (OR=13.46, p=0.002), preoperative thrombocytopeniaformance on the validation set is achieved. The training (OR=9.04, p=0.002) and compromised blood pressureset was accordingly split randomly into 60% training, on arrival (OR=3.89, p=0.042).40% validation. Using these four variables, the neural network

The records were converted from the original data- correctly predicted individual survival, on the testbase format into comma-delimited files suitable for dataset of 51 patients, with a sensitivity of 86.4%importing into the neural network software. The inputs and specificity of 79.3%. The positive and negativeto the network were selected on the basis of the predictive values of the neural network were 82.6%most significant independent predictors of survival and 88.5%, respectively.identified by multivariate analysis. These comprised

systolic blood pressure on arrival (>91 mmHg, or

<91 mmHg), thrombocytopenia (platelet count

<100×109/l, or >100±109/l), preoperative cardiac ar- Discussionrest (yes, or no) and type of rupture (intraperitoneal,

or retroperitoneal) (Table 2). The output comprised Many factors currently influence the widely reported

survival data following surgery for RAAA.31 Outcometwo classes – survival (at 30 days) or death.

Several single hidden layer networks, ranging in is, in part, determined by the patient’s preoperative

Eur J Vasc Endovasc Surg Vol 19, February 2000

AAA: a Novel Method of Outcome Prediction 187

Table 3. Univariate association with mortality on logistic regression (data for missing variables not analysed).

Variable Patients analysed p-Value OR 95% CI

Male Gender 102 0.688 1.19 0.51–2.74PT >18 seconds 63 0.499 1.41 0.52–3.84Age >76 years 102 0.013 2.86 1.24–6.56Creatinine >300 lmol/l 102 0.076 3.41 0.88–13.22PCV <25% 102 0.000 7.89 3.00–20.74BP systolic <91 mmHg 102 0.000 9.46 3.83–23.35Amylase >340 units/l 56 0.004 10.50 2.10–52.58Haemoglobin <8 g/dl 102 0.000 10.75 3.69–31.35Preoperative cardiac arrest 102 0.002 11.50 2.50–52.82Platelet count <100×109/l 102 0.000 15.77 5.90–42.12Intraperitoneal rupture 102 0.000 18.70 5.84–59.84APTT >60 seconds 61 0.001 34.00 4.03–286.66

PT=prothrombin time; PCV=packed cell volume; BP=blood pressure; APTT=activated partial thromboplastin time; OR=odds ratio;CI=confidence interval.

Table 4. Multivariate model of variables independently associated with mortality.

Variable p-Value OR 95% CI

Creatinine >300 lmol/l 0.760 1.33 0.21–8.35PCV <25% 0.682 1.37 0.30–6.15Age >76 years 0.505 1.58 0.41–6.01Haemoglobin <8 g/dl 0.372 2.07 0.42–10.19Male gender 0.211 2.72 0.57–13.06BP systolic <91 mmHg 0.042 3.89 1.05–14.42Platelet count <100×109/l 0.002 9.04 2.20–37.22Intraperitoneal rupture 0.002 13.46 2.52–71.95Preoperative cardiac arrest 0.017 15.06 1.61–140.93

PCV=packed cell volume; BP=blood pressure.

status, the surgeon, and occurrence of postoperative operation notes of 13 patients in the study. Ad-

ditionally, although no policy existed to exclude anycomplications.11,17,18 At present both referral patterns

and the proportion of cases accepted for operation patients with RAAA from operation during the study

period, it is not possible to identify if this ever occurred.differ between hospitals.8 The subsequent reporting of

survival data for purposes of comparative audit are We would hope that a prospective study would im-

prove the accuracy of individual prediction by theconsequently strongly influenced by the case-mix.4

Many perioperative variables may have a strong in- neural network even further.

Three of these variables are available preoperativelyfluence on patient outcome, but there has long been a

need for these to be both identified and validated. and the presence of intraperitoneal rupture at lap-

arotomy. This type of system could be re-trained toThe knowledge of the functioning artificial neural

network is built on learning and experience from utilise variables available only in the accident and

emergency room, to offer a very rapid prediction ofhistorical data. Based on this prior knowledge, the

artificial neural network can predict non-linear re- outcome in the emergency situation. Future pro-

spective validation of the neural network is now beinglations found in newly presented datasets and offer

an advantage over simple logistic regression analysis.32 performed, to define its role in these different areas.

These may include deciding whether to proceed withWe have shown that outcome following surgery for

RAAA can be predicted for an individual patient, surgery, or on the appropriate use of an intensive care

unit bed postoperatively, in patients predicted not tousing a limited amount of perioperative data, with a

high accuracy. However, this was a retrospective re- survive.

In conclusion, we have confirmed and validatedview and subject to possible bias being introduced to

the neural network. We were unable to trace the the highly significant independent association of 4

Eur J Vasc Endovasc Surg Vol 19, February 2000

E. P. L. Turton et al.188

aortic aneurysms – prognostic indicators and complications af-perioperative variables with mortality followingfecting mortality. A local experience. South African Journal of

RAAA using an artificial neural network. These vari- Surgery 1995; 33: 21–25.ables should at least be reported in future studies as 17 Katz SG, Kohl RD. Ruptured abdominal aortic aneurysms. A

community experience. Arch Surg 1994; 129: 285–290.an indication of case-mix and may have a future18 Meesters RC, van der Graaf Y, Vos A, Eikelboom BC. Rup-

application in patient selection.tured aortic aneurysm: early postoperative prediction of mor-tality using an organ system failure score. Br J Surg 1994; 81:512–516.

19 Bauer EP, Redaelli C, von Segesser LK, Turina MI. Rupturedabdominal aortic aneurysms: predictors for early complicationsand death. Surgery 1993; 114: 31–35.Acknowledgement

20 Davies MJ, Murphy WG, Murie JA et al. Preoperative co-agulopathy in ruptured abdominal aortic aneurysm predicts

The authors would like to acknowledge the support of the Yorkshirepoor outcome. Br J Surg 1993; 80: 974–976.Vascular Surgical Research Fund for this project. This work was

21 Rantakokko V, Havia T, Inberg MV, Vanttinen E. Abdominalundertaken by St James’s University Hospital, which receives fund-aortic aneurysms: a clinical and autopsy study of 408 patients.ing from the NHS Executive; the views expressed are those of theActa Chirurgica Scandinavica 1983; 149: 151–155.authors and not necessarily those of the Executive.

22 Kattan MW, Beck JR. Artificial neural networks for medicalclassification decisions. Arch Pathol Lab Med 1995; 119: 672–677.

23 Bishop JB, Szpalski M, Ananthraman SK, McIntyre DR, PopeMH. Classification of low back pain from dynamic motioncharacteristics using an artificial neural network. Spine 1997; 22:

References 2991–2998.24 Tourassi GD, Floyd CE, Coleman RE. Acute pulmonary em-

bolism: cost-effectiveness analysis of the effect of artificial neural;1 Reitsma JB, Pleumeekers HJ, Hoes AW et al. Increasing in-networks on patient care. Radiology 1998; 206: 81–88.cidence of aneurysms of the abdominal aorta in The Netherlands.

25 An CS, Petrovic LM, Reyter I et al. The application of imageEur J Vasc Endovasc Surg 1996; 12: 446–451.analysis and neural network technology to the study of large-2 Thomas PR, Stewart RD. Abdominal aortic aneurysm. Br J Surgcell liver-cell dysplasia and hepatocellular carcinoma. Hepatology1988; 75: 733–736.1997; 26: 1224–1231.3 Ernst CB. Abdominal aortic aneurysm. N Engl J Med 1993; 328:

26 Gross GW, Boone JM, Bishop DM. Pediatric skeletal age: de-1167–1172.termination with neural networks. Radiology 1995; 195: 689–695.4 Callam MJ, Haiart D, Murie JA, Ruckley CV, Jenkins AM.

Ruptured aortic aneurysm: a proposed classification. Br J Surg 27 Tourassi GD, Floyd CE, Sostman HD, Coleman RE. Artificial1991; 78: 1126–1129. neural network for diagnosis of acute pulmonary embolism:

5 Chen JC, Hildebrand HD, Salvian AJ, Hsiang YN, Taylor DC. effect of case and observer selection. Radiology 1995; 194: 889–893.Progress in abdominal aortic aneurysm surgery: four decades of 28 Bottaci L, Drew PJ, Hartley JE et al. Artificial neural networksexperience at a teaching center. Cardiovascular Surgery 1997; 5: applied to outcome prediction for colorectal cancer patients in150–156. separate institutions. Lancet 1997; 350: 469–472.

6 Bradbury AW, Makhdoomi KR, Adam DJ et al. Twelve-year 29 Beale R, Jackson T. Neural Computing – An Introduction. Bristolexperience of the management of ruptured abdominal aortic IOP, 1992.aneurysm. Br J Surg 1997; 84: 1705–1707. 30 Rumelhart DE, Hinton GE, Williams RJ. Learning internal

7 Johansen K, Kohler TR, Nicholls SC et al. Ruptured abdominal representations by error propagation. Parallel Distributed Pro-aortic aneurysm: the Harborview experience. J Vasc Surg 1991; cessing Vol 1. MIT Press. 1986.13: 240–247. 31 Rutledge R, Oller DW, Meyer AA, Johnson GJ, Jr. A statewide,

8 Hewin DF, Campbell WB. Ruptured aortic aneurysm: the de- population-based time-series analysis of the outcome of rupturedcision not to operate. Ann R Coll Surg Engl 1998; 80: 221–225. abdominal aortic aneurysm. Ann Surg 1996; 223: 492–503.

9 Lazarides MK, Arvanitis DP, Drista H, Staramos DN, Day- 32 Dorsey SG, Waltz CF, Brosch L et al. A neural network modelantas JN. POSSUM and APACHE II scores do not predict the for predicting pancreas transplant graft outcome. Diabetes Careoutcome of ruptured infrarenal aortic aneurysms. Ann Vasc Surg

1997; 20: 1128–1133.1997; 11: 155–158.

33 van Dongen HPA, Leusink JA, Moll FL, Brons FM, de Boer10 Koskas F, Kieffer E. Surgery for ruptured abdominal aortic

A. Ruptured abdominal aortic aneurysms: factors influencinganeurysm: early and late results of a prospective study by the

postoperative mortality and long-term survival. Eur J Vasc Endo-AURC in 1989. Ann Vasc Surg 1997; 11: 90–99.

vasc Surg 1998; 15: 62–66.11 Sayers RD, Thompson MM, Nasim A et al. Surgical management

34 Martinez R, Garces D, Podeur L et al. Ruptured abdominalof 671 abdominal aortic aneurysms: a 13 year review from aaortic aneurysm. A ten year experience. J Cardiovasc Surg 1997;single centre. Eur J Vasc Endovasc Surg 1997; 13: 322–327.38: 1–6.12 Chen JC, Hildebrand HD, Salvian AJ et al. Predictors of death

35 Berridge DC, Chamberlain J, Guy AJ, Lambert D. Prospectivein nonruptured and ruptured abdominal aortic aneurysms. Jaudit of abdominal aortic aneurysm surgery in the northernVasc Surg 1996; 24: 614–620.region from 1988 to 1992. Northern Vascular Surgeons Group.13 Darling RC, 3rd, Cordero JA, Jr, Chang BB et al. AdvancesBr J Surg 1995; 82: 906–910.in the surgical repair of ruptured abdominal aortic aneurysms.

36 Lambert ME, Baguley P, Charlesworth D. Ruptured ab-Cardiovascular Surgery 1996; 4: 720–723.dominal aortic aneurysms. J Cardiovasc Surg 1986; 27: 256–261.14 Hardman DT, Fisher CM, Patel MI et al. Ruptured abdominal

37 Gloviczki P, Pairolero PC, Mucha P, Jr et al. Ruptured ab-aortic aneurysms: who should be offered surgery? J Vasc Surgdominal aortic aneurysms: repair should not be denied. J Vasc1996; 23: 123–129.Surg 1992; 15: 851–857.15 Jaakkola P, Hippelainen M, Oksala I. Infrarenal aortofemoral

38 Halpern VJ, Kline RG, D’Angelo AJ, Cohen JR. Factors thatbypass surgery: risk factors and mortality in 330 patients withaffect the survival rate of patients with ruptured abdominalabdominal aortic aneurysm or aortoiliac occlusive disease. Annaortic aneurysms. J Vasc Surg 1997; 26: 939–945.Chir Gynaecol 1996; 85: 28–35.

16 Browning NG, Long MA, Barry R et al. Ruptured abdominal 39 Bagley JS, Tyler MPH, Cooper GG. Hyperamylasaemia in

Eur J Vasc Endovasc Surg Vol 19, February 2000

AAA: a Novel Method of Outcome Prediction 189

ruptured aortic aneurysm: incidence and prognostic im- 42 Subramaniam P, Bennett RC, Campbell IA. Infrarenal aorticaneurysm surgery in a rural surgical service: risk factors forplications. Br J Surg 1994; 81: 31–32.mortality. Aust NZ J Surg 1998; 68: 25–28.40 Murphy JL, Barber GG, McPhail NV, Scobie TK. Factors

43 Miani S, Mingazzini P, Piglionica R, Biasi GM, Ruberti U.affecting survival after rupture of abdominal aortic aneurysm:Influence of the rupture site of abdominal aortic aneurysms witheffect of size on management and outcome. Canadian Journal ofregard to postoperative survival rate. J Cardiovasc Surg 1984; 25:Surgery 1990; 33: 201–205.414–419.31 Crew JR, Bashour TT, Ellertson D, Hanna ES, Bilal M.

Ruptured abdominal aortic aneurysms: experience with 70 cases.Clinical Cardiology 1985; 8: 433–436. Accepted 24 August 1999

Eur J Vasc Endovasc Surg Vol 19, February 2000