A neural network approach to the diagnosis of morbidity outcomes in trauma care

Artificial Intelligence in Medicine 15 (1999) 299–307

A neural network approach to the diagnosis ofmorbidity outcomes in trauma care

Robert P. Marble a,*, James C. Healy b,1

a College of Business Administration, Creighton Uni6ersity, 2500 California Plaza, Omaha,NE 68178-0308, USA

b Department of Pathology, Creighton Uni6ersity School of Medicine, 601 N. 30th Street, Omaha,NE 68131, USA

Received 30 March 1998; received in revised form 14 July 1998; accepted 18 August 1998

Abstract

This paper introduces the application of artificial neural networks to trauma complicationsassessment. The potential financial benefits of improving on trauma center diagnosticspecificity in complications assessment are illustrated and the operational feasibility of theuse of diagnostic neural models across institutions is discussed. A prototype neural networkmodel is described, which, after training, succeeds in diagnosing the complication of sepsis invictims of traumatic blunt injury. Its diagnostic performance with 100% sensitivity and 96.5%specificity is accomplished with test data from a regional trauma center. The model is furthershown to have correctly detected, during training, incorrectly coded data. The potential thissuggests, for parsimonious database scrubbing through the use of neural network models, isdiscussed. © 1999 Elsevier Science B.V. All rights reserved.

Keywords: Neural networks; Trauma care; Diagnostic decision aids; Morbidity outcomes;Complications assessment

* Corresponding author. Tel.: +1-402-2802215; Fax: +1-402-2802172; e-mail:[email protected].

1 Tel.: +1-402-4494630; Fax: +1-402-4495252; e-mail: [email protected]

0933-3657/99/$ - see front matter © 1999 Elsevier Science B.V. All rights reserved.

PII: S0933 -3657 (98 )00059 -1

R.P. Marble, J.C. Healy / Artificial Intelligence in Medicine 15 (1999) 299–307300

1. Introduction

The work described in this paper represents a preliminary address of thedesirability and the potential for application of artificial neural network models intrauma complications assessment. The work is motivated by the possible diagnosticand cost savings benefits that can result from an improved ability to recognizepatterns in trauma complications data. It utilizes the ability of neural networks torecognize complex and highly non-linear relationships, such as are likely tocharacterize medical circumstances.

A neural network model is elucidated in this paper, which succeeds in learning todiagnose the complication of sepsis from victims of blunt injury trauma. The modelwas constructed and trained with data elements that come from a standardrepresentation scheme and are regularly collected by regional trauma centers. It isintended that this will allow successful neural models to be applied universally.

After the beginning section gives a brief overview of the area of pursuit, thesecond section makes an economic case for applying neural networks to traumacomplications assessment. The third section discusses the prospects for neuralmodels’ applicability across institutions. The experimental model is described in thenext section and its results in the fifth section. The sixth section contains remarksabout future work in continuing the pursuits of the paper and it is followed byreferences.

2. Background

In 1991, Baxt [3] demonstrated the predictive reliability of artificial neuralnetwork models in medical diagnosis. He constructed a neural model with inputvariables selected from the presenting symptoms, the past history findings, and thephysical and laboratory findings of adult patients presenting to an emergencydepartment with anterior chest pain. This model improved on the predictivereliability of attending physicians markedly. The recent past has now witnessed amyriad of applications of neural models for prediction and diagnosis in variousfields of medicine, including oncology [10,12,23,32], radiology [13,21,31], andcardiology [14,19].

McGonigal et al. [25] have pioneered the application of neural network models inthe area of trauma scoring, by using such a model to estimate the probability ofsurvival for patients with penetrating trauma. In a carefully constructed comparisonstudy, he was able to show a significant improvement in sensitivity over the TRISS[6] and ASCOT [7] methods of survival prediction. While noting the increasedability of neural models to characterize the nonlinear behavior of biologic systems,McGonigal also pointed out that this was accomplished without requiring dataelements that other approaches don’t already utilize. Others have followed withvarious neural network models for trauma outcomes prediction. See, for instance,[11,16,20], and [28]. Rutledge [27] employed neural networks in the assessment ofinjury severity by using them to predict the probability of survival in trauma cases.

R.P. Marble, J.C. Healy / Artificial Intelligence in Medicine 15 (1999) 299–307 301

Also showing predictive improvement over standard methods, his study was aimedat overcoming the lack of available data for application of the standard methods inmany state and regional trauma centers.

While the work of McGonigal and others has shown the potential for successfuluse of artificial neural network models in the trauma care environment, itsorientation has been in the area of survival scoring and its motivation appears tobe in the area of extramural quality control assessment. Indeed, much recent traumacare outcome analysis work has concentrated in these areas. (See, for instance [29].)In fact, the Major Trauma Outcome Study, establishing national norms for traumacare [8] has concentrated on the patient outcomes of survival/death at discharge fromthe acute care hospital and length of stay in the intensive care unit and in the hospital.

The pursuit described in this paper is part of a program of study that extends theabove orientation to analysis of specific complication outcomes. Hoyt et al. [15] havefacilitated this orientation shift with their study of the incidence and types ofcomplication outcomes in trauma care. Their study was directed toward theidentification of thresholds for provider-related and disease-specific morbidity out-comes, in the interest of quality improvement. They have contributed an importantstandardized and well-defined categorization of complications for more specificoutcome analysis pursuits. The Committee on Trauma of the American College ofSurgeons has recognized the importance of expanding outcomes research beyond itspreviously limited scope of straight survival, to include the monitoring of all typesof morbidity [26].

The modeling effort described in this paper is intended to result eventually inrefined systems capable of functioning as aids to existing trauma care tools andheuristics in prehospital, acute care, and rehabilitation settings. The core of thesesystems will consist of various artificial neural network models intended forrecognition of patterns in trauma patient data routinely recorded in regional traumacenter TRACS [26] databases. Their contribution will be to reliably indicate thepresence or absence of various complication outcomes of the trauma care situationswith which they are presented.

3. Potential cost savings

A great potential exists for realizing cost-savings by utilizing this approach in thetrauma setting. It can be evinced by improving the clinical predictability that acomplication or morbid condition is unlikely to be present and thus preclude the useor initiation of an expensive workup or treatment strategy. When the pretest (orpre-evaluation) probability of disease is low, test specificity has the greatest effect onthe total number of positive test results, since most positive test results are falsepositive. We therefore wish to decrease the number of false positive determinationsresulting from the evaluation of the patient early in the course of their resuscitationand treatment following trauma. This has the effect of increasing the positivepredictive value of the evaluation. We are also assuming that our improvement inspecificity will not be coupled with a decrease in sensitivity.


The potential benefits of increased specificity might best be illustrated with anexample, considering the complication of sepsis. Savings would be realized fromprevention of initiation of antibiotic prophylaxis for suspected sepsis. Hoyt [15]reported that 1.9% of trauma patients in his institution developed sepsis. If weassume diagnostic sensitivity and specificity of 90%, and assuming that the patientssubsequently identified as having sepsis were the only patients with sepsis (diseaseprevalence of 1.9%) then approximately 11% of patients received antibiotics whenonly 1.9% were septic. Improving specificity by 5% would decrease the total numberof patients receiving antibiotics to about 6% (as shown in Table 1), thus decreasingoverall expenditures. Assuming a hospital charge of $100 per dose of antibioticsand 7 days of treatment with three doses per day, reducing the number of patientsreceiving antibiotics by 49 would save approximately $103 000 per 1000 traumapatients. Of course these figures are somewhat speculative and actual savings willdepend on local treatment patterns, but the potential savings are quite dramatic.

As Baxt has shown [2,3,5], neural networks can provide significant improvementin diagnostic accuracy when compared to clinicians judgements in the diagnosis ofmyocardial infarction. Neural networks also provide a significant improvement indiagnostic performance over other computer-based strategies [4]. In Baxt’s data,sensitivity improved from 88 to 92% and specificity improved from 71 to 96%.These results are important and relevant to our work since Baxt evaluated theperformance of his neural network utilizing input data similar to the pre-admissiondata available from TRACS. His data are also important in that both increasedsensitivity and specificity were achieved.

4. Universal use

To provide a scenario in which these savings could be universally realizable,models are constructed here that derive from commonly collected and availabledata. The general orientation of this work is towards models that are applicable

Table 1A 5% improvement in diagnostic specificity illustrated for 1000 patients

ActualActual Actual Actualnegative positivepositive negative

(19) (981)(19)(981)

17 17 49Diagnosed 98 Testedpositive (66)positive (115)

2 932Tested883Diagnosed 2negative (885) negative (934)

Sensiti6ity SpecificitySpecificitySensiti6ity(883/981)(17/19)Before (932/981)(17/19)After

95%89.47%90%89.47%


across institutions, rather than towards those whose utility will only exist inrestricted circumstances. This reflects (and drives) the increasing recognition oftrauma data registries as an important priority in health care. This need forcommon standards for collection and dissemination of trauma survival outcomesdata was well expressed by MacKenzie et al. [22] and Champion et al. [8]. The callwas expanded by Rice and Rutledge [26] to include trauma complication outcomesdata and its results have provided the basis for the research of this paper.

The commonly available data elements for the model of this research are drawnfrom the National TRACS design [26] and include data elements on demographics(13 elements), injury (11), prehospital findings (22), referring hospital findings (15),emergency department admission (19), emergency department treatment (13), hos-pital diagnoses (22), operations (36), quality assurance indicators (17), complica-tions (28), and outcomes (9). This structure, proposed by the Committee onTrauma of the American College of Surgeons, has become a standard for traumaregistries. It is available nationally with tools for its implementation and proceduresfor reporting data to state and central trauma registries [30].

Given that uniformity exists in data structure definitions between institutionsusing the TRACS design, it has been argued that nevertheless models developedwith data from one trauma registry may not lead to valid conclusions for a differentdata set [18]. Variables that are difficult to quantify, such as referral patterns orsurgical skill level, could affect patient outcomes without being explicitly repre-sented in the database. Lim and colleagues [20] cite aspirations for diagnosticdecision aids that can be adapted for differing operational conditions. They pointout that geographical and demographical conditions, as well as differing clinicalpractices from site to site, may confound the portability of such systems. This leadsus to the objective of universal model structures that draw from commonlyavailable data elements, but whose parameters can be freshly established for eachtrauma center to which they would be applied. Neural network models with aback-propagation learning algorithm represent the universal models suggested here.Their connection weights, newly trainable for every trauma center, represent theparameters.

5. The prototype model

A prototype neural model was constructed and evaluated [1], to make a starttoward the above described goals. The complication outcome investigated in themodel is that of sepsis. Sepsis was selected because of its impact on the length ofstay in the hospital of a trauma patient, as well as its effect on subsequentmorbidity and mortality outcomes. It is also more likely to be captured in thedatabase or on subsequent chart review. It more frequently results in symptoms andmultiple physical and laboratory findings will point to its presence in a patient.

The data were selected from among 912 trauma cases recorded between July,1994 and April, 1995 in the TRACS database of the Creighton University MedicalCenter. Preliminary filtering methods [24] resulted in a data set of 515 cases, 13 of


which (2.5%) were coded as having involved the complication of sepsis. Theindicator variable for sepsis represents the single target output for this back-propa-gation neural network model. Our clinical judgement led us to select 18 dataelements as neural model inputs with reasonable possibility of deterministic rele-vance to sepsis. The variables selected are: patient age; prehospital pulse rate,respiratory rate, systolic blood pressure, Glasgow Coma Score, and RevisedTrauma Score; emergency department temperature, pulse rate, respiratory rate,systolic blood pressure, Glasgow Coma Score, and Revised Coma Score; Hemat-ocrit and base deficit values (recorded from blood testing in the emergencydepartment); manual and derived injury severity scores; functional independencemeasure score; and a hospital referral indicator variable.

All non-indicator input variables were normalized to a scale with extremes of −4and 4. The first 256 cases were used for supervised network training; the remaining259 cases were reserved as a test set. The training set contains nine of the 13 sepsiscases. The prototype model (with 18 input nodes and one output node) containstwo hidden layers of ten and four nodes, respectively.

6. Results

The model trained and tested well. Using sigmoid transfer functions with upperand lower limits of 1 and −1 and a constant learning rate of 0.2, the networktrained to a training set mean squared error of 0.005 with some 1700 training setpresentations. While the trained network still failed to recognize one training caseinvolving sepsis (for a sensitivity of 89%), it successfully learned all 247 non-sepsistrauma cases of the training set. When presented with the test set, the networkrecognized all four sepsis cases (for 100% sensitivity in untrained cases) and issuedonly nine false positive conclusions, evincing a specificity of 96.5%. Table 2 showsthese results.

Table 2Neural network sensitivity and specificity

ActualActualActual Actualpositivepositive negative negative(4) (255)(9) (247)

Tested8 4 9Trained 0positive (13)positive (8)

0 246Tested247Trained 1negative (248) negative (246)

Sensiti6ity Sensiti6ity SpecificitySpecificity(247/247)(8/9)Training set (246/255)(4/4)Test set

96.5%100%100%89%


The results indicate that there is promise for using neural networks in the traumacomplications context. A necessary step in the further evaluation of this modelwould be a thorough chart review for all trauma cases used. Information in thecomplete medical records for these cases may lead to some recoding of data. Inaddition to corrections in interpretive entries by emergency department personnel,the charts will also help determine whether the clinical team recognized sepsis in allcases.

As a first step toward such evaluating, the training case involving sepsis whichthe model never recognized was selected for review. Medical records for this casewere reviewed and yielded a conclusion that sepsis was not a clinical diagnosis atany time during the patient’s hospitalizations. The case had been miscoded. Thecoding error was corrected in the training data set and the model was retrained.

The model with revised training set trained to a mean squared error of 0.000589after only 59 training set presentations. This is particularly startling, becausevarious replications with the original training data set, while always successful,would sometimes require as many as 5000 training set presentations (depending onthe effects of the fresh connection weight randomizations). With corrected trainingdata, the model never required more than 70 training set presentations to train tothe 0.0005 MSE level after re-randomization of weights.

Further training of the model to MSE levels of 0.000169, 0.0000208, and0.0000124 required total training set presentations of 79, 219 and 788, respectively.Beyond this level, the network seemed to stabilize.

With this vast improvement in training time, no change in the 100% sensitivityand the 96.5 specificity, with respect to the test data set, was evinced. It appearsthat the original model had successfully generalized the patterns in the data, evenwith a coding error present in those data.

7. Remarks

An interesting question arises regarding the integrity of the data recorded intrauma registries. While some researchers have expressed a view that these costlyand resource-intensive efforts have not led to useful results, there exists a generalconsensus about the high value of trauma information [9]. Problems and limitationswith trauma registries often seem to center around the interpretive coding necessaryto populate them. Rutledge [27] notes that a medically knowledgeable person mustconduct the coding for each trauma case and that this coding requires a lot of time.Unfortunately, as Shapiro and colleagues point out [30], this is happening in acontext of health care reform, characterized more by cutbacks than by additionalresource outlays. Jones [17] adds to concerns about the problems with maintainingsuch ambitious data repositories, by expressing the commonly held view that priorto data analysis based on their contents, extensive checking is routinely necessary‘to remove all inaccuracies and inconsistencies’.

If regional data are to be used in models such as that of this paper, nontrivialconsideration must be given at each locale to controlling for the quality of data.


This might otherwise introduce prohibitive costs, since medical personnel arerequired for checking data. If, however, neural models can flag the cases that maymost deserve checking—as the work of this paper indicates—perhaps the need forexhaustive data checking can be obviated. A possible approach to achieving theintended results would be to use the local training of a standard neural networkarchitecture to scrub the data. Cases about which cycling occurs during training orcases that are otherwise suspect from the results of training would be candidates forchart review. Then the finally trained neural model, with weights that reflect localpatterns, would be applied to the evaluation of new trauma cases. Periodicadditional training could also be undertaken. The project surrounding the proto-type reported here will need to be continued with data from another regionaltrauma center, as its components receive further scrutiny.

References

[1] Bailey DL, Thompson DM. How to develop neural-network applications. AI Expert 1990;5:38–47.[2] Baxt WG. Use of an artificial neural network for data analysis in clinical decision-making: the

diagnosis of acute coronary occlusion. Neural Comput 1990;2:480–9.[3] Baxt WG. Use of an artificial neural network for the diagnosis of myocardial infarction. Ann Intern

Med 1991;115:843–8.[4] Baxt WG. Complexity, chaos and human physiology: the justification for non-linear neural

computational analysis. Cancer Lett 1994;77:85–93.[5] Baxt WG, Skora J. Prospective validation of artificial neural network trained to identify acute

myocardial infarction. Lancet 1996;347:12–5.[6] Boyd CR, Tolson MA, Copes WS. Evaluating trauma care: the TRISS method. J Trauma

1987;27:370–8.[7] Champion HR, Copes WS, Sacco WJ, et al. A new characterization of injury severity. J Trauma

1990;30:539–46.[8] Champion HR, Copes WS, Sacco WJ, et al. The major trauma outcome study: establishing national

norms for trauma care. J Trauma 1990;30:1356–65.[9] Champion HR, Copes WS, Sacco WJ. Injury severity scoring again. J Trauma 1995;38:94–5.

[10] De Laurentiis M, Ravdin PM. A technique for using neural network analysis to perform survivalanalysis of censored data. Cancer Lett 1994;77:127–38.

[11] Dombi GW, et al. Prediction of rib fracture injury outcome by an artificial neural network. JTrauma 1995;39:915–21.

[12] Downs J, Harrison RF, Kennedy RL, Cross SS. Application of the fuzzy ARTMAP neural networkmodel to medical pattern classification tasks. Artif Intell Med 1996;8:403–28.

[13] Hamilton D, Miola UJ, Mousa D. Interpretation of captopril transplant renography using a feedforward neural network. J Nucl Med 1996;37:1649–52.

[14] Heden B, Edenbrandt L, Haisty WK, Pahlm O. Artificial neural networks for the electrocardio-graphic diagnosis of healed myocardial infarction. Am J Cardio 1994;74:5–8.

[15] Hoyt DB, Hollingsworth-Fridlund P, Fortlage D, et al. An evaluation of provider-related anddisease-related morbidity in a level I university trauma service: directions for quality improvement.J Trauma 1992;33:586–601.

[16] Izenberg SD, Williams MD, Luterman A. Prediction of trauma mortality using a neural network.Am Surg 1997;63:275–81.

[17] Jones JM. An approach to the analysis of trauma data having a response variable of death orsurvival. J Trauma 1995;38:123–8.


[18] Jones JM, Redmond AD, Templeton J. Uses and abuses of statistical models for evaluating traumacare. J Trauma 1995;38:89–93.

[19] Kennedy RL, et al. An artificial neural network system for diagnosis of acute myocardial infarction(AMI) in the accident and emergency department: Evaluation and comparison with serum myo-globin measurements. Comput Methods Programs Biomed 1997;52:93–103.

[20] Lim CP, Harrison RF, Kennedy RL. Application of autonomous neural network systems tomedical pattern classification tasks. Artif Intell Med 1997;11:215–39.

[21] Lo JY, et al. Predicting breast cancer invasion with artificial neural networks on the basis ofmammographic features. Radiology 1997;203:159–63.

[22] MacKenzie EJ, Morris JA, Smith GS, Fahey M. Acute hospital costs of trauma in the UnitedStates: implications for regionalized systems of care. J Trauma 1990;30:1096–101.

[23] Mango LJ. Computer-assisted cervical cancer screening using neural networks. Cancer Lett1994;77:155–62.

[24] Marble RP, Healy JC, Johnson TJ. Neural network models for improving diagnostic specificity intrauma complications assessment, Proceedings of the 26th Annual Meeting of the Decision SciencesInstitute (Boston, 1995), pp. 561–563.

[25] McGonigal MD, Cole J, Schwab CW, et al. A new approach to probability of survival scoring fortrauma quality assurance. J Trauma 1993;34:863–70.

[26] Rice CL, Rutledge R. Trauma registry. American College of Surgeons-Committee on Trauma,Resources for Optimal Care of the Injured Patient, 1993:97–101.

[27] Rutledge R. Injury severity and probability of survival assessment in trauma patients using apredictive hierarchical network model derived from ICD-9 codes. J Trauma 1995;38:590–601.

[28] Selker HP, et al. A comparison of performance of mathematical predictive methods for medicaldiagnosis: Identifying acute cardiac ischemia among emergency department patients. J Investig Med1995;43:468–76.

[29] Shackford SR, Mackersie RC, Hoyt DB, et al. Impact of a trauma system on outcome of severelyinjured patients. Arch Surg 1987;122:523–7.

[30] Shapiro MJ, Cole KE, Keegan M, Prasad CN, Thompson RJ. National survey of state traumaregistries—1992. J Trauma 1994;37:835–42.

[31] Tourassi GD, Floyd CE, Sostman HD, Coleman RE. Acute pulmonary embolism: artificial neuralnetwork approach for diagnosis. Radiology 1993;189:555–8.

[32] Wilding P, Morgan MA, Grygotis AE, et al. Application of backpropagation neural networks todiagnosis of breast and ovarian cancer. Cancer Lett 1994;77:145–53.

.

Documents

A neural network approach to the diagnosis of morbidity outcomes in trauma care