AN ANATOMICAL INDEX IN BLUNT TRAUMA

Abstract : The case histories of 2,135 patients with acute blunt trauma, consecutively admitted to a hospital over a 48-month period, were studied in attempt to produce an objective scale for the severity of injury. All injuries were coded using the Hospital Adaptation of International Classification of Disease (H-ICDA) prior to computer analysis. An 'effective' probability of survival (PE) for each injury code was tested for its ability to predict outcome on random subsets of patients. When used as an injury 'score' the PE can predict survival rates to within + or - with individual misclassification rates of 12%.


CONTENTS
In tho Bioresponse to Trauma Research Program of the USA Armament Command, a new methodology is evolving for assessing wound severity from blunt and penetrating injuries.The rationale consists of characterizing injuries by strings of "diagnostic codes."The strings are correlated with mortality using similarly coded data on traumatic injuries from US Army and civilian data bases.A report documenting the methodology and its applications to the Bioresponse to Trauma Research Program is in preparation.This report discusses the development and validation of an anatomical index for multiplh.:lunt trauma injuries.
The methodology has a wide range of potential military and civilian applications including validation of other injury assessment methods (such as the use of medical assessors), triage of patients, and evaluation of health care.
The present impetus to improve the management of trauma victims has provoked several attempts to develop a system for quantitating injury.The difficulties of characterizing a miscellany of injured patients are well recognized.Paradoxically, injury is more susceptible to quantitation than most disease processes because,, despite potential complications, the inflicted injury is not progressive, and the anatomical disruption provides a static data base.The elements of anatomical quantitation which remain problematic include an agreement on definitions for labeling, assignment of a scoring system, and the synergistic effect of multiple injuries.
Factors which alter with time and influence the prognosis can be encompassed by the term "physiological response" to injury, initially described by Cuthbertsonl but here used to include all metabolic and physiological responses to acute trauma.Such variables reflect not only the severity of the total trauma and the time elapsed since injury, but also the patient's age and pre-or co-existing diseases, both of which may affect the response to injury and the eventual outcome.
Existing quantitative systems have employed the degree of anatomical injury, 2 -4 elements of the physiological and biochemical response, 5 -7 and combinations of the two. 8' 9 The anatomical approach usually involves the arbitrary assignment of numbers to a subjective evaluation of the severity of the injury.The Abbreviated Injury Scale (AIS) 4 is a ranking of injuries by their severity, and is used internationally by researchers, including multi-disciplinary accident investigation teams established by the United States Department of Transportation.The injury grades were based on an arbitrary scale developed by approximately 50 physicians, engineers, and researchers.No verification has established that a "3" assigned to urethral or pericardial contusion is truly equivalent to that "3" attached to a hemothorax, or that the number is meaningfully relative to the "5" of a tracheal avulsion.The methodology has been extended to include the additive effect of multiple injuries by using a quadratic equation which correlates with actual mortality figures.10The AIS has also been used to evaluate the Comprehensive Injury Scale 1 1 which includes estimates of energy dissipated, degree of impairment, and other factors not previously included in injury quantitation.In brief, the AIS is ;m alternative total assessment to that based on clinical judgment and derives from a consensual sibjective assessment of anatomical injury on an arbitrary scale.It is a useful but limited tool when precise anatomical diagnosis is available either through surgery or postmortem examination.
Other attempts to quantify trauma have been specifically directed towards triage.The Trauma Index described by Kirkpatrick and Youmens 7 combines superficial anatomical assessment with some measure of physiological response in the form of pulse, blood pressure, cyanosis, and level of consciousness, Although the scoring system has not been validated, the index has been tested in Japan' 2 and Pennsylvania 1 3 with a good correlation between the index rating and the clinical state of the injured patient I week later.The index appears to be of value J in triage by paramedical personnel or in the emergency room, but lacks the precision required to compare management or to evaluate care.tI This paper is an attempt to further the quest for an acceptable, practical system for quantitation of injury, The methodologies used were mathematical!y'erived estimates of the probability of survival associated with injury to provide an objective 'a,ýsment of the degree of trauma involved.An attempt was made to provide a system that could be easily applied to a widespread variety of needs including triage, comparison of therapeutic modalities, evaluation of health care, and validation of other indices.
All patients with acute trauma admitted to a single referral center over a 4-year period (1972-1975) were studied.Upon discharge or death, a detailed diagnosis was provided by the attending physician and coded according to the Hospital Adaptation of the International Classification for Disease Adapted for use in the United States (H-ICDA).'4"This coding was checked against diagnoses in the hospital chart in triplicate: by medical records personnel, by computer card punchers, and by medical students, When autopsy findings were available, any necessary alterations in coding were made, All analyses were performed on a Univac 1108 computer.
Initially, 2,833 patients were included in this study.Patients injured by weapons were excluded as were patients with injuries (lacerations, fractures, dislocations, various nmusculoskeletal injuries-intrathoracic, intraabdominal, intracranial, vascular, nerve, and spinal U 1 cord injuries) that were not within the H-ICDA code range of 800.0 to 959.0,After subtracting these exclusions, 2,135 patients were left for analysis.A random selection of 1,884 of these patients was used to establish the statistical methodology, and this group of patients was called the Training Set, Data on the remaining 251 patients (Test Set) were withheld to validate the methodology.

A.
Methodology for the Trainin Set.
The training set of 1,884 patients provided a computed "conditional" probability of survival, PC, and an "effective" probability of survival, PE, for each injury code (in the range 800.0 to 959.9).The PC was derived as the proportion of survivals associated with each injury code.The conditional probabilities were used to rank the severity of the Injury codes by the decision rule that injury X was less severe than injury Y if the conditional probability of survival for injury X exceeded the conditional probability of survival for injury Y.
The ranking of injuries provided by PC was then used to cor.mpute the PE for each injury code.The PE for a given injury code is the proportion of survivors in the subset oc patients for whom this injury is the most severe injury sustained, the severity ranking being established by the PC's.For each of five random subgroups of the training set and for the test set, the PE for each code was used as a basis to validate the methodology.The probability of survival for each patient was estimated to be the PE associated with the most severe injury, The PE for each patient was used to compute the expected number of survivors for each subgroup of the training set and for the test set.The expected survival rate for a set of patients was computed by summing the probabilities of survival for all patients in that set.These values were compared with the actual number of survivors.
A decision rule predicting survival of a patient if the PE associated with his most severe injury was greater than 0. Using the total set of 2,135 patients, the procedure was repeated.The PC was computed for each Injury code.These were used to rank the codes and to recompute a PE for each code.

D.
Computation of Effective Probability of Survival for Anatomical Groups of Codes.
The injury codes were grouped anatomically.An effective probability of survival was associated with each group (G).Each patient whose most severe injury was an injury code in G was assigned to a set (SG).The PE for G was computed to be the proportion of survivors In SG.In this way, probabilities of survival were obtained for subclassifications of various anatomical groups.

E, Validation of Effective Probability of Survival.
The values for PE, associated with each H-ICDA diagnostic code, were used to predict the survival rates of five random patient goups comprising the total study set.Individual deaths were predicted using the decision rule previously applied to validate the methodology, and misclassification rates were calculated.

Ill. RESULTS.
The 2,135 patients studied were assigned 259 different injury codes.There was at least one and up to 14 codes for each patient (table A-i, appendix A).A weighted regression line (Y = 0.984 -0.127X) was computed for these data.Of the 2,135 patients, 1,751 (83%) survived.Data from the training set of 1,884 random patients provided the probability calculations.The percentage of survivors in the training set was 82%.In this set -and for each of the 259 injury codes -the PC, and subsequently the PE were derived.
Of' the 1,884 patients, 1,535 involved at least one of 40 diagnostic codes with a PC lcs.thaii 1.0 (i.e., the injury rccurred at least once in a patient who died), The remaining 349 patients involved at least one of It17 codes which were never associated with fatalities.Conditional probabilities were used to rank the severity of the injury codes and to provide the basis for the computation of the PEo A.
Effective Probability of Survival.
Of' the 1,884 patients with computed effective probabilities of survival, 349 possessed one or more of 117 codes which were never associated with a patient death.Each of 1,171 patients was labelled with one of 72 injury codes which occurred as a highest ranked code.For each of these 72 Injury codes, there was associated at least one death.Each of 364 patients had one of 61 codes which occurred as a highest ranked code; for patients experiencing these codes, there were no deaths, Seven codes existed which were associated with death although not in those patients where they occurred as the highest ranked code.

B. Test Set and Validation of Methodology.
The effective probabilities of death thus obtained for the training set were used on the 251 patients in the test set to predict the expected survival.
The survival rate predicted for the test set was 0.81.The survival rates predicted for the five random subsets of patients and the test set were consistently within 94% of the actual survival rate (table A-2).Misclassification rates for individual patients (including both false positives and false negatives) appear in table A-3.

C.
Effective Probability on Whole Data Set.
The training set and test set patients combined provided a total population of 2,135 patients with acute blunt trauma.The PC was used to rank and to compute the PE associated with each diagnostic code.A summary of the data is shown in table A-4.
The results of the validation process for PE are given in tables A-5 and A-6 with expected and actual survival rates and individual misclassification rates.The expected misclassification rate for the anatomical index was 0.12 and the actual rate experienced varied between 0,13 and 0.18 (table A-6).An example of the application of PE for various anatomical groupings is shown in figure B, appendix B.

IV. DISCUSSION.
The PC, associated with a given injury code, was computed as the proportion of survivors possessing a code.This reflects the survival associated with a given injury in the presence of any number of other injuries and, consequently, reflects both the severity of the individual injury and the frequency of its association with other injuries.
The PC provides a statistical basis for the ranking of one injury against another in a manner which reflects their occurrence in the patient population studied.The PC cannot be used as a severity score because certain incongruities occur where less severe injuries (such as a fractured humerus) are commonly associated with more severe injuries (such as a ruptured liver and cerebral contusion), resulting in an unreasonably high value for PC.Application of the methodology to a larger data base from a variety of treatment centers, such as the Illinois Trauma Registry, 1 5 may result in precise ranking.Although infrequent, the effects of such incongruous values for PC have predominantly been eliminated by the use of PE as the "score." The PE was obtained by excluding (from the data set) all patients with injuries achieving a lower ranking probability of survival than the one under analysis.It Is an attempt to estimate the impact of a specific injury in a real world setting, in that such injuries, when they do not occur alone, will have their maximum impact when they are the dominant injury.
Both PC and PE are objective values, unlike the arbitrary assignment forming the basis for the AIS and other current methods of quantitation.Internal consistency and statistical validity tests show that PE can be used to predict expected survival rates in patient groups to within ±5% although the methodology is insufficiently developed to predict accurate individual outcomes, Attempts to predict the survival for an individual patient will often result in a misclassification.The range of the effective probabilities of survival for the injuries is 0.17 to 1 .0.Thus, for a given injury with 600% chance of survival, prediction fcr an individual will result in misclassification 4 out of 10 times using our decision rule that predicts death if the probability of survi al is less than 0.5, A 12% actual misclassification rate was found when the PE from the training set was applied to the test set using a decision rule which predicts death if PE < 0.5 and otherwise predicts survival.This compares favorably with the misclassification rates associated with existing decision rules and is the same as the expected misclassification rate, An empirical comparison with a random decision rule (RDR), based on the a priori probability of survival (p) for the patient population studied, is of some interest.The RDR predicts survival for a patient if a random number, r, chosen from a uniform distribution of numbers on the unit interval is less than or equal to p; if r is greater than p, the RDR predicts death, The expected survival rate associated with the RDR would be p, and the misclassification rate would be 2p (I -p).This latter quantity is obtained as follows: Misclassiflcation rate = probability (r < p and the patient dies) + probability (r > p and the patient survives) = p0(-p)+(Ip)p 2p (I -p) "In our patient population p = 0.82 and 2p (I -p) = 2 (0.82) (0.18) -0.30,The decision rule utilized in this study thus decreases the misclassification rate by y8'r (30% to 12%) over a random prediction based on a knowledge of our patient population.
While the PE value undoubtedly underestimates the probability of survival associated with isolated injury, it is a mean probability of survival for an individual injury associated with other less severe trauma and thus incorporates the interaction of multiple less severe injuries and the injury under scrutiny.Estimates for PE within confidence limits of :1:5% require samples of 60 to 1,000 patients per injury depending on the value of PE.If applied to regional trauma registries, the PE could form that valid objective basis for evaluation of care and achievement so long elusive to the medical profession, Characterization of interactions which are present in multiple trauma is complex.Our data (table A-i) show surprisingly little overall effect In this context, while a separate study from the same centert and affirmed by Baker 10 showed a marked increase in the mortality rate when a severely damaged organ from another body system was added to a spectrum of injuries, but little effect from minor injuries Multiple Injuries within one body system occur most frequently within the abdomen and musculoskeletal system.On the basis of relative frequency and severity, the additive effect of an intraabdominal or musculoskeletal Injury will thus, In general, be less than that of a thoracic or central nervous system injury, A comprehensive model for multiple trauma should account for the most critical injury and the number and relative importance of the other injuries.
The results reported here might serve as a template on which a number of scales of injury could be formulated.Data on single injuries and Injury combinations could eventually be computed with a high degree of confidence and with automatic updating.It Is questionable whether such a degree of resolution, though Intellectually appealing, would be of practical value and significantly improve the predictive capability reported here for a selected population.
The widely used H-ICDA is a code for recording injury and may form a realistic labeling basis for such a system of quantitation.In its present form, however, subjectivity enters the process of coding to some extent because the descriptive terminology lacks specificity and sensitivity for certain injuries.Minor modifications, and definitions to aid in the assignment of such labels as "moderate cerebral contusion," would ameliorate this problem.
The ability to make a definitive diagnosis in trauma depends on the training and skill of the diagnostician and on the facilities available to aid in the diagnosis, Thus, a paramedic may discern that a patient has a chest injury; a physician may suspect a hemothorax; but the diagnosis is only confirmed by aspiration of blood and clearing of the effusion seen on chest X-ray.By combining the diagnostic groups as In figure B, we are Introducing a refinement of the statistical methodology which, when combined with a single parameter of physiological response (e.g., respiratory rate or level of consciousness), may be of benefit in triage or may add refinement to scoring systems already existing for this purpose.
The PE offers a mathematically derived data-based scale for measuring the effect of injury on the basis of anatomical disruption to individual organs or groups of organs.It has been derived from and tested on a select limited patient population.The statistical methodology has been validated.The methodology has a wide range of potential applications from validation of other injury-scoring systems to evaluation of health care delivery.It offers a system of scoring based on a specific diagnosis of the most severe injury as opposed to one based on an arbitrary assignment to a group of presumed equivalent injuries.
Baker et al, 1 0 added a new dimension to quantitation by correlating the arbitrary severity scores of the AIS with mortality.By utilizing the H-ICDA coding system, our methodology can easily be integrated into the medical records system of a hospital.A prospective comparison with the AIS would be another tentative step towards widespread systematic measurement of injury and treatment.
5 was used as the basis for individual patient prediction.Misclassification rates (MR), based on this decision rule, were calculated from the formula MR -PsPFP + PDPFN where PS = a priori probability of survival Pa f s Number of patients predicted to die, but survived PP=probability of false positives a ubro uvvr Number of survivors PD = a priori probability of death a I -PS PFN = probability of false negatives = Number of patients predicted to live, but died Number of deaths This calculated misclassification rate was compared with an expected misclassification rate (EMR) derived from the formula EMR -E p (L/Code) p (Code) + Z (I -p [L/Codel p (Code) RIR 2 where p (Code) = probability of the code appearing III all codes for which p (L/Code) <0,5 R 2 all codes for which p (L/Code) >0,5 p (L/Code)-probability of code appearing and patient surviving 9 9 .i................. gf C. ('omputation of Effective Probability of Survival Using the Entire Data Set.

Table A -
I. Distribution of Injury and Attendant Survival In Population Studied

Table 2 .
Predicted Death Rates in Five Random Subsets of Training Set Patients and a Test Set of 251 Patients using PE from Training Set

Table A - 5
Predicted Death RAte in Subseto of Total Patient Population Using PE from the Total Patient Population Table A.6.Misclanifi•tion Rates In Subsets of Total Patient Population Using PE from the Total Patient Population LIST FOR TR 16082 (Contd) Pi=rT'ECEDIID PAGE BLANK-NOT FILMED DISTRIBUTION