William G. Baxt

Use of an artificial neural network for data analysis in clinical decision-making: the diagnosis of acute coronary occlusion

A nonlinear artificial neural network trained by backpropagation was applied to the diagnosis of acute myocardial infarction (coronary occlusion) in patients presenting to the emergency department with acute anterior chest pain. Three-hundred and fifty-six patients were retrospectively studied, of which 236 did not have acute myocardial infarction and 120 did have infarction. The network was trained on a randomly chosen set of half of the patients who had not sustained acute myocardial infarction and half of the patients who had sustained infarction. It was then tested on a set consisting of the remaining patients to which it had not been exposed. The network correctly identified 92% of the patients with acute myocardial infarction and 96% of the patients without infarction. When all patients with the electrocardiographic evidence of infarction were removed from the cohort, the network correctly identified 80% of the patients with infarction. This is substantially better than the performance reported for either physicians or any other analytical approach.

The trauma triage rule: A new, resource-based approach to the prehospital identification of major trauma victims

STUDY OBJECTIVE To develop a new trauma decision rule. DESIGN Retrospective clinical review. SETTING Level I trauma center. TYPE OF PARTICIPANTS 1,004 injured adults. MEASUREMENTS AND MAIN RESULTS A new trauma decision rule was derived from 1,004 injured adult patients using a new operational definition of major trauma. The rule, termed the Trauma Triage Rule, defines a major trauma victim as any injured adult patient whose systolic blood pressure is less than 85 mm Hg; whose motor component of the Glasgow Coma Score is less than 5; or who has sustained penetrating trauma of the head, neck, or trunk. Using the operational definition of major trauma, the rule had a sensitivity of 92% and a specificity of 92% when tested on the 1,004-patient cohort. CONCLUSION The Trauma Triage Rule may significantly reduce overtriage while only minimally increasing undertriage. This approach must be validated prospectively before it can be used in the prehospital setting.

The impact of a regionalized trauma system on trauma care in San Diego County

A review of autopsy reports on traumatic deaths in 1986 was conducted to determine the impact on trauma mortality of the regionalized trauma system instituted in San Diego County in 1984. Determination of preventable death was made by a panel of experts and compared with an identical review of traumatic deaths in 1979, five years before the institution of regionalized trauma care. Of 211 traumatic deaths reviewed from 1986, two (1%) were classified as preventable, compared with 20 of 177 (11.4%) deaths in 1979 (P less than .001). A breakdown of trauma deaths into central nervous system and noncentral nervous system categories revealed the overall decline was in large part a consequence of the decline in non-central nervous system deaths from 16 of 83 in 1979 to one of 62 in 1986 (P less than .005). The decrease in central nervous system-related preventable deaths from four of 94 in 1979 to one of 149 in 1986 (P less than .10) was not statistically significant. Although it is likely the trauma system introduced in 1984 contributed to the decline in preventable death, it is not possible to isolate this variable from other changes that occurred during the interval between studies. A review of trauma deaths over the same time interval in a community with similar demographics but without a trauma system might help determine the relative contribution of the trauma system.

Improving the accuracy of an artificial neural network using multiple differently trained networks

When either detection rate (sensitivity) or false alarm rate (specificity) is optimized in an artificial neural network trained to identify myocardial infarction, the increase in the accuracy of one is always done at the expense of the accuracy of the other. To overcome this loss, two networks that were separately trained on populations of patients with different likelihoods of myocardial infarction were used in concert. One network was trained on clinical pattern sets derived from patients who had a low likelihood of myocardial infarction, while the other was trained on pattern sets derived from patients with a high likelihood of myocardial infarction. Unknown patterns were analyzed by both networks. If the output generated by the network trained on the low risk patients was below an empirically set threshold, this output was chosen as the diagnostic output. If the output was above that threshold, the output of the network trained on the high risk patients was used as the diagnostic output. The dual network correctly identified 39 of the 40 patients who had sustained a myocardial infarction and 301 of 306 patients who did not have a myocardial infarction for a detection rate (sensitivity) and false alarm rate (1-specificity) of 97.50 and 1.63%, respectively. A parallel control experiment using a single network but identical training information correctly identified 39 of 40 patients who had sustained a myocardial infarction and 287 of 306 patients who had not sustained a myocardial infarction (p = 0.003).

The failure of prehospital trauma prediction rules to classify trauma patients accurately

Clinical prediction rules are used extensively by most regionalized trauma systems to identify which patients have sustained major injuries. Because of reported high misclassification rates of some of these rules and the known global difficulty of transporting prediction rules, four such rules (the Trauma Score, the CRAMS Scale, the Revised Trauma Score, and the Prehospital Index) and two newly derived rules were statistically analyzed using a cohort of 2,434 injured patients. All rules accurately predicted mortality with a minimum sensitivity and specificity of 85%. However, not one of the rules was able accurately to identify surviving patients who had sustained major injuries. In this instance, no rule was able to achieve a sensitivity of at least 70% while achieving a specificity of 70%. These results suggest that the problem with trauma prediction rules lies in the inherent limitations of the clinical data on which they are based. In view of this, the usefulness of existing prehospital trauma predictive rules must be questioned.

Analysis of the clinical variables driving decision in an artificial neural network trained to identify the presence of myocardial infarction.

STUDY OBJECTIVE To determine which clinical variables drive the output of an artificial neural network trained to identify the presence of myocardial infarction. DESIGN Partial output analysis. SETTING Tertiary university teaching center. PARTICIPANTS Seven hundred six patients more than 18 years old presenting with anterior chest pain. MEASUREMENTS Differential network output analysis. MAIN RESULTS A methodology was developed as the first step in measuring the impact input clinical variables have on the output (diagnosis) of an artificial neural network trained to identify the presence of acute myocardial infarction. The methodology revealed that the network used the presence of ECG findings, as well as the presence of rales, syncope, jugular venous distension, response to trinitroglycerin, and nausea and vomiting, as major predictive sources. Although this first-step analysis studied individual variables, it must be stated that the network comes to clinical closure based on the settings of all variables in a pattern and that the impact of a single variable cannot be taken out of the context of a pattern. CONCLUSION An artificial neural network trained to recognize the presence of myocardial infarction appears to place diagnostic importance on clinical variables that have not been shown previously to be highly predictive for infarction.

Bootstrapping confidence intervals for clinical input variable effects in a network trained to identify the presence of acute myocardial infarction

The artificial neural network has been successfully applied to a broad range of clinical settings (Widrow and Hoff 1960; Rumelhart et al. 1986; McClelland et al. 1988; Weigend et al. 1990; Hudson et al. 1988; Smith et al. 1988; Saito and Nakano 1988; Kaufman et al. 1990; Hiraiwa et al. 1990; Cios et al. 1990; Marconi et al. 1989; Eberhard et al. 1991; Mulsant and Servan-Schreiber 1988; Bounds et al. 1990; Yoon et al. 1989). Such a network has been adapted for use as an aid to the clinical diagnosis of acute myocardial infarction (Baxt 1990, 1991, 1992a; Harrison et al. 1991) (heart attack). Both initial retrospective and subsequent prospective studies have revealed that this network performed more accurately than either physicians or other electronic data processing technologies (Baxt 1990, 1991; Goldman et al. 1988). Since nonlinear artificial networks are known to be capable of identifying relationships between input data that are not apparent to human analysis (Weigend et al. 1990), one hope has been that the network could be utilized to identify relationships in clinical data that have not been revealed by previous study. The inherent problem in this hope has been the inability easily to identify how artificial neural networks derive their output. One indirect way that this can be approached is by the stepwise perturbation of isolated individual input variables across a large number of patterns coupled with an analysis of the effect this has on network output. Prior application of this analysis to the artificial neural network trained to identify the presence of acute myocardial infarction revealed that one could gain a

The lack of full correlation between the Injury Severity Score and the resource needs of injured patients.

STUDY OBJECTIVE To determine whether the Injury Severity Score (ISS) correlates with the resource requirements of severely injured patients by studying the association of the ISS with three major interventions (fluid resuscitation, invasive central nervous system monitoring, and acute operative repair) trauma centers routinely provide severely injured patients. DESIGN Retrospective clinical review. SETTING Level I trauma center. TYPE OF PARTICIPANTS Eight hundred fourteen adult injured patients. MEASUREMENTS AND MAIN RESULTS When an ISS of more than 9 was used as the definition of major trauma, the ISS undercorrelated 11% of the time with the need for any one procedure. When an ISS of more than 14 was used as the definition, it undercorrelated 20% of the time. CONCLUSION The ISS may not be completely correlated with the resource requirements of injured patients and should not be used as the sole means by which to define major injury.

Complexity, chaos and human physiology: the justification for non-linear neural computational analysis

Background is presented to suggest that a great many biologic processes are chaotic. It is well known that chaotic processes can be accurately characterized by non-linear technologies. Evidence is presented that an artificial neural network, which is a known method for the application of non-linear statistics, is able to perform more accurately in identifying patients with and without myocardial infarction than either physicians or other computer paradigms. It is suggested that the improved performance may be due to the networks better ability to characterize what is a chaotic process imbedded in the problem of the clinical diagnosis of this entity.

An autopsy study of traumatic deaths: San Diego County, 1979☆

All traumatic deaths in San Diego County were analyzed for the year of 1979. Death certificates, coroners reports, and autopsy data served as the basis for this review. A total of 177 deaths were studied, of which 94 were associated with CNS injury and 83 were not. Sixteen (20 percent) of the deaths not CNS-associated and four (5 percent) of the CNS-associated deaths were classified as preventable. One hundred seventeen deaths were due to motor vehicle accidents, of which 11 of 35 (31 percent; all not CNS-associated) were deemed preventable. Preventable causes of death included hemorrhage, unrecognized hemopneumothorax, and unrecognized epidural hematoma.