Diagnosis of myocardial infarction: integration of serum markers and clinical descriptors using information theory

We examine the use of information theory applied to a single cardiac troponin T (cTnT) (first generation monoclonal; Boehringer Mannheim Corp., Indianapolis, Indiana) used with the character of chest pain, electrocardiography (ECG) and serial ECG changes in the evaluation of acute myocardial infarction (AMI). We combined a single measure of cTnT (blinded to the investigators) with a creatine kinase MB isoenzyme (CK-MB) measurement to discover the best decision value for this test in a study of 293 consecutive patients presenting to the emergency department with symptoms warranting exclusion of AMI.

The decision value for determining whether cTnT is positive or negative was determined independently of the final diagnosis by examining the information in the cTnT and CKMB data. Using information theory, an autocorrelation matrix with a one-to-one pairing of the CKMB and troponin T was constructed. The effective information, also known as Kullback entropy, assigned the values for troponin T and for CKMB that have the lowest frequency of misclassification error. The Kullback entropy is determined by subtracting the data entropy from the maximum entropy of the data set in which the information has been destroyed. The assignment of the optimum decision values was made independently of the clinical diagnoses without the construction of a receiver-operator characteristic curve (ROC). The final diagnosis of AMI was independently determined by the clinicians and entered into the medical record.

The decision value for cTnT was 0.1 ng/ml as determined by the the information in the data. The method was validated within the same study by mapping the results so obtained into the diagnoses obtained independently by the clinicians using all of the methods at their disposal. The cTnT was different in AMI (n = 60) compared with non-AMI patients (n = 233) (2.08 +/- 0.21 vs. 0.07 +/- 0.10; p < .0001).

Information theory provides a strong framework and methodology for determining the decision value for cTnT which minimizes misclassification errors at 0.1 ng/ml. The result has a strong correlation with other features in detecting AMI in patients presenting with chest pain.