Transferability of neural network-based decision support algorithms for early assessment of chest-pain patients

The present investigation concerns methodological and epidemiological aspects of the transferability of artificial neural network-based algorithms, as key-components for classification in decision support systems (DSS). The prevalence of pathological conditions to be detected must be known in order to tune an artificial neural networks (ANN)-decision algorithm so that the predictive values of the outcome fulfil medical requirements. Another aspect of transferability, when clinical laboratory results are used, concerns differences in analytical performance of measuring instruments. The relative bias between two instruments is not known exactly, but must be estimated and corrected for. A general method, based on original measured data sets and statistical modeling, was developed for simulating the impact of various correction procedures when using different analytical instruments. The simulation methodology was applied to a real clinical problem of ruling-in/ruling-out of patients with suspected acute myocardial infarction (AMI) by biochemical monitoring. The recommended correction procedure was based on method comparison with use of five duplicate measurements on a common set of patient samples covering the relevant measuring interval. Transferability of laboratory data over time was also studied. The design of quality assurance procedures should be based on analytical quality requirement specifications related to medical needs. Limits of critically sized systematic errors were assessed by calculating the decrease in diagnostic performance of the ANN-algorithm as a result of temporary analytical disturbances. The consequences for the design of QA procedures was illustrated. It is concluded that the actual ANN-decision algorithm for early assessment of chest-pain patients should be possible to transfer to new sites under realistic conditions.