Developing result approval algorithms for biochemical tests with critical values: Learning algorithms

In the field of laboratory medicine, minimizing the errors and establishing standardization is possible by pre-defined processes. Assessment and decision algorithms are needed to use time efficiently especially for reporting those laboratory tests which have critical values. Very few studies are present in literature on the development of such algorithms and their integration to laboratory information systems. In these studies, "simple rule based" approach which does not allow cross-checking is used. The aim of this study was to build an experimental and decision algorithm model open to improvement which would efficiently and timely evaluate biochemical test results with critical values by evaluating multiple factors concurrently. The experimental model was built by WEKA® software based on the method of artificial neural network. Demographical and analytical data were received from Dokuz Eylül University Central Laboratory for a total of 252,847 samples with glucose, calcium, magnesium, potassium, sodium, uric acid requests between the dates 01.01.2013 and 01.03.2014. Data switched to virtual computer system were evaluated by the laboratory specialist. "Training sets" were developed for our experimental model to teach the evaluation criteria derived from literature information and specialists' views. After training the system, with the "test sets" developed for different conditions, validity of the model was assessed by statistical methods. By the learned algorithm, which was developed by training three times, no result was verified which was refused by the laboratory specialist. The rate of false rejection of those tests which needed to be accepted by the laboratory specialist was 0.5%. The sensitivity of this model was 91% and specificity was 100%. The estimated kappa score examining the agreement between the results of our model and the specialists' was 0,950. According to our literature search, this is the first study based on artificial neural network approach to build an experimental assessment and decision algorithm model open to improvement which will aid the specialist in verifying the results of the biochemical parameters. By integrating our trained algorithm model to laboratory information system, it could be possible to reduce work-load without compromising patient safety.