Computer-assisted detection of monoclonal components: results from the multicenter study for the evaluation of CASPER (Computer Assisted Serum Protein Electrophoresis Recognizer) algorithm.

Background: To investigate the potential use of Artificial Neural Network (ANN) in the evaluation of serum protein electrophoresis, we set up a multicenter study involving six Italian laboratories. For this purpose, we developed an algorithm named CASPER (Computer Assisted Serum Protein Electrophoresis Recognizer). Methods: A total of 59,516 samples from the six centers were divided into three groups. Training and validation sets were used to develop the neural network, whereas evaluation set was used to test the performance of CASPER in recognizing abnormal electrophoretic profiles. Results: CASPER showed 93.0% sensitivity and 47.4% specificity. CASPER sensitivity and specificity ranged in the six sites from 88% (site 3) to 97% (site 5) and from 36% (site 6) to 53% (site 3), respectively. Sensitivity for g zone was 94.6%, for beta zone 89.7% and for oligoclonal patterns 92.0%. Conclusions: The sensitivity of the CASPER algorithm does not allow us to recommend its use as a replacement for the visual inspection, but it could be helpful in avoiding accidental misclassifications by the operator. Moreover, the CASPER algorithm may be a useful tool for training operators and students. This study evidenced a high inter-observer variability, which should be addressed in a dedicated study. Data set to train and validate ANNs should contain a huge range and an adequate number of different abnormalities.