Background: Patient-based real-time quality control (PBRTQC), a complement to traditional QC, may eliminate matrix effect from QC materials, realize real-time monitoring as well as cut costs. However, the accuracy of PBRTQC has not been satisfactory as physicians expect till now. Our aim is to set up a artificial intelligence-based QC for small error detection in real laboratory settings. Taking tPSA as our unique research subject, data extraction, data stimulation, data partition, model construction and evaluation were designed.
Methods: 84241 deidentified results for tPSA were extracted from Laboratory Information System of Aviation General Hospital. The data set was accumulated by way of data simulation. Independent training and test datasets were separated. After three classification models (RF, SVM and DNN) in ML constructed and weighted by information entropy, a multi-model fusion algorithm was generated. Performance of the fusion model was evaluated by comparing with optimal PBRTQC.
Results: For 4 PBRTQC methods, MovSO showed overall better performance for 0.2 μg/L bias and optimal MNPed was equal to 200. For the fusion model, MNPeds were less than 12 for all biases, and ACC surpassed MovSO nearly 100 times. Except for 0.01 μg/L bias, ACC was more than 0.9 for the rest of biases. FPR was apparently lower than MovSO, only 0.2% and 0.1%.
Conclusion: The fusion model shows outstanding performance and reduces incorrect and omitting error detection, adaptable for the real settings.
Keywords: Information entropy; Machine learning; PBRTQC; Quality control; Small shift.
Copyright © 2022 The Authors. Published by Elsevier Ltd.. All rights reserved.