B-154 Development of Deep Learning Pipeline for Identification of IV Fluid Contamination in Basic Metabolic Panels

Background Intravenous (IV) fluid contamination is a common preanalytical error that occurs in clinical testing, which may result in alterations in measured analyte concentrations, misdiagnosis, and delays in patient care. Methods for detecting contamination vary across laboratories, which often involve a combination of manual review and automated rules such as delta flags, and typically rely on laboratory technologist expertise and non-specific indicators. Deep learning algorithms have the potential to transform healthcare, utilizing both unsupervised and supervised machine learning to solve problems involving large and complex data. Artificial Neural Networks (ANNs) can be developed and applied to identify IV fluid contamination, reducing the dependence on manual review by technologists and improving laboratory efficiency by identifying errors poorly detected by current clinical workflows. A real-time integrated methodology to support laboratory technologists in identifying IV fluid contamination would reduce a significant source of preanalytical error, improve clinical operations and efficiency, and benefit patient care. Methods As a primary key development step, a validation study using published, anonymized Basic Metabolic Panel (BMP) results was conducted offline in a Jupyter Notebook utilizing Python and Tensorflow, a powerful framework for developing and deploying machine learning and artificial intelligence models. A Long Short Term Memory (LSTM) model, an improved version of a recurrent neural network (RNN), was selected for its excellence in classification tasks and in capturing long-term dependencies. 17,350 samples of current and prior BMP results obtained from a public data repository were split into training and testing datasets. Following expert review criteria, 17,117 were labeled as non-IV contaminated and 253 were labeled as IV contaminated. Statistical classification metrics in addition to Matthews Correlation Coefficient (MCC) and area under the receiver operating characteristic curve (auROC) were measured to evaluate model performance in classifying IV fluid contamination. The results of this analysis are utilized to develop a scalable pipeline to be validated and integrated into the current clinical workflow. Results Accuracy, sensitivity, specificity, and positive predictive value (PPV) were 99.481%, 99.674%, 84.403% and 99.801%, respectively. The false positive rate (FPR) was 15.596%. The Matthews Correlation Coefficient (MCC) was 0.802, and the area under the receiver operating characteristic curve (auROC) was 0.88. Conclusion The deep learning algorithm is evaluated and compared for its advantages and disadvantages in identifying IV fluid contamination in Basic Metabolic Panel results. The model correlates with existing published findings and outperforms comparable methodologies, suggesting feasibility of using this approach to identify contaminated results. Further development and integration into clinical workflows is pending additional validation.