B-136 Indirect Estimation and Verification Tool for Reference Intervals Using Unsupervised Machine Learning

Abstract

Abstract Background Determining the Reference Intervals (RI) of quantitative laboratory tests is crucial. This aspect is also part of the accreditation process, which can be challenging. The goal of our study was to create and validate a model using the R programming language that can estimate and verify Reference Intervals using an indirect method by mining data from the Laboratory Information System. Methods The new model integrates algorithms that carry out preliminary processing, grouping, and statistical reduction of results. The algorithm identifies unusual outcomes by analyzing asymmetry and kurtosis and applying the Expectation-Maximization (EM) algorithm. During the grouping and reduction phases, the EM algorithm computes the mean, standard deviation, and weight of the distribution for each group. Additionally, the model integrates algorithms that determine the global mean/median, modes, antimodes, kurtosis, and skewness, to find values that best represent the reference population’s distribution. The distribution is then truncated to the defined IR partition through the combination of algorithms (Fig. 1). The Bias Ratio was used to evaluate the difference between the RI determined by our model and those obtained from the Kosmic and Refine R models. The model’s validity was tested using simulated and real data from anonymous reference individuals.Fig. 1. Results The effectiveness of the model was evaluated using was assessed for each test under varying levels of contamination. The overall performance showed excellent results with an accuracy of 0.996, kappa of 0.882, sensitivity of 0.945, specificity of 0.979, positive predictive value of 0.864, negative predictive value of 0.992, and F1-score of 0.896. Conclusions The indirect estimation and verification tool for Reference Intervals is a cost-free and valuable resource for clinical laboratories. It supports the accurate interpretation of patient´s results.