Answers are grounded in this paper and cite their sources.
Modern medicine is heavily reliant on laboratory results for diagnostics. The interpretation of these results hinges on the comparison of measured values of analytes with reference intervals (RIs) established for healthy populations. However, the traditional means of determining these RIs are costly and time-consuming, often leading to the application of one-size-fits-all reference intervals despite compelling evidence that factors such as sex, age, and ethnicity significantly impact their validity. At the same time, clinical laboratory data routinely generated in everyday medical practice holds great potential for enhancing patient diagnostics and care. In particular, it can be used to determine RIs for biomarkers with much higher granularity and precision at a fraction of the cost. This study leverages unsupervised machine learning to differentiate normal from pathological results in unlabeled screening laboratory tests of hepatic biomarkers. The results distribution is modeled as a transformed Gaussian mixture of healthy and pathological components, with Bayesian Gaussian mixture modeling used to estimate the parameters of the healthy distribution from which the RIs are derived. Furthermore, this methodology extends naturally to the joint distribution of multiple biomarkers. Reference surfaces (RSs), representing the 95% boundary of the multidimensional joint distribution, offer a more holistic assessment compared to traditional one-dimensional RIs. These RSs correspond to level sets of the inverse survival function, defined via Mahalanobis distance, and allow the abnormal flag to be generalized into a survival probability that provides a quantitative measure of abnormality. Applying this approach, we estimated RIs for six hepatic biomarkers in the Puerto Rican population. Both univariate RIs (as marginals of the joint distribution) and multivariate RSs were obtained, with the latter capturing correlations between biomarkers that rectangular combinations of univariate RIs do not account for. This work demonstrates the feasibility of using real-world clinical laboratory data and probabilistic modeling to construct multivariate RSs. While univariate RIs remain useful for comparability with current practice, RSs extend the concept of reference intervals to multiple dimensions and account for correlations among biomarkers; their diagnostic utility remains to be established against clinical outcomes.
No takes yet. Share an insight, caveat, or question.
Velev et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: