Abstract
Luo Y, Szolovits P, Dighe AS and Baron JM. Am J Clin Pathol 2016; 146: 778–788.
Clinical diagnoses are largely made via the expert interpretation of numerous laboratory test results. Despite this, clinical laboratories report the vast majority of their results as individual values. Writing in the American Journal of Clinical Pathology in June, Luo et al. present a clinical decision support algorithm that integrates multiple data elements in order to enhance the diagnostic efficiency of multivariate laboratory test panels. Machine learning techniques were applied to both impute a laboratory test result (ferritin) and predict patient classification (iron status) from a training data-set of haematological and biochemical results for 3590 cases.
Luo et al. demonstrated the ability of these methods to accurately predict patients’ iron statuses in a test dataset of 1538 cases. Their analysis also indicated that substantial redundancy existed within test panels. Interestingly, they demonstrated that the imputed ferritin values – calculated from the overall pattern of the other analytes – may, in some cases, better represent the patients’ true iron status than the measured value. This study represents an elegant proof-of-concept for the application of computational modelling techniques to laboratory data. Analysis of the entirety of the patient’s data as opposed to focusing on single results (cf AKI alerts) has the potential to improve the diagnostic power of laboratory data. Further work will determine if these methods enjoy similar success in other clinical scenarios.