Assessment of machine-learning techniques on large pathology data sets to address assay redundancy in routine liver function test profiles.

ABSTRACT

BACKGROUND: Routine liver function tests (LFTs) are central to serum testing profiles, particularly in community medicine. However there is concern about the redundancy of information provided to requesting clinicians. Large quantities of clinical laboratory data and advances in computational knowledge discovery methods provide opportunities to re-examine the value of individual routine laboratory results that combine for LFT profiles. METHODS: The machine learning methods recursive partitioning (decision trees) and support vector machines (SVMs) were applied to aggregate clinical chemistry data that included elevated LFT profiles. Response categories for γ-glutamyl transferase (GGT) were established based on whether the patient results were within or above the sex-specific reference interval. Single decision tree and SVMs were applied to test the accuracy of GGT prediction by the highest ranked predictors of GGT response, alkaline phosphatase (ALP) and alanine amino-transaminase (ALT). RESULTS: Through interrogating more than 20,000 individual cases comprising both sexes and all ages, decision trees predicted GGT category at 90% accuracy using only ALP and ALT, with a SVM prediction accuracy of 82.6% after 10-fold training and testing. Bilirubin, lactate dehydrogenase (LD) and albumin did not enhance prediction, or reduced accuracy. Comparison of abnormal (elevated) GGT categories also supported the primacy of ALP and ALT as screening markers, with serum urate and cholesterol also useful. CONCLUSIONS: Machine-learning interrogation of massive clinical chemistry data sets demonstrated a strategy to address redundancy in routine LFT screening by identifying ALT and ALP in tandem as able to accurately predict GGT elevation, suggesting that GGT can be removed from routine LFT screening.