An unsupervised learning method to identify reference intervals from a clinical database

Abstract

Reference intervals are critical for the interpretation of laboratory results. The development of reference intervals using traditional methods is time consuming and costly. An alternative approach, known as an a posteriori method, requires an expert to enumerate diagnoses and procedures that can affect the measurement of interest. We develop a method, LIMIT, to use laboratory test results from a clinical database to identify ICD9 codes that are associated with extreme laboratory results, thus automating the a posteriori method. LIMIT was developed using sodium serum levels, and validated using potassium serum levels, both tests for which harmonized reference intervals already exist. To test LIMIT, reference intervals for total hemoglobin in whole blood were learned, and were compared with the hemoglobin reference intervals found using an existing a posteriori approach. In addition, prescription of iron supplements were used to identify individuals whose hemoglobin levels were low enough for a clinician to choose to take action. This prescription data indicating clinical action was then used to estimate the validity of the hemoglobin reference interval sets. Results show that LIMIT produces usable reference intervals for sodium, potassium and hemoglobin laboratory tests. The hemoglobin intervals produced using the data driven approaches consistently had higher positive predictive value and specificity in predicting an iron supplement prescription than the existing intervals. LIMIT represents a fast and inexpensive solution for calculating reference intervals, and shows that it is possible to use laboratory results and coded diagnoses to learn laboratory test reference intervals from clinical data warehouses.

Reference Intervals Generated by Electronic Medical Record Data Mining with Clinical Exclusions: Age-Specific Intervals for Thyroid-Stimulating Hormone from 33038 Euthyroid Patients

“…Some studies have used repeat testing as an assumption of illness and, therefore, an exclusion criterion (11,12), whereas Kouri et al (1) used discharge diagnoses related to the analyte of interest to exclude patients. Poole et al (13) recently used unsupervised computer learning to identify ICD9 codes associated with extreme laboratory values and then used these codes to exclude individuals from the reference population. Here, we demonstrate how data mining coupled with disease-specific exclusion criteria was used in the a posteriori sampling technique to select a large reference population of euthyroid patients from the laboratory database to establish age-based thyroid-stimulating hormone (TSH) reference intervals.…”

Indirect Reference Intervals: Harnessing the Power of Stored Laboratory Data

“…This may simply involve excluding values beyond an arbitrary limit, such as those more than 10 times the upper reference limit 13 or involve a statistical test, such as that of Tukey, or others. [14][15][16] Another data pre-processing step that may be used is the exclusion of data from particular referral sites where there is a high likelihood that the patients have significant disease, such as intensive care units and oncology departments. 13,17 It may also be appropriate to exclude data from additional referral sites depending on the analyte of interest, for example lipid and renal clinics.…”

“…This was done by the Laboratory Mining for Individualized Threshold (LIMIT) study, which used an unsupervised machine learning algorithm to identify diagnostic codes that were significantly associated with outlier results for the analyte of interest. 14 The 'learning' component of the algorithm involved setting values for 4 parameters (one of which, for instance, governed the sensitivity to outlier detection). These values were set using data for serum sodium because of its well-established reference interval.…”

Deep Learning in Medicine

“…In this case, the very notion of reference limits would change, and ML, by leveraging and improving other statistical approaches, could help limit the misinterpretation of values outside of reference limits or of apparently normal data but also diagnostic for some conditions (e.g. [25]). Furthermore, some envision an ML-based clinical decision support that, by predicting correlated test results and enhancing the diagnostic value of multianalyte sets of test results, could help to reduce redundant laboratory testing [26] and, hence, lower healthcare costs, which are estimated to total $5 billion yearly in the United States alone [27].…”