B-159 Data Mining With the Bhattacharya Method as a Tool for Reference Interval Determination in Populations With Thyroid Disorders

Abstract

Abstract Background Establishing a precise reference interval (RI) is crucial for laboratory diagnosis, population studies, and aiding physicians in patient follow-up treatment. Data mining and Bhattacharya analysis can be an efficient and cost-effective way to determine RIs for thyroid disorders based on large datasets. Bhattacharya analysis is a simple resolution method that detects distributions into Gaussian components. Unlike traditional reference interval studies using parametric or non-parametric statistics, Bhattacharya analysis does not exclude results from an “unhealthy” population before analysis. This study aimed to define the RI for thyroid-stimulating hormone (TSH) and free thyroxine (fT4) from three distinct regions in São Paulo, Brazil, using the Bhattacharya method. Methods The study was a 20-month retrospective analysis of a large laboratory database in São Paulo from February/2021 to September/2022. Individuals aged 21 or older with available TSH or fT4 results were selected, including only outpatient results. Three groups from different regions (R1, R2, R3) were studied and the indirect Bhattacharya method was used to determine each region’s RI values. The midpoint, upper and lower reference limits (URL and LRL), were calculated, and a data plot was created using Microsoft Excel. One-way ANOVA was used for multiple group comparisons and results with *P < 0.01 were considered statistically significant. Results A total of 234 903 TSH and fT4 records were obtained. The three groups showed significant differences in their mean age and TSH and fT4 results (P < 0.001). The mean age was 49.41 ± 16.9 for R1, 46.9 ± 9.6 for R2, and 53 ± 16 years old for R3. The average TSH value for all three regions was 2.44 µIU/mL (SD 0.18, CV 7.2%), and the average fT4 value was 1.17 ng/dL (SD 0.06, CV 4.7%). The central bin was selected for the best data distribution and close to the average or median for each group. The derived RIs for R1, R2, and R3 were compared for consideration of a common RI. Other criteria, such as the best bin size and at least four bins for the graph curve, were also applied. Altogether, Bhattacharya’s analysis identified the following estimates from all collection points for TSH and fT4: TSH Average LRL 0.43 µIU/mL (SD 0.07, CV 16.6%)/TSH Average URL 6.45 µIU/mL (SD 0.85, CV 13.1%)/TSH Average Range 6.03 µIU/mL (SD 0.88, CV 14.6%)/fT4 Average LRL 0.82 ng/dL (SD 0.07, CV 8.1%)/fT4 Average URL 1.47 ng/dL (SD 0.07, CV 4.4%)/fT4 Average Range 0.65 ng/dL (SD 0.02, CV 3.5%). 9.7% of individuals in R3 had results outside the accepted RIs for TSH and fT4, compared to R2 as a standard population with the lowest possibility of relevant thyroid disorders. Conclusion Bhattacharya’s non-parametric method could be useful to determine reference values and laboratory analysis of populations with distinct characteristics. The study highlights the importance of choosing the correct control group in comparison studies, and the potential bias in unifying distinct regions to determine RIs. The low cost of data mining makes it possible for nearby laboratories to perform similar studies using the Bhattacharya method to establish RI results suitable for specific patient populations.