From objective grouping to fuzzy reference intervals: A standardized machine learning approach for thyroid function tests.

Background: Accurate interpretation of thyroid function tests (TFTs) requires reliable reference intervals (RIs).

Indirect methods based on retrospective laboratory data are increasingly used, but current strategies face major

limitations, including rigid age cut-offs, inconsistent partitioning, and lack of objective evaluation of subgrouping

criteria. To address these challenges, we developed a novel machine learning (ML)-based framework for the

objective, data-driven determination of continuous and personalized RIs.

Methods: A total of 48,397 records (2019–2021) from individuals aged ≥18 years were retrieved from the

hospital laboratory information system. After applying inclusion and exclusion criteria, 9455 individuals

constituted the reference sample group. Partitioning was based on age, sex, thyroid-stimulating hormone (TSH),

and free thyroxine (fT4). The Elbow method suggested the optimal number of clusters; K-means clustering was

used to form subgroups, and the Extra Trees Classifier quantified the relative importance of partitioning criteria.

RIs were estimated using the non-parametric percentile method, and Fuzzy C-Means clustering was applied to

smooth sharp transitions between groups.

Results: Six subgroups were identified, with age as the dominant determinant (feature importance score = 0.96),

whereas sex had a negligible effect. Fuzzy clustering generated continuous and personalized RIs. In clinical

evaluation, applying fuzzy RIs reclassified 7 % of patients, predominantly shifting diagnoses toward

hyperthyroidism.

Conclusions: This study offers a universal and adaptable ML-based framework for generating continuous,

personalized RIs. This advance toward precision laboratory medicine can enhance diagnostic accuracy, reduce

misclassification, and support more patient-centered care.