A Retrospective Study on Thyroid Disease Unsupervised Anomaly Detection.

Abstract

 

Introduction:
Thyroid illness is a prevalent endocrine ailment that impacts the synthesis and control of thyroid hormones. Enhancing patient outcomes can be achieved by early detection of thyroid problems. In order to find atypical instances and investigate thyroid status determinants, the current work sought to perform unsupervised anomaly detection on data related to thyroid disease.

Methods:  
A dataset of 6916 patient records with clinical and demographic data was used for a retrospective review. To find unusual records, two outlier identification algorithms showing local Outlier Factor and Isolation Forest were used. The predictors of thyroid medication use (on thyroxine) and thyroid disease status were analyzed using logistic regression and Cox regression using SPSS version 27.

Results:
250 records (3.6%) were flagged by the outlier identification algorithms as possible anomalies. Between outliers and typical cases, there were notable differences in a number of characteristics. Goitre, pregnant women, lithium, and query_on_thyroxine were found to predict on thyroxine use by logistic regression. After controlling for confounders, Cox regression revealed that older age, pregnancy, and query_on_thyroxine were risk factors for thyroid illness. It was reasonable to assume that proportionate hazards were satisfied.

Conclusion:
Unsupervised anomaly detection found a tiny subset of patient records with unusual features related to thyroid conditions. The use of thyroid medications and the state of the disease were linked to a number of sociodemographic and clinical parameters. More information on outlier instances may be able to clarify uncommon conditions and validate unusual results. The retrospective approach and the possibility of inadequate records are among the limitations.