Machine learning-based model for identifying liver injury in patients with thyroid-associated ophthalmopathy

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: Random forest model: AUC 0.977 (cross-validation), AUC 0.973 (test set), accuracy 0.914. SVM model: AUC 0.986 (test set), accuracy 0.914. Logistic regression: accuracy 0.914 (test set).

• Primary Outcome: Classification of liver injury versus normal liver function in TAO patients using machine learning models

• Nnt Or Sensitivity: Sensitivity, specificity, PPV, NPV, and Youden index stated as evaluated but specific values not provided in abstract; AUC 0.973–0.986 on test set across best-performing models

• Confidence Interval: Not reported for any metric

Clinical Application

The model requires clinical and biochemical input variables that are routinely collected in TAO management. However, without a published, externally validated, and openly accessible model with defined input requirements, clinical implementation is not currently feasible. Routine liver function testing (LFTs) remains the standard and most direct approach to detecting liver injury. In Australian practice, TAO is managed according to RACGP and Endocrine Society of Australia guidelines, with antithyroid drug therapy (carbimazole, propylthiouracil) known to carry hepatotoxicity risk — a key reason liver monitoring is already standard. The TGA-approved antithyroid agents carry PBS-listed indications with mandatory liver function monitoring. An ML model identifying liver injury risk could theoretically guide monitoring frequency, but would require validation in Australian populations with different genetic backgrounds, comorbidity profiles, and treatment patterns before adoption. The model's high apparent performance in a Chinese tertiary cohort cannot be assumed to generalise to Australian primary or secondary care settings without prospective external validation. Adult patients with confirmed TAO under specialist ophthalmology or endocrinology care, particularly those at risk of concurrent liver dysfunction (e.g., those on antithyroid drugs, with autoimmune comorbidities, or with elevated baseline liver enzymes)

Abstract

PURPOSE: To explore the relationship between thyroid-associated ophthalmopathy (TAO) and liver injury and establish a model for identifying liver injury by using machine learning so as to provide an effective diagnostic tool for liver injury in patients with TAO. METHOD: A single-center retrospective study was conducted to collect the clinical data of 318 patients with TAO admitted to a hospital from 2016 to 2022. The patients were divided into a TAO liver injury group (104 cases) and TAO normal liver function group (214 cases) according to whether the patients had normal liver function. The multivariate binomial logistic regression model was used to analyse the risk factors for liver injury in patients with TAO. Feature selection was performed using the random forest algorithm. The research data were divided into a training set and a test set at a ratio of 6:4. Taking whether accompanied by liver injury (0 = no, 1 = yes) as the outcome variable, models were established based on logistic regression, random forest, support vector machine, and decision tree. The performance of the models was evaluated using metrics, including sensitivity, specificity, positive predictive value, negative predictive value, Youden index, and accuracy. RESULT: The random forest method was used for feature screening, and variables ranked among the top 10 according to either mean decrease accuracy or mean decrease Gini were selected for model construction. In five-fold cross-validation, the RF model showed the highest accuracy of 0.937 and AUC of 0.977. In the test set, the RF model also showed good discrimination, with an AUC of 0.973, and the SVM model showed the highest AUC of 0.986, while SVM and LR achieved the highest accuracy of 0.914. CONCLUSION: This study shows that a classification model based on machine learning can effectively identify the risk of liver damage in patients with TAO.

References

1. 1.Guan, Y., Zhu, J., Zhu, H., Zhang, S., Wu, L., & Lin, W. (2026). Machine learning-based model for identifying liver injury in patients with thyroid-associated ophthalmopathy. International Ophthalmology. https://doi.org/10.1186/s13062-020-00285-0