Determination of clinically significant Gilbert syndrome genotype using machine learning tools based on routine laboratory tests

Background. Exploring the potential of artificial intelligence (AI) and machine learning (ML) is an important task aimed at creating an effective screening strategy, identifying risk groups and using more accessible and cost-effective laboratory tests for the diagnosis of Gilbert syndrome (GS).
Aim. To develop and evaluate the accuracy of a prognostic model for detecting mutations in the UGT1A1 gene based on demographic data (age, gender) and laboratory test results: bilirubin fraction, hemoglobin, alanine aminotransferase (ALT) and aspartate aminotransferase (AST).
Methods. An anonymized laboratory database was used: 1499 patients, genotypes: 6TA/6TA (n=179), 6TA/7TA (n=496), 7TA/7TA (n=824). The classification model was developed using the LightAutoML platform, combining linear regression and gradient boosting on decision trees.
Results. On the test sample, the developed model showed high diagnostic efficiency in predicting the 7TA/7TA genotype: precision 78%, recall 88%, F1-score 83%. On the validation sample of patients with mechanical jaundice, the model demonstrated high specificity without classifying the 7TA/7TA genotype, which emphasizes its potential for differential diagnosis of hyperbilirubinemia.
Conclusion. The model is promising for clinical use, especially in screening and medical decision support systems (MDSS).

1. Vítek L., Tiribelli C. Gilbert’s syndrome revisited. J Hepatol. 2023;79(4):1049-1055. doi: 10.1016/j.jhep.2023.06.004.

2. Veeravalli R. S. S. Data-Driven Exploration of Phenotypes in UK Electronic Health Records for Symptom Identification and Diagnosis of Rare or Monogenic Disease. Diss. UCL (University College London). 2024.

3. Amina D., Murveta D., Amel D. et al. Diagnosis of Different Types of Hyperbilirubinemia Using Artificial Neural Network. International Conference on Medical and Biological Engineering. 2021: 199-207. https://doi.org/10.1007/978-3-030-73909-6_22

4. Vakhrushev A., Ryzhkov A., Savchenko M. et al. Lightautoml: Automl solution for a large financial services ecosystem. arXiv preprint arXiv. 2021; 2109.01528. https://doi.org/10.48550/arXiv.2109.01528

5. Ke G., Meng Q., Finley T. et al. Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems. 2017; 30: 3149-57.