[A machine learning diagnostic model for hereditary hearing loss based on GJB2 and SLC26A4 genes: construction and interpretability analysis].

Objective: To construct a machine learning diagnostic model for hereditary hearing loss based on GJB2 and SLC26A4 genes and perform interpretability analysis using SHapley Additive explanations (SHAP). Methods: The data of genetic variants and hearing status were collected from 1 539 individuals at the Deafness Molecular Diagnosis Center of the Chinese PLA General Hospital from June 2015 to August 2024. Participants were categorized by expert diagnosis as hereditary hearing loss patients or non-hereditary hearing loss individuals, and were randomly assigned to a training set (n=1 077) and a test set (n=462) in a 7∶3 ratio using a computer-generated random sequence. Using non-zero coefficient variants of GJB2 and SLC26A4 genes screened by least absolute shrinkage and selection operator (LASSO) regression, six machine learning models including logistic regression, decision tree, random forest, gradient boosting (GB), eXtreme Gradient Boosting and k-nearest neighbors were constructed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) with the DeLong test, and accuracy, precision, sensitivity, F1 score and specificity were calculated. The best model was compared against intermediate-level physicians to assess its clinical value, and SHAP was applied for interpretation. Results: A total of 748 hereditary hearing loss patients (391 males and 357 females) aged 29 (13, 36) years and 791 non-hereditary hearing loss individuals (405 males and 386 females) aged 27 (12, 36) years were included. LASSO regression yielded 121 non-zero coefficient variants: 34 in the GJB2 gene and 87 in the SLC26A4 gene. The GB model produced an AUC of 0.975 (95%CI: 0.967-0.983), outperforming each of the other five models (all P<0.05). Moreover, the GB model also showed higher precision [98.5% (95%CI: 97.8%-99.2%) vs 92.6% (95%CI: 90.4%-94.4%)] and specificity [98.7% (95%CI: 98.1%-99.3%) vs 93.2% (95%CI: 91.2%-94.8%)] than intermediate-level physicians (both P<0.001). SHAP identified the top 10 impactful variants in the GB model: nine pathogenic variants (GJB2: p.Leu79CysfsTer3, p.His100ArgfsTer14, p.Val37Ile, p.Gly59AlafsTer18; SLC26A4: c.919-2A>G, p.His723Arg, p.Asn392Tyr, p.Thr410Met, p.Val659Leu) and one benign variant (GJB2: p.Val27Ile). The model tends to diagnose individuals carrying pathogenic homozygous or compound heterozygous variants as hereditary hearing loss patients. Conclusion: The machine learning models incorporating the GJB2 and SLC26A4 genes are of referential value for the auxiliary diagnosis of hereditary hearing loss, with the GB model demonstrating the best diagnostic performance.