A clinical predictive model for hearing recovery after middle ear cholesteatoma surgery based on machine learning

Yahui Zhao^1,2*shengnan ye¹

• ¹First Affiliated Hospital of Fujian Medical University, Fuzhou, China
• ²Heping Hospital Affiliated to Changzhi Medical College, Changzhi, China

Objective: To explore various factors influencing postoperative hearing recovery in patients with middle ear cholesteatoma and to construct and validate a clinical prediction model for postoperative hearing recovery. Methods: Clinical data from 548 patients diagnosed with middle ear cholesteatoma, gathered between May 2019 and December 2023, were randomly split into a training cohort and a validation cohort in a ratio of 7:3. To enhance feature selection, we utilized univariate logistic regression analysis, multivariate logistic regression analysis, and the Least Absolute Shrinkage and Selection Operator (LASSO) regression model to identify significant variables and develop the prediction model. The model's ability to predict outcomes was assessed using the Receiver Operating Characteristic (ROC) curve, while its clinical relevance was evaluated through calibration curves and clinical decision curves. Ultimately, the study findings were visually illustrated with a nomogram. Results: The findings from both univariate and multivariate logistic regression analyses suggest that several predictive factors are significant. These factors encompass the completeness of the ossicular chain, granulation tissue presence within the ossicular chain, the use of ossicular prostheses, eustachian tube functionality, instances of mixed hearing loss, ear conditions (either dry or wet), diabetes, and hypertension. For the training cohort, the area under the curve (AUC) was calculated to be 0.992 (95% CI 0.84–0.99), with the Hosmer-Lemeshow test yielding X2=10.54 and P=0.29. In the validation cohort, the AUC was 0.977 (95% CI 0.82–0.98), and the Hosmer-Lemeshow test revealed X2=8.54 and P=0.42. After implementing strict post-split preprocessing to mitigate overfitting and data leakage risks, the model was re-evaluated. The bootstrap-corrected AUC for the training cohort was 0.980 (95% CI: 0.82-0.99), and the cross-validated, optimism-corrected AUC for the validation cohort was 0.965 (95% CI: 0.80-0.98). A nomogram has been developed to visually forecast postoperative hearing recovery in individuals diagnosed with middle ear cholesteatoma. Additionally, the calibration curve, along with the clinical decision curve, indicates that this predictive model is both stable and trustworthy. Conclusion: This nomogram is an effective tool for predicting hearing recovery in patients with middle ear cholesteatoma, providing evidence-based support for clinical practice.