Deep Learning-Assisted Prediction of Hearing Outcomes After Anatomically Successful Type I Tympanoplasty

IMPORTANCE: Type I tympanoplasty restores hearing in patients with simple tympanic membrane (TM) perforations, but reliable tools to predict postoperative outcomes remain limited. OBJECTIVE: To develop and evaluate a deep learning-assisted model integrating automated TM image features and clinical data to predict postoperative air-bone gap (ABG) closure and residual ABG. DESIGN: Diagnostic and prognostic model development and validation study. SETTING: A tertiary referral medical center in northern Taiwan. PARTICIPANTS: A total of 1285 otoendoscopic images were collected, of which 1014 intact and 150 perforated TMs were used to train the mask region-based convolutional neural network (Mask R-CNN) segmentation model. Prognostic analysis included 121 patients with simple perforations and anatomically successful type I tympanoplasty (complete TM closure), with 83 preoperative images for training and 38 for independent internal testing. Demographic, clinical, and audiometric data were recorded.Intervention or Exposures:Automated image features extracted by Mask R-CNN, combined with clinical and audiometric variables, were used to develop prognostic models. MAIN OUTCOME MEASURES: ), root mean square error (RMSE), mean absolute error, and predictive accuracy, defined as a predicted ABG within 10 and 5 dB of the measured value. RESULTS: values of 0.418 for ABG closure and 0.363 for residual ABG, with corresponding RMSEs of 4.39 and 4.36 dB. Prediction accuracy reached 97% within 10 dB and 74% within 5 dB, significantly outperforming baseline mean-value prediction (P < .05). CONCLUSIONS: Deep learning-assisted analysis of TM images showed modest predictive ability for hearing outcomes after anatomically successful type I tympanoplasty. RELEVANCE: This image-based approach may modestly assist preoperative counseling in otologic practice.