2026 Proffered Presentations
S292: PREDICTING SURGICAL OUTCOMES IN SUPERIOR SEMICIRCULAR CANAL DEHISCENCE USING MACHINE LEARNING
Aryan Pandey; Anubhav Chandla; Mahlet Mekonnen; Laila Khorasani; Simon Han; Quinton S Gopen; Isaac Yang; UCLA
Introduction: Superior semicircular canal dehiscence (SSCD) is characterized by thinning or absence of the bone covering the superior semicircular canal, resulting in debilitating auditory and vestibular symptoms. Surgical repair is the standard treatment, but predicting postoperative outcomes from preoperative clinical and demographic variables remains a challenge.
Objective: To evaluate the performance of machine learning (ML) algorithms in predicting postoperative outcomes for SSCD patients and identify key predictive features.
Methods: A retrospective cohort of 600 surgically treated SSCD patients was analyzed. Clinical and demographic factors were used as model inputs. Predictive algorithms included gradient boosting, random forests, logistic regression, and support vector machines. Model performance was assessed using accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC), validated with 10-fold cross-validation.
Results: Gradient boosting achieved the best predictive performance (AUC = 0.87, accuracy = 82%, sensitivity = 85%, specificity = 79%), outperforming random forests (AUC = 0.84, accuracy = 79%), logistic regression (AUC = 0.78, accuracy = 74%), and support vector machines (AUC = 0.76, accuracy = 72%). Feature importance analysis revealed that patient age strongly influenced predictions, although outcome prediction relied more on combined feature patterns than on individual variables.
Conclusions: ML algorithms, particularly gradient boosting, demonstrated robust accuracy in predicting postoperative outcomes in SSCD. These predictive tools hold promise for enhancing surgical planning, improving patient counseling, and advancing personalized care. Prospective, multi-center validation studies are warranted to confirm clinical applicability and facilitate integration of ML-driven decision support into SSCD management.
Table 1. Demographics and symptom presentation of the cohort of SSCD repair patients.
Table 2. Machine learning model analysis of predicting symptom resolution based on demographic and symptom presentation factors.
Table 3. Risk decile calibration table demonstrating the mean predicted risk of developing dizziness versus the observed improvement by decile for out-of-fold predictions.
Table 4. Prediction breakdown by model for true positive, false positive, true negative, and false negative along with precision (PPV) and NPV.