Longitudinal data preprocessing for regression-based machine learning models in idiopathic sudden sensorineural hearing loss

Idiopathic sudden sensorineural hearing loss (ISSNHL) shows highly variable recovery, and clinicians lack tools to provide individualized, frequency-specific prognostic information. Most previous machine learning studies in ISSNHL have focused on binary outcomes using baseline features at a single time point. This retrospective single-center study reframed ISSNHL prognosis as a frequency-specific, multi-output regression task and used longitudinal data preprocessing (LDP) to encode irregular follow-up timing and cumulative treatment exposure. We analyzed 408 eligible ISSNHL patients from 592 initially identified cases; LDP transformed their irregular follow-up visits into 656 visit-level instances. A temporal patient-level hold-out split was used. In the primary six-frequency regression analysis, XGBoost achieved the highest mean frequency-wise R 2 (0.7017), followed by CatBoost (0.6863) and Random Forest (0.6758). After mapping continuous predictions to Siegel's criteria, Random Forest and XGBoost achieved the highest Siegel accuracy (70.40%), XGBoost achieved the highest Siegel macro-F1 (0.64), and Random Forest achieved the best binary endpoint performance (accuracy 73.60%, F1-score 0.73, balanced accuracy 0.75). Additional analyses incorporated dB-scale regression errors, patient-level bootstrap uncertainty, class-distribution checks, calibration analyses, direct-classification baselines, and sensitivity analyses for target scope and within-patient dependence; these analyses were used to contextualize the primary held-out results rather than replace them. SHAP-based explainability highlighted baseline hearing thresholds, age, time to diagnosis, intratympanic steroid therapy, and systemic steroid therapy as predictive features. Treatment-related SHAP findings were interpreted as predictive associations rather than causal interpretations because confounding by indication is likely. Overall, the proposed framework provides an interpretable treatment- and time-updated approach for frequency-specific ISSNHL prognosis, but prospective multi-center validation is required before clinical decision-support use.