Technical Summary: Interpretable Machine Learning for Subtype-Specific Prognosis in Sudden Hearing Loss

Problem Statement

Sudden sensorineural hearing loss (SSNHL) is a heterogeneous clinical condition characterized by unpredictable recovery rates. Current management is complicated by the lack of precision in prognostic assessment, as existing models often function as "black boxes" that fail to account for distinct audiometric subtypes or provide actionable, interpretable thresholds for clinical intervention. While factors such as age, vertigo, and hearing loss severity are known to influence outcomes, the specific pathophysiological mechanisms and prognostic trajectories vary significantly across different audiogram configurations (e.g., ascending, descending, flat, profound, and atypical). There is a critical need for a high-precision, interpretable machine learning (ML) framework that can stratify patients by subtype and quantify the nonlinear impact of key predictors to guide timely, evidence-based interventions.

Methodology

The study employed a retrospective cohort design utilizing the largest reported clinical dataset for SSNHL to date, comprising 3,957 patients diagnosed between 2010 and 2017 at the Eye & ENT Hospital of Fudan University.

Data Processing and Subtyping:
Patients were stratified into five distinct cohorts based on audiogram configuration using operational definitions:

1. Ascending: Low-frequency loss (threshold difference ≥20 dB HL where 250/500 Hz > 4000/8000 Hz).
2. Descending: High-frequency loss (threshold difference ≥20 dB HL where 4000/8000 Hz > 250/500 Hz).
3. Flat: Uniform impairment across 250–8000 Hz with Pure-Tone Average (PTA) ≤90 dB HL.
4. Profound: Severe-to-profound loss across all frequencies with PTA >90 dB HL.
5. Atypical: Configurations not meeting the above criteria.

Model Development:

• Feature Selection: A composite approach utilizing Random Forest intrinsic importance, Lasso regression, and Recursive Feature Elimination (RFE) identified the top 10 predictive features. A core set of five factors—disease duration, onset PTA, age, WHO hearing classification, and systemic glucocorticoid therapy—was consistently ranked highest across subtypes.
• Algorithm Training: Eight ML algorithms were trained and validated for each subtype. The data was split into an 80% training set and a 20% hold-out test set. Class imbalance was addressed using Synthetic Minority Over-sampling Technique (SMOTE). Hyperparameter tuning and internal validation were performed via 5-fold cross-validation within the training set.
• Interpretability: To overcome the "black-box" nature of ML, the study employed SHapley Additive exPlanations (SHAP) for feature contribution analysis, Partial Dependence Plots (PDPs) to visualize nonlinear relationships, and Individual Conditional Expectation (ICE) plots to assess heterogeneity. Bootstrap resampling was used to determine the statistical significance and stability of identified threshold points.

Evaluation Metrics:
Model performance was assessed using accuracy, Area Under the Receiver Operating Characteristic Curve (AUC), Brier score, sensitivity, and specificity. Treatment outcomes were dichotomized based on Siegel criteria into "recovery" (complete or partial) and "non-recovery" (slight improvement or no improvement).

Key Results

Model Performance:
The study successfully developed subtype-specific models with robust performance.

• Optimal Algorithms: Support Vector Machine (SVM) emerged as the optimal classifier for ascending, descending, and atypical subtypes, while Logistic Regression was optimal for flat and profound subtypes.
• Accuracy: Test accuracies ranged from 0.706 (profound) to 0.774 (descending).
• Best Performance: The SVM model for the descending subtype achieved the highest test accuracy (0.774), the lowest Brier score (0.146), and a test AUC of 0.828. The Logistic Regression model for the flat subtype achieved a test accuracy of 0.732 and an AUC of 0.782.

Prognostic Factors and Thresholds:

• Core Predictors: Five factors were consistently identified as the most influential: disease duration, pure-tone average (PTA), age, WHO hearing classification, and systemic glucocorticoid therapy.
• Nonlinear Dynamics: PDP and ICE analyses revealed significant nonlinear declines in recovery probability as disease duration and initial PTA increased.
• Actionable Thresholds: The study quantified specific treatment windows where the probability of recovery drops precipitously. The threshold for a 50% recovery probability varied by subtype, ranging from 7.5 to 13.1 days.
  • Example: In the atypical subtype, patients with a disease duration ≤10.6 days had a 43.6% recovery rate, which dropped to 8.5% for those exceeding 18.4 days.
  • Exception: The profound subtype showed uniformly poor recovery rates (≤4.6%) across all duration groups, indicating that disease duration was not a significant prognostic factor for this specific group (p=0.639).
• Treatment Impact: Systemic glucocorticoid therapy showed a clear positive association with recovery, particularly in atypical and ascending subtypes.

Deployment:
The models were integrated into a locally deployed clinical decision support tool ("Precision Prognosis System for Sudden Hearing Loss Subtypes") via Flask, providing real-time, patient-specific prognostic probabilities.

Key Contributions

1. Subtype-Specific Precision: Unlike previous studies that treat SSNHL as a monolithic condition, this work establishes distinct, validated ML models for five specific audiometric subtypes, acknowledging their likely different underlying pathophysiologies.
2. Interpretability and Transparency: By utilizing SHAP, PDPs, and ICE plots, the study deconstructs the "black-box" nature of ML, providing clinicians with transparent insights into how specific variables influence outcomes.
3. Quantification of Therapeutic Windows: The research moves beyond general associations to define statistically validated, nonlinear threshold points (e.g., specific day counts) where the likelihood of recovery significantly diminishes, offering concrete guidance for early intervention.
4. Large-Scale Validation: The study leverages the largest reported SSNHL cohort (n=3,957) with rigorous internal validation, addressing common limitations in existing prognostic models such as small sample sizes and lack of robust validation.

Significance and Claims

The authors claim that this framework represents a significant advancement in the precision management of SSNHL. By translating complex algorithmic predictions into actionable, subtype-specific clinical insights, the study provides a tool that reinforces the necessity of early systemic glucocorticoid therapy. The identification of specific, quantifiable intervention thresholds allows otologists to move beyond probabilistic uncertainty toward evidence-based, timely decision-making.

The study positions itself as a step toward "precision medicine" in otology, offering a real-time decision support system that can assist in triaging patients and tailoring therapeutic approaches. However, the authors maintain a modest stance regarding limitations, acknowledging the single-center, retrospective nature of the data and the need for multi-center, prospective external validation to confirm the generalizability of the identified prognostic thresholds. They explicitly state that while the tool offers immediate clinical utility, further validation is required before widespread adoption.