Interpretable Machine Learning for Identifying High SBS-Related Symptom Burden Among Endoscopy-Center Staff: A Multicenter Cross-Sectional Study Across 184 Hospitals

Technical Summary: Interpretable Machine Learning for Identifying High SBS-Related Symptom Burden Among Endoscopy-Center Staff

Problem Statement
Endoscopy center staff face complex occupational health risks due to chronic exposure to chemical disinfectants (e.g., glutaraldehyde, ortho-phthalaldehyde), biological hazards, and aerosols. While previous research has focused on diagnosed diseases, workers often experience persistent, nonspecific discomfort (fatigue, irritation, sleep disturbances) prior to overt disease onset. This state, characterized by Sick Building Syndrome (SBS)-related symptoms, is multifactorial, involving interactions between chemical exposure, indoor air quality, and psychosocial stress. Existing studies are limited by their focus on single exposure factors, single-center designs, and the use of conventional linear models that struggle with the non-linear, high-dimensional, and correlated nature of real-world occupational data. There is a need for a multidimensional approach to identify high-risk individuals and actionable determinants for early intervention.

Methodology
The study employed a multicenter cross-sectional design involving 319 valid questionnaires from endoscopy-center staff across 184 hospitals in five Chinese provinces.

• Data Collection & Outcome Definition: A structured questionnaire assessed demographics, lifestyle, health status, workplace environment, and work-related perceptions. The primary outcome was a "total SBS-related symptom burden score" derived from 29 symptom items (ocular, respiratory, neurologic/sleep, dermatologic, and gastrointestinal/emotional) reported over the previous two weeks. Scores were dichotomized using the sample-specific 75th percentile (P75 = 24.5) to define a "high symptom burden" group (score ≥ 24.5) versus a low burden group.
• Feature Selection: Recursive Feature Elimination with Cross-Validation (RFECV) using a Support Vector Machine (SVM) as the base estimator was employed to select the most predictive variables from the initial pool. This process reduced the feature set to 12 candidates, which were further refined to 10 based on SHAP importance analysis (excluding gender and drinking status which showed zero importance).
• Modeling Strategy: Nine supervised learning algorithms (XGBoost, Logistic Regression, LightGBM, Random Forest, AdaBoost, KNN, SVM, Decision Tree, GBDT) were compared using five-fold cross-validation. Hyperparameters were tuned via grid search. Logistic Regression was selected as the final model due to its superior Area Under the Curve (AUC) and calibration (lowest Brier score) on validation folds.
• Interpretability: The final model was interpreted using SHapley Additive exPlanations (SHAP) to quantify feature contributions at both global (population-level importance) and individual levels.
• Validation: The dataset was split into a training set (70%) and an independent test set (30%). Model performance was evaluated using AUC, accuracy, sensitivity, specificity, positive/negative predictive values, F1 score, and Cohen's kappa.

Key Results

• Participant Characteristics: The cohort was predominantly female (93.7%) with a median age of 34. The high symptom burden group (25.1% of participants) reported significantly lower satisfaction with workplace temperature, humidity, and fresh air, higher job stress, and poorer sleep quality compared to the low burden group.
• Model Performance: Logistic Regression demonstrated the best performance. In the independent test set, the model achieved:
  • AUC: 0.891
  • Accuracy: 0.823
  • Sensitivity: 0.750
  • Specificity: 0.847
  • Negative Predictive Value (NPV): 0.910
  • Positive Predictive Value (PPV): 0.621
  • F1 Score: 0.679
  • Cohen's Kappa: 0.558
    The model showed good calibration (Brier score: 0.113) and decision curve analysis indicated a net benefit over "treat-all" or "treat-none" strategies across a broad range of threshold probabilities.
• Key Predictors (SHAP Analysis): The most influential predictors for high symptom burden were:
  1. Workplace odor (highest impact)
  2. Perceived air quality
  3. Coworker similar symptoms
  4. Job stress
  5. Satisfaction with fresh-air supply
     Other significant factors included sleep quality, home ventilation, sleep duration, hospital visits for symptoms, and pre-existing diseases. The analysis revealed that adverse conditions (e.g., strong odors, poor air quality, high stress) shifted predictions toward high symptom burden, while protective factors (e.g., good sleep, no coworker symptoms) shifted predictions toward low burden.

Significance and Claims
The authors claim that this study successfully developed and validated an interpretable machine learning model capable of identifying endoscopy-center staff with a high SBS-related symptom burden at a pre-disease stage. The significance of the work lies in three main areas:

1. Multidimensional Integration: The model moves beyond single-factor analysis, demonstrating that high symptom burden arises from the cumulative impact of workplace air conditions, psychosocial load, and individual recovery status.
2. Actionable Insights: By using SHAP, the study identifies specific, actionable determinants (e.g., odor, fresh air supply, job stress) rather than just abstract risk scores. This allows for targeted interventions, such as optimizing ventilation management and addressing psychosocial stressors.
3. Practical Utility: The model relies on variables obtainable through routine questionnaires, suggesting it can serve as a practical tool for early occupational health risk stratification. It enables the detection of department-level hazards (via "coworker similar symptoms") and individual risk screening, facilitating proactive health surveillance and environmental optimization in endoscopy centers.

The authors maintain a modest stance, acknowledging that the cross-sectional design precludes causal inference and that the reliance on self-reported environmental data may introduce subjective bias. They note that the model's generalizability to other occupational roles or healthcare settings requires further evaluation.