A machine-learning-derived online diagnostic model for Parkinson’s disease integrating eye-tracking features and clinical indicators: A single-center case-control study

Technical Summary: A Machine-Learning-Derived Online Diagnostic Model for Parkinson's Disease

Problem Statement
Parkinson's disease (PD) requires early and accurate diagnosis to facilitate timely intervention. Traditional diagnostic methods often rely on clinical observation and subjective assessments, which can be delayed until motor symptoms are prominent. This study addresses the need for an objective, non-invasive diagnostic tool by integrating eye-tracking features with standard clinical indicators. The research aims to develop a machine-learning-based model capable of distinguishing PD patients from controls with high accuracy, leveraging the known oculomotor deficits associated with the disease.

Methodology
The study employed a single-center case-control design involving participants from multiple hospitals in Guangdong, China. The dataset included both clinical indicators (demographics, cognitive scores such as MMSE, MoCA, HAM-A, HAM-D, PSQI) and high-dimensional eye-tracking features derived from saccadic and smooth pursuit tasks (e.g., overlap saccade accuracy/latency/velocity, anti-saccade metrics).

• Data Preprocessing and Feature Selection:

  • Univariate Analysis: Initial screening was conducted using univariate logistic regression to identify significant predictors. Variables such as educational level, visuospatial/executive function, attention, language, delayed recall, orientation, MoCA total score, and specific eye-tracking metrics (e.g., mean overlap saccade velocity, anti-saccade correction rate) showed statistical significance.
  • Multicollinearity Handling: A Variance Inflation Factor (VIF) analysis revealed severe multicollinearity in the MoCA total score (VIF = 45.402) and moderate multicollinearity in several sub-scores. The study acknowledged these correlations, noting that variables with severe collinearity are typically candidates for removal or combination to ensure model stability.
  • Imputation and Balancing: The study compared model performance using Multiple Imputation by Chained Equations (MICE) versus complete case analysis, finding that MICE yielded slightly higher specificity and AUC. Additionally, the impact of SMOTE (Synthetic Minority Over-sampling Technique) was evaluated; models trained with SMOTE showed improved specificity and balanced metrics compared to those without.

• Model Development:

  • Multiple machine learning algorithms were trained and compared: Logistic Regression (LR), Random Forest (RF), XGBoost, Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), and Decision Tree (DT).
  • Explainability: The Random Forest model, which demonstrated superior performance, was interpreted using SHAP (SHapley Additive exPlanations) values to quantify feature importance and contribution to predictions.

• Validation:

  • The models were evaluated on an independent test set.
  • Performance metrics included Sensitivity, Specificity, Positive/Negative Predictive Values (PPV/NPV), Precision, Recall, F1-score, Accuracy, Kappa, AUC (Area Under the ROC Curve), and PR-AUC (Precision-Recall AUC).
  • Statistical significance between model performances was assessed using pairwise DeLong tests.
  • Subgroup analyses were conducted based on age and gender to assess robustness across demographics.

Key Results

• Model Performance: The Random Forest (RF) model emerged as the top performer. In the independent test set, the RF model achieved an AUC of 0.9711, significantly outperforming Logistic Regression (AUC 0.8438, p=0.002), XGBoost (AUC 0.9306, p=0.094), and SVM (AUC 0.9144, p=0.024).
• Training Set Metrics: In the training set, the RF model demonstrated an AUC of 0.996 (95% CI 0.991-1.000), a PR-AUC of 0.9973, and a Brier score of 0.033, indicating excellent calibration and discrimination.
• Feature Importance: The univariate analysis identified that lower scores in cognitive domains (visuospatial/executive function, attention, language, delayed recall, orientation, MoCA) and specific eye-tracking anomalies (e.g., reduced mean overlap saccade velocity, increased mean overlap saccade duration, reduced anti-saccade correction rate) were strongly associated with PD.
• Robustness:
  • SMOTE: The use of SMOTE oversampling improved the model's specificity (0.9560 vs. 0.9121) and F1-score (0.9537 vs. 0.9412).
  • Imputation: MICE imputation resulted in a perfect specificity (1.000) and PPV (1.000) in the test set, with an AUC of 0.998, compared to 0.996 without imputation.
  • Subgroups: The model maintained high performance across age groups (AUC 0.963–0.998) and genders (AUC 0.957–0.979), suggesting generalizability within the studied cohort.

Significance and Claims
The paper claims to present a robust, machine-learning-derived online diagnostic model for Parkinson's disease that effectively integrates eye-tracking biomarkers with clinical indicators. The primary contribution is the demonstration that a Random Forest model, utilizing these combined features, can achieve high diagnostic accuracy (AUC > 0.97) in an independent test set.

The study emphasizes the utility of eye-tracking features as objective markers that complement cognitive assessments. By employing SHAP analysis, the model offers interpretability, highlighting which specific oculomotor and clinical variables drive the diagnosis. The authors position this work as a step toward objective, non-invasive diagnostic tools that could potentially aid in the early detection of PD, supported by rigorous statistical validation including multicollinearity checks, multiple imputation strategies, and subgroup analyses. The work is framed as a single-center case-control study, with the results suggesting a strong potential for clinical application pending further external validation.