Evaluating Machine Learning Stability in Predicting Depression and Anxiety Amidst Subjective Response Errors.

Area of Science:

• Computational psychiatry
• Machine learning in healthcare
• Mental health informatics

Background:

• Major Depressive Disorder (MDD) and Generalized Anxiety Disorder (GAD) significantly impact individuals and society.
• Accurate prediction of MDD and GAD is crucial for timely intervention and treatment.
• Machine learning (ML) models using electronic health records and survey data show promise for forecasting these conditions, but are vulnerable to subjective data inaccuracies.

Purpose of the Study:

• To evaluate the reliability of five ML algorithms in predicting MDD and GAD under varying levels of subjective survey response inaccuracies.
• To identify ML algorithms that demonstrate resilience and maintain predictive accuracy when faced with data unreliability.

Main Methods:

• Assessed five ML algorithms: Convolutional Neural Network (CNN), Random Forest, XGBoost, Logistic Regression, and Naive Bayes.
• Utilized a dataset comprising biomedical, demographic, and self-reported survey information.
• Simulated subjective response inaccuracies (memory recall bias, subjective interpretation) to test algorithm performance.

Main Results:

• All algorithms performed well with high-quality survey data.
• Performance diverged significantly when encountering erroneous or biased responses.
• The CNN demonstrated superior resilience, maintaining and even enhancing accuracy, Cohen's kappa, and positive precision for both MDD and GAD prediction.

Conclusions:

• Algorithmic resilience is critical for accurate mental health prediction, especially with subjective self-reported data.
• The CNN shows a robust capability to handle data unreliability in mental health predictions.
• Careful algorithm selection is essential, with the CNN emerging as a promising candidate for predicting MDD and GAD under data uncertainties.