Handling varying amounts of missing data when classifying mental-health risk levels

Area of Science:

• Clinical Data Science
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Classifying clinical data presents challenges due to missing features.
• Traditional methods like imputation or data deletion can reduce accuracy with high missingness.
• Existing approaches struggle with datasets having substantial missing values and high variability.

Purpose of the Study:

• To propose a new methodology for handling clinical datasets with a high percentage of missing values.
• To develop a classification approach that accommodates high variability in missing data patterns.
• To improve the accuracy and interpretability of predictive models in clinical settings.

Main Methods:

• Sequential feature selection based on correlation with the dependent variable and selected features.
• Individual classification model generation for each test case using available training data.
• Application to real-world mental health data for suicide risk prediction.

Main Results:

• The proposed method demonstrates favorable comparison with alternative approaches.
• The methodology effectively handles datasets with significant missing data and high variability.
• Explanations for risk predictions are derived solely from provided data, not imputed values.

Conclusions:

• The new methodology offers an effective solution for classifying clinical data with extensive missing features.
• This approach enhances the reliability of clinical decision support systems by using actual patient data.
• The method ensures transparency and interpretability in predicting clinical outcomes like suicide risk.