Explainability Enhanced Machine Learning Model for Classifying Intellectual Disability and...

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing (NLP)

Background:

• Psychological test reports are crucial for diagnosing intellectual disability (ID) and attention-deficit/hyperactivity disorder (ADHD).
• Current reports are often unstructured, subjective, and prone to human error, leading to diagnostic challenges.
• Physicians may not review full reports, and report volume often trails diagnostic needs.

Purpose of the Study:

• To develop explainable predictive models for classifying ID and ADHD from patient reports.
• To address limitations of traditional report analysis using NLP and AI.

Main Methods:

• Utilized NLP to analyze 1,475 patient reports, incorporating physician diagnoses.
• Employed SHAP and permutation importance for feature selection, ensuring model explainability.
• Developed an n-gram feature-based search system to reconstruct human-readable text from model outputs.

Main Results:

• Achieved a maximum model accuracy of 0.92 in classifying ID and ADHD.
• Successfully restored 80 human-readable texts from four distinct models.
• Demonstrated accurate classification even with limited report data.

Conclusions:

• The developed models accurately classify ID and ADHD, offering explainable predictions.
• Enhanced model explainability aids physician understanding and supports evidence-based decision-making.
• This approach offers a scalable solution for improving diagnostic efficiency and accuracy.