Machine Learning and Explainable Artificial Intelligence Using Counterfactual Explanations for Evaluating Posture...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Medical Imaging

Background:

• Postural deficits like hyperlordosis and hyperkyphosis present diagnostic challenges due to subjective assessments.
• Existing machine learning (ML) approaches for posture analysis have limited human-friendly interpretations.
• Explainable artificial intelligence (XAI) offers potential for objective, data-driven medical decision support.

Purpose of the Study:

• To develop an objective, data-driven ML system for diagnosing hyperlordosis and hyperkyphosis.
• To enhance ML interpretability using counterfactual explanations (CFs) for user-friendly insights.
• To improve diagnostic accuracy and support personalized therapeutic adaptations.

Main Methods:

• Collected posture data from 1151 subjects using stereophotogrammetry.
• Employed a Gaussian process classifier trained on expert-classified posture data.
• Utilized counterfactual explanations (CFs) for model interpretation and confident learning for label re-evaluation.

Main Results:

• Achieved very good classification performance for both hyperlordosis and hyperkyphosis.
• Re-evaluation and correction of test labels significantly improved classification performance (M_PRAUC = 0.97).
• Statistical evaluation indicated the plausibility of the counterfactual explanations.

Conclusions:

• The proposed ML system offers an objective approach to diagnosing postural deficits, reducing diagnostic errors.
• The human-friendly CF interpretations can aid clinicians in decision-making and personalized treatment planning.
• This approach could form the basis for preventive posture assessment applications.