Supporting AI-Explainability by Analyzing Feature Subsets in a Machine Learning Model.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Computational Medicine

Background:

• Machine learning (ML) algorithms are increasingly used in medicine for pattern recognition in complex data.
• The "black box" nature of many ML models raises concerns regarding their use in sensitive medical applications.
• A need exists for transparent and interpretable ML models in clinical settings.

Purpose of the Study:

• To demonstrate the benefits of aggregated and systematic feature analysis for ML models in medicine.
• To enhance the interpretability of ML models used in medical decision-making.
• To facilitate the formulation and examination of new clinical hypotheses based on model insights.

Main Methods:

• Introduction of a grouped permutation importance analysis.
• Evaluation of the influence of entire feature subsets within ML models.
• Application of expert-defined subgroups for assessing decision-making processes.

Main Results:

• The grouped permutation importance analysis provides a systematic way to evaluate feature subsets.
• This method allows for the assessment of clinically relevant feature groups.
• Results can guide the understanding of how specific feature combinations impact model predictions.

Conclusions:

• Aggregated feature analysis, using the proposed method, improves the transparency of ML models in medicine.
• This approach supports the validation of expert knowledge and aids in generating novel research questions.
• Enhanced interpretability of ML models is crucial for their safe and effective adoption in healthcare.