PERform: assessing model performance with predictivity and explainability readiness formula.

Area of Science:

• Artificial Intelligence
• Computational Toxicology
• Cheminformatics

Background:

• Traditional AI model evaluation often separates performance and explainability, leading to subjective assessments.
• Existing quantitative metrics primarily focus on model performance, neglecting interpretability.
• There is a need for integrated, quantitative measures to assess both predictivity and explainability.

Purpose of the Study:

• To introduce PERForm, a unified formula for quantitatively measuring both predictivity and explainability in AI models.
• To provide a standardized method for evaluating and selecting AI models based on integrated performance and interpretability.
• To advance transparency and interpretability in artificial intelligence applications.

Main Methods:

• Developed the PERForm formula, incorporating explainability as a weighting factor into existing statistical performance metrics.
• Applied the generic PERForm formula across diverse datasets (DILIst, Tox21, MAQC-II) and various modeling algorithms.
• Evaluated 73 distinct endpoints to demonstrate the formula's applicability and utility.

Main Results:

• PERForm successfully integrated explainability into quantitative AI model assessment.
• Demonstrated varied model performances across datasets; AdaBoost excelled in DILIst prediction, while linear regression was superior for most Tox21 endpoints.
• Provided quantitative evidence of the trade-offs between model performance and explainability.

Conclusions:

• PERForm offers a robust, quantitative framework for evaluating AI models, balancing predictivity and explainability.
• This approach facilitates more informed model selection, application, and development.
• The research significantly contributes to enhancing transparency and interpretability in artificial intelligence.