An Artificial Intelligence-based model for cell killing prediction: development, validation and explainability analysis of the ANAKIN model

ANAKIN demonstrates superior predictive accuracy compared to established clinical models: The AI model outperformed the Microdosimetric Kinetic Model (MKM) and Local Effect Model (LEM version III)—the only two radiobiological models currently used in clinical practice—across all tested biological endpoints for predicting cell survival after radiation exposure.

The model was trained on a substantial and diverse dataset: ANAKIN was developed using 513 cell survival experiments from the publicly available PIDE database, covering multiple types of ion beams and numerous cell lines, which supports the generalizability of its predictions across different biological contexts.

Explainability analysis confirms the model learns genuine biological principles: Using Explainable AI (XAI) techniques, researchers demonstrated that ANAKIN's predictions can be understood and interpreted, and importantly, that the model reproduces well-known biological patterns such as the "overkilling effect"—suggesting it captures real biological mechanisms rather than relying on spurious correlations.

This work bridges the gap between AI performance and clinical interpretability: By combining superior predictive accuracy with explainability analysis, ANAKIN addresses a key concern in responsible AI adoption in healthcare: the ability to understand why an AI system makes its predictions, which is essential for clinical validation and physician trust.