On Interpretability in Multimodal Biomedical Image Analysis

InterpretabilityInterpretabilityis a critical property for machine learning models used in high-stakes biomedical applications, where understanding the rationale behind automated decisions is essential for clinical trust, regulatory compliance, and ethical accountability. In this chapter, we argue that interpretabilityinterpretabilityshould be understood not as a binary label, but as a continuum that reflects the structural, semanticsemantic, and cognitive accessibility of a model. We focus specifically on two core dimensions: simulatabilitysimulatability, the ability of a human to mentally reproduce a model’s reasoningreasoningprocess, and decomposabilitydecomposability, the extent to which a model can be broken down into meaningful and inspectable components. Through case studies involving symbolic approaches, specifically genetic programming pipelines generated with the MultimodalMultimodalAdaptive Graph Evolution method, we demonstrate how interpretabilityinterpretabilitycan emerge by design, particularly when models are constructed using human-readable operations and type-safe modular structures. We contrast these with deep convolutional neural networkneural networkmodels where interpretabilityinterpretabilityremains limited despite post hoc explanation methods such as Grad-CAM. Finally, drawing on insights from software engineering and system design, we explore how decomposabilitydecomposabilitycan be formalized and measured, laying the groundwork for optimization strategies that balance predictive performance with structural transparency.