Clinical decision support systems have a long history of limited adoption — not primarily because their recommendations are wrong, but because clinicians cannot see why the system made a specific recommendation, cannot assess whether its reasoning applied to their specific patient context, and cannot override it with confidence when their clinical judgement differs. Explainability is not a nice-to-have in clinical decision support; it is the mechanism through which trust is earned and maintained.
The Explainability Problem with Statistical Models
Machine learning models trained on clinical data can achieve impressive predictive accuracy on held-out test sets. Their limitation is that the internal representation of knowledge in a neural network — a high-dimensional vector of learned weights — does not correspond to any clinical concept that a human can evaluate. When such a model recommends against a specific drug combination, it cannot say why: it can offer feature attribution scores, but these are post-hoc approximations of model behaviour, not causal explanations of clinical reasoning.
Knowledge Graph Reasoning as Inherently Explainable
Recommendations derived from knowledge graph reasoning are inherently explainable because the reasoning is itself a sequence of explicit assertions. "This drug is not recommended for this patient because: (1) the drug has-contraindication renal-failure; (2) the patient has condition eGFR < 30 which is-subtype-of renal-failure; therefore (3) the contraindication applies." Each step in this reasoning chain can be presented to the clinician, traced to its source in the knowledge graph, and evaluated against the patient's actual clinical state. The clinician can agree, override, or request additional evidence — and that interaction can be logged as feedback for future knowledge graph improvement.
Combining Statistical and Symbolic Reasoning
The most capable clinical decision support architectures combine statistical models for pattern recognition — identifying which patient features are predictive of specific outcomes — with knowledge graph reasoning for the application of clinical rules, contraindications, and dosing guidelines. The statistical layer identifies risk; the knowledge layer contextualises it and generates the recommendation. The combination produces recommendations that are both data-driven and clinically interpretable — the standard that leading health systems and regulators are beginning to require for AI-assisted clinical tools.
