From Graphs to Reasoning in Medicine: The Role of Structured Knowledge in Deep Learning Systems

Deep learning has become a central methodology in biomedical research, yet its expansion into medicine is shaped by a fundamental tension: the most powerful neural architectures presuppose large, clean, well-annotated datasets, while clinical practice, constrained by pri- vacy regulations, the cost of expert annotation, and the rarity of many conditions, can rarely supply them. Medicine is, however, a domain of extraordinary opportunity in another re- spect: decades of clinical research have produced an exceptionally rich corpus of structured knowledge, encoded as ontologies, knowledge graphs, and decision pathways that articulate the very reasoning patterns that human clinicians follow. This thesis develops a research program centred on a single argument: that structured knowledge constitutes a principled and increasingly necessary inductive bias for deep learning systems applied to medicine, and that its role evolves along a coherent trajectory of increasing structural commitment. At the first stage, knowledge is implicit, and the burden of generalization falls entirely on data; here, data-centric techniques can extend what neural models achieve in scarce-label regimes, but they simultaneously reveal a structural ceiling that no amount of statistical refinement can transcend. At the second stage, knowledge becomes an explicit inductive bias, embed- ded into neural architectures so as to align their representational space with the functional organization of the domain; the resulting models are not merely more accurate, but more clinically coherent, in that their predictions respect the priorities encoded in established medical protocols. At the third stage, knowledge becomes the primary object of learning: when deep models are trained directly on biomedical knowledge graphs with appropriately chosen geometric inductive biases, their latent spaces inherit, by virtue of algebraic closure, a compositional inferential capacity that exceeds the training distribution and supports multi- step queries that no single training example could justify in isolation. At the fourth stage, knowledge becomes a normative yardstick: structured decision pathways drawn from clinical sources are turned into a controlled benchmark against which the reasoning capabilities of state-of-the-art Large Language Models are measured, revealing that the structural, and not merely the informational, character of graph-based knowledge is what closes the gap between parametric recall and stepwise clinical inference. Read in sequence, the four stages describe a single trajectory in which structured knowledge plays a progressively more central role, and in which a recurring principle emerges across very different methodological settings: that structure, more than scale, is the operative variable in domains characterised by limited data, strong prior organization, and high requirements of robustness, interpretability, and reason- ing fidelity. By tracing this trajectory from small clinical cohorts to large-scale biomedical knowledge graphs, and ultimately to the evaluation of how generative models reason when confronted with the very structured pathways that human clinicians follow, the thesis offers a unified perspective on how the next generation of biomedical artificial intelligence can be designed to be simultaneously data-efficient, robust, interpretable, and capable of reasoning in the structured terms that medicine itself demands.