INTELLIGENT HEALTHCARE DATA LAKEHOUSES: INTEGRATING MEDALLION ARCHITECTURE, DATA QUALITY, AND AI-READY ANALYTICS

Healthcare organisations hold large volumes of clinical, administrative, and claims data, yet routinely find that data unusable for analytics or machine learning without substantial remediation. The obstacles are structural: heterogeneous source semantics, incomplete and temporally inconsistent records, weak provenance, and privacy constraints that limit the free movement of data across environments. This paper presents a layered lakehouse architecture for healthcare analytics that combines a bronze–silver–gold (medallion) refinement model with an embedded, declarative data-quality regime and a terminology standardisation stage aligned to established interoperability and common-data-model standards. The contribution is threefold. First, we specify what each medallion layer must guarantee in a healthcare context, replacing the informal "raw / cleaned / curated" description with explicit per-layer contracts covering immutability, identity resolution, temporal semantics, terminology binding, and de-identification state. Second, we define a data-quality framework mapping the harmonised electronic health record quality dimensions of conformance, completeness, and plausibility onto executable assertions positioned at specific layer boundaries, with quarantine rather than silent drop as the default failure action. Third, we specify AI-readiness criteria for the gold layer - point-in-time correctness, label provenance, cohort reproducibility, and documented population coverage - that distinguish a table that is merely queryable from one that can responsibly support model development. We describe implementation on a transactional lakehouse, propose an evaluation protocol, and analyse limitations including the persistence of unstructured clinical narrative and the risk that quality filtering itself introduces population bias.