Artificial intelligence (AI) diagnostic systems demonstrate exceptional performance in controlled laboratory settings yet consistently fail to translate into equitable and reliable clinical tools. This thesis identifies and analyzes the structural roots of this translation gap, arguing that the pervasive challenges of algorithmic bias and poor generalizability are not isolated technical failures but predictable outcomes of a development paradigm that prioritizes narrow accuracy metrics over robust, equitable performance. Through a systematic analysis of evidence across medical specialties, this research demonstrates how models trained on geographically concentrated and demographically homogeneous data systematically underperform for marginalized populations and fail when deployed in new contexts. The compounding of bias (differential performance across groups) and poor generalizability (performance degradation across settings) creates an "equity paradox" wherein AI tools perform best for populations with the least need and worst for those who could benefit most from improved diagnostic access. This thesis reveals how current regulatory frameworks, economic incentives, and organizational structures actively reinforce these problematic practices. It moves beyond technical mitigation strategies to propose a fundamental reorientation of the AI development lifecycle that centres equity and generalizability as non-negotiable requirements. The proposed framework includes proactive data diversity, mandatory multi-site and intersectional validation, fairness-aware optimization, and robust governance structures. The findings necessitate a paradigm shift from accuracy-focused to equity-centred AI development, with implications for researchers, regulators, healthcare institutions, and policymakers. Ultimately, this thesis contends that the technical capacity for building equitable AI diagnostics exists; what is required is the collective commitment to treat equity not as an aspirational goal but as a fundamental criterion for clinical deployment.
Johnson Gbenga Oyeniyi (Wed,) studied this question.