Artificial Intelligence for Physiologically Informed Precision Care in Sickle Cell Disease: Diagnosis, Risk Prediction, Personalised Treatment and Equitable Implementation

Abstract

Sickle cell disease (SCD) is a monogenic haemoglobinopathy whose clinical expression is shaped by multiscale interactions among haemoglobin S polymerisation, erythrocyte rheology, haemolysis, inflammation, vascular dysfunction, organ reserve, treatment exposure and social context. Artificial intelligence (AI) is increasingly being applied to these heterogeneous data, but high algorithmic performance does not by itself establish clinical usefulness. This critical narrative review evaluates how AI can support physiologically informed precision care across diagnosis, phenotyping, prediction of vaso-occlusive and organ complications, personalised treatment, curative-therapy decision support and equitable implementation. Literature published from 1 January 2015 through 28 June 2026 was identified through multidisciplinary biomedical and scholarly sources, supplemented by citation searching and a South-East Asian regional index. The evidence is strongest for automated screening and image-based phenotyping, where deep-learning systems can classify sickle morphology or haemoglobin-separation patterns with high internal accuracy. Yet many studies depend on small or reused datasets, image-level rather than patient-level validation, and limited external testing. Predictive models based on physiological signals, electronic health records and multimodal clinical variables show clinically plausible associations with pain, acute organ failure, kidney decline, readmission and mortality, but calibration, transportability and prospective utility remain insufficiently established. Personalised treatment is less mature: pharmacokinetic-guided hydroxyurea dosing provides a mechanistically grounded precision-care benchmark, whereas SCD-specific AI treatment-selection evidence remains sparse, although transplant outcome models are emerging. Biophysical imaging, erythrocyte dynamics, metabolomics and genetic modifiers offer a route towards models that represent disease mechanisms rather than correlations alone. Equitable translation is a central validity requirement because SCD burden is concentrated in populations and health systems under-represented in most AI development datasets, while physiological sensors and healthcare-derived labels can encode measurement and structural bias. The field should therefore prioritise prospective multicentre validation, physiology-aware multimodal modelling, decision-impact studies, uncertainty estimation, fairness auditing and implementation designs suited to high-burden settings.