Data-Centric Approaches in Precision Health

Abstract

Precision health aims to deliver the right intervention to the right patient at the right time by leveraging individual-level molecular, clinical, and behavioural data. While model-centric AI research focuses on developing better algorithms, the data-centric paradigm recognises that data quality, completeness, representativeness, and labelling accuracy are the primary determinants of health AI performance. Electronic health records (EHRs) contain structured and unstructured data on hundreds of millions of patients but suffer from missingness (30-70% of clinical variables missing per encounter), label noise (ICD codes have 15-25% error rates), population bias (underrepresentation of minorities, rural populations, and rare diseases), and temporal distribution shift (coding practices and treatment guidelines change over time). We present the Data-Centric Precision Health Framework (DCPHF), evaluating five data-centric strategies -- systematic data cleaning and imputation, active learning for efficient labelling, data augmentation and synthetic generation, federated learning for multi-institutional integration, and fairness-aware data curation -- across four precision health applications (clinical risk prediction, treatment response modelling, rare disease diagnosis, and health disparity reduction). Our Data-Centric Health Score (DCHS) measures data quality improvement, downstream model performance, label efficiency, population representativeness, and fairness impact. Federated learning achieves the highest DCHS (0.924) by enabling multi-institutional model training without centralising patient data, while fairness-aware curation achieves the highest equity impact (0.960) by systematically addressing representation gaps in training data.