Enhancing learning health care systems with real-world data harmonization frameworks

Abstract

The use of digital tools has enabled the capture of clinical data that can be leveraged to produce real-world evidence, complementing traditional trial data and addressing evidence gaps in oncology. Electronic health records offer rich, longitudinal data, although recording practices, collection methods, and quality may vary over time and across settings, which can lead to inconsistencies. Data harmonization is key to overcoming fragmentation, and common data models offer structured methods for unifying diverse health care data. Various mapping strategies standardize data balancing, automation, flexibility, and semantic precision: ‘upstream semantic mapping' emphasizes standardization, ‘downstream semantic mapping' allows tailored adjustments by local experts, and ‘direct dataset mapping' focuses on creating analysis-ready datasets without intermediate data transformation. The best approach depends on organizational needs, data characteristics, and intended use cases, but harmonization must be a continuous process that adapts to evolving data sources, variable data types, new collection methods, and shifting transformation practices. Clinical expertise must ensure that cohorts mirror real-world scenarios and capture local nuances, more often with an artificial intelligence–assisted framework, while epidemiologists must consider harmonization practices in the study design and analysis. Data strategies must protect patient privacy while supporting health care improvements, with secure data handling paired with accessible sharing. By engaging medical professionals at every stage, health care institutions can become learning health systems driven by high-quality observational research, ultimately advancing knowledge and improving patient outcomes.