Missingness Is a Workflow Outcome: An Equity-Aware Framework for Clinical Machine Learning

Clinical machine learning often represents missing values as technical defects to impute, encode, or exclude. In electronic health records, however, missingness is also produced by healthcare access, clinical indication, staffing, workload, documentation practices, interface design, organizational policy, and the timing of data exchange. These conditions are not distributed equally across patients, clinicians, sites, or shifts. A model that treats all missing values as equivalent may therefore learn local patterns of access and work while appearing statistically accurate. We propose the WORK-Missingness framework for equity-aware evaluation of clinical machine learning: Window of opportunity, Observation process, Record production, Known group distribution, and Model consequence. The framework asks investigators to define when a datum could have been produced, identify the people and systems involved in producing it, distinguish clinical absence from data unavailability, examine distributions across groups and settings, and test how missingness handling changes model performance and downstream eligibility. Published studies show that electronic record completeness differs by race, ethnicity, language, sex, socioeconomic status, site, and clinical workflow, with consequences for prediction and decision support. The framework is a position proposal, not a validated reporting standard. It offers a practical agenda for treating missingness as a sociotechnical and equity-relevant outcome rather than a neutral property of a dataset.