JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) systems in health care increasingly rely on electronic health record (EHR) data for predictive modeling. However, EHR data often fail to capture real-world nursing workflows and clinical reasoning. In inpatient settings, gaps in glucose management documentation may reflect either care delivered but not documented (missing data) or care not delivered (missed care), yet these scenarios appear indistinguishable in EHR data. Without understanding how care is delivered, AI models may misclassify missing documentation as missed care, reducing accuracy and creating safety risks.

Objective:

To explore how inpatient nurses manage and document insulin-related care, informing clinical knowledge graphs of nursing decision-making.

Methods:

We conducted a qualitative study using semi-structured interviews and focus groups with inpatient registered nurses across the United States. Data were analyzed using a hybrid deductive–inductive approach informed by the Theory of Missed Nursing Care, with GPT-assisted thematic synthesis reviewed by nurse experts.

Results:

Nineteen nurses participated across 10 sessions. Six themes explained why EHR gaps arise and are misinterpreted: (1) technology friction and interoperability gaps; (2) competing demands forcing documentation workarounds; (3) clinical decisions adjusting to mealtime variability; (4) informal, undocumented communication; (5) under documented but necessary protocol deviations; and (6) narrative documentation supplementing inadequate structured fields.

Conclusions:

Documentation gaps often reflect adaptive, context-dependent nursing work rather than missed care. Structured EHR fields fail to capture workarounds and informal reasoning, meaning AI models built without domain expertise risk conflating undocumented care with absent care, threatening both accuracy and patient safety. Inpatient insulin management is context-dependent and cannot be reliably interpreted without adequate domain expertise. Distinguishing missing data from missed care requires representing these dependencies explicitly, offering a practical foundation for safer, more interpretable AI models built on nurse-generated EHR data.