Predicting Glucose Test Ordering in Hospitalized Patients Using Temporal Models of Clinical Context Embeddings

The overuse of laboratory tests is a persistent challenge in healthcare, driving unnecessary costs, patient discomfort, and low-value care. Glucose testing, one of the most common diagnostics, exemplifies this issue in hospital settings. We present a deep learning framework that integrates structured and unstructured electronic medical record data to predict whether a glucose test will be ordered in the next AM/PM time bin. Using multi-hospital data from the GEMINI dataset, we combine Long Short-Term Memory models with Clinical BioBERT embeddings to capture both the timing and clinical context of testing. On held-out test data, our best model achieved ROC-AUC of 0.92 and PR-AUC of 0.67, and generalized across sites in leave-one-hospital-out evaluation (ROC-AUC 0.84). Embedding-based models outperformed traditional feature representations, though adding more tests and vitals did not always yield further gains. By contrast, introducing a simple temporal recency cue (bin counter) improved performance. An exploratory regression task for predicting glucose values performed worse, likely due to class imbalance and reliance on forward-filled values; Random Forest achieved R^2 of 0.80 under masked evaluation, indicating a need for more frequent or diverse test data. Predicting laboratory test ordering is the first step toward evaluating the usefulness of laboratory test use and establishes a foundation for future real-time decision support to reduce unnecessary lab use in hospitals.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.