Hypoglycemia Score Helps Pharmacists Rule Out Low-Risk Inpatients

Predictive Model Development

Authors:Liren Li; Jiayu Chen; Tongshu Guan, et al.

Journal:Diabetes Research and Clinical Practice

DOI:10.1016/j.diabres.2025.112993

A 10-variable CatBoost model turns routine EMR data into a dynamic hypoglycemia risk radar for hospitalized patients with diabetes, using high negative predictive value to safely rule out many low-risk inpatients so pharmacists can focus scarce monitoring and medication-review resources on patients most likely to develop hypoglycemia, pending local validation.

November 10, 2025

Quick Take

CatBoost predicted inpatient hypoglycemia with area under the curve (AUC) 0.85 and negative predictive value (NPV) 0.89 (positive predictive value (PPV) 0.75) in a held-out test set of 2,845 hospitalized patients (418 events, 14.7%).

Pharmacy impact: a 10-variable electronic medical record (EMR) score (insulin use, HbA1c, creatinine, BMI class, urea, triglycerides, albumin, ALT, indirect bilirubin, urine protein) could help pharmacists rule out low-risk patients (high NPV) and prioritize monitoring and medication review for flagged higher-risk inpatients.

Why it Matters

Inpatient hypoglycemia is a frequent, high-risk adverse event; most existing tools rely on static pre-admission data and are not optimized for time-sensitive, in-hospital risk changes.

Pharmacy teams must triage limited staff and monitoring resources; insulin exposure and fluctuating renal or nutritional labs materially change short-term risk and make manual review resource intensive and error prone.

A dynamically updated, EMR-based risk score could concentrate point-of-care glucose checks and pharmacist-led medication review where they are most needed, but deployment requires clinical decision support (CDS) integration, automated data pipelines, and planning for alert burden and operational constraints.

What They Did

Retrospective, single-center EMR study at Nanfang Hospital (November 2021–December 2022); from an initial 37,966 inpatients with diabetes, predefined exclusions produced a final cohort of 2,845 adults (excluded: pre-admission hypoglycemia, pregnancy/death, ≥30% missing data).

Candidate predictors were routine EMR items: demographics, medication orders, comorbidities, and laboratory results; features with >50% missing or extreme imbalance were dropped and remaining gaps imputed using a random-forest method.

Feature selection used a stepwise forward approach guided by XGBoost to identify a 10-variable panel; models were developed on an 80/20 train/test split with 10-fold cross-validation on the training set.

Ten machine learning algorithms were compared (including SVM, XGBoost, random forest, LGBM, transformer); model choice prioritized AUC and calibration, and temporal leakage was mitigated by using only pre-event data for cases.

What They Found

CatBoost outperformed other algorithms with AUC = 0.85, PPV = 0.75 and NPV = 0.89 on the held-out test set (n = 2,845; 418 hypoglycemia events, 14.7%) and had the best calibration (Brier = 0.088).

A 10-variable panel (BMI class; insulin use; HbA1c; creatinine; urea; triglycerides; albumin; alanine aminotransferase (ALT); indirect bilirubin (IDIL); urine protein (PRO)) reached peak stepwise-selection AUC = 0.81 and formed the input for the final CatBoost model that produced the reported metrics.

SHAP (SHapley Additive exPlanations) attribution identified insulin use as the single strongest predictor; lower BMI, higher HbA1c, and renal dysfunction markers (creatinine, urea, proteinuria) also meaningfully increased predicted risk.

Pharmacy-relevant conclusion: the model’s high NPV and favorable decision-curve net benefit suggest it could reliably rule out many low-risk inpatients and help prioritize pharmacist medication review and targeted glucose monitoring for patients the model flags as higher risk. The incremental signal was driven chiefly by insulin exposure and dynamic renal/nutritional laboratory markers.

Takeaways

For pharmacists, the model functions as a dynamic risk radar for inpatients with diabetes, combining insulin exposure with a small set of routinely collected labs to estimate in-hospital hypoglycemia risk.

Because the model’s negative predictive value is high (0.89 in this cohort), its most practical use is ruling out low-risk patients so teams can concentrate monitoring, dose adjustments, and nutrition review on the smaller, model-flagged higher-risk group.

Operational caution: the model mirrors clinical heuristics (insulin and renal/nutritional signals) and, given the single-center retrospective development, should complement—not replace—pharmacist judgment and be interpreted in context of local practice patterns and bedside assessment.

Strengths and Limitations

Strengths:

Robust ML development workflow: large EMR-derived cohort after filtering, explicit temporal alignment to avoid label leakage, systematic feature selection, 10-model comparison, and 10-fold cross-validation on the training set.

Parsimonious, interpretable final model: CatBoost uses only 10 routinely collected labs/medication variables, demonstrated good calibration (Brier = 0.088) and SHAP-based interpretability that aligns with clinical risk factors.

Limitations:

Single-center, retrospective design without multicenter external validation limits transportability; differences in hypoglycemia incidence, insulin practices, and monitoring protocols will likely require local recalibration.

Detection bias risk: glucose testing frequency was not standardized—patients perceived as higher risk (for example, those on insulin) are tested more often, which can inflate measured performance.

Selection bias from missing data: 512 eligible patients were excluded for ≥30% missingness, so the model was trained and tested on a relatively data-complete cohort that may not reflect real-world EMR messiness.

Temporal-mismatch/immortal-time risk: for cases predictors were drawn from the pre-event window but for controls from the entire hospitalization, which can introduce bias if longer observation windows correlate with lower apparent risk.; Confounding by indication: insulin use is the dominant predictor, meaning the model partly identifies patients receiving high-risk therapy rather than independent, unmodifiable patient susceptibility.

Operational constraints: the model recalculates risk only when required labs are updated; it does not produce continuous hourly risk trajectories and would require automated data pipelines and CDS integration for real-time use.

Bottom Line

Promising internally validated, parsimonious tool to rule out low-risk inpatients and help prioritize pharmacist review, but it needs multicenter external validation, local recalibration, and careful operational planning (detection-bias mitigation, CDS design, and alert-thresholding) before clinical deployment.