Developing and Evaluating a Natural Language Processing Algorithm to Detect Insulin Pump and Continuous Glucose Monitor Use in Electronic Health Records of Patients with Type 1 Diabetes
Authors' conclusion
We showed that NLP can accurately identify insulin pump and CGM users from unstructured EHR notes, substantially outperforming billing code-based methods. This scalable approach can support system- and population-level evaluations of diabetes technologies.
Study context
- Population
- Patients with T1D; outpatient endocrinologist clinical notes, University of California Los Angeles (UCLA) holdout dataset (most recent note per patient)UCLA holdout dataset
- Participants
- 667 patients
- Comparison
- Natural language processing (NLP) algorithms vs billing codes for identifying insulin pump and continuous glucose monitor (CGM) use
- Outcome
- Sensitivity and specificity for identifying insulin pump and CGM users vs manual chart review
Methodological detailsrule-based NLP algorithms, holdout dataset, external validation in a second health system, manual chart review as gold standard
- Rule-based NLP algorithms
- Holdout dataset
- External validation in a second health system
- Manual chart review as gold standard
Performance
As reported in the abstract
Diagnostic accuracy vs manual chart review (UCLA holdout dataset)
| Test or model | Sensitivity | Sensitivity, insulin pump (ICD/CPT) | Sensitivity, CGM (CPT) | Sensitivity, combined device use | Specificity | Specificity, insulin pump (ICD/CPT) | Specificity, CGM (CPT) | Specificity, combined device use |
|---|---|---|---|---|---|---|---|---|
| Insulin pump algorithm | 0.90 | — not stated in the abstract | — not stated in the abstract | — not stated in the abstract | 0.89 | — not stated in the abstract | — not stated in the abstract | — not stated in the abstract |
| CGM algorithm | 0.85 | — not stated in the abstract | — not stated in the abstract | — not stated in the abstract | 0.84 | — not stated in the abstract | — not stated in the abstract | — not stated in the abstract |
| Combined algorithm | 0.76 | — not stated in the abstract | — not stated in the abstract | — not stated in the abstract | 0.92 | — not stated in the abstract | — not stated in the abstract | — not stated in the abstract |
| Billing codes | — not stated in the abstract | 0.09 | 0.68 | 0.06 | — not stated in the abstract | 1.00 | 0.86 | 1.00 |
Sensitivity
Insulin pump algorithm0.90
CGM algorithm0.85
Combined algorithm0.76
Billing codes— not stated in the abstract
Sensitivity, insulin pump (ICD/CPT)
Insulin pump algorithm— not stated in the abstract
CGM algorithm— not stated in the abstract
Combined algorithm— not stated in the abstract
Billing codes0.09
Sensitivity, CGM (CPT)
Insulin pump algorithm— not stated in the abstract
CGM algorithm— not stated in the abstract
Combined algorithm— not stated in the abstract
Billing codes0.68
Sensitivity, combined device use
Insulin pump algorithm— not stated in the abstract
CGM algorithm— not stated in the abstract
Combined algorithm— not stated in the abstract
Billing codes0.06
Specificity
Insulin pump algorithm0.89
CGM algorithm0.84
Combined algorithm0.92
Billing codes— not stated in the abstract
Specificity, insulin pump (ICD/CPT)
Insulin pump algorithm— not stated in the abstract
CGM algorithm— not stated in the abstract
Combined algorithm— not stated in the abstract
Billing codes1.00
Specificity, CGM (CPT)
Insulin pump algorithm— not stated in the abstract
CGM algorithm— not stated in the abstract
Combined algorithm— not stated in the abstract
Billing codes0.86
Specificity, combined device use
Insulin pump algorithm— not stated in the abstract
CGM algorithm— not stated in the abstract
Combined algorithm— not stated in the abstract
Billing codes1.00
— not stated in the abstract
Further findings
External validation in a second health system
- Similarly strong algorithm performance in the second health system
Original publication
Developing and Evaluating a Natural Language Processing Algorithm to Detect Insulin Pump and Continuous Glucose Monitor Use in Electronic Health Records of Patients with Type 1 Diabetes. Diabetes Technol Ther. Published 29 September 2025. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →