Exploring the reliability of inpatient EMR algorithms for diabetes identification

Area of Science:

• Health Informatics
• Clinical Data Science
• Diabetes Research

Background:

• Accurate, real-time identification of medical conditions in inpatients is vital for health systems.
• Current comorbidity algorithms face delays due to administrative data integration.
• Electronic Medical Record (EMR) data offers a promising alternative for timely condition phenotyping.

Purpose of the Study:

• To develop and evaluate EMR data-based algorithms for inpatient diabetes phenotyping.
• To compare the performance of different EMR data sources for diabetes identification.
• To assess the utility of free-text data in improving diabetes case identification.

Main Methods:

• A chart review of 3040 individuals, with 583 having diabetes, served as the gold standard.
• EMR data, including laboratory results, medications, and free-text notes, were used to develop five distinct diabetes algorithms.
• Algorithm performance was evaluated based on sensitivity (SN) and positive predictive value (PPV).

Main Results:

• The 'all document types' algorithm achieved high performance: SN 0.95, specificity (SP) 0.98, PPV 0.94, and negative predictive value (NPV) 0.99.
• The medication and laboratory algorithm demonstrated strong performance with SN 0.90, SP 0.95, PPV 0.80, and NPV 0.97.
• The ICD-coded data algorithm showed SN 0.84, SP 0.98, PPV 0.93, and NPV 0.96.

Conclusions:

• EMR-based diabetes algorithms, particularly those utilizing free-text data, can achieve performance comparable or superior to traditional ICD-coded methods.
• These algorithms can supplement existing methods, offering a more timely approach to inpatient case identification.
• Inpatient EMR-based algorithms are crucial for efficient resource planning and timely care delivery.