Identification of Cohorts with Inflammatory Bowel Disease Amidst Fragmented Clinical Databases via Machine Learning.

Area of Science:

• Gastroenterology and Hepatology
• Health Informatics
• Clinical Epidemiology

Background:

• Accurate Inflammatory Bowel Disease (IBD) cohort identification is crucial for patient care and research.
• Traditional methods relying on billing codes (e.g., ICD-10) are often incomplete due to database fragmentation and missing data.
• Specialist centers, like those with intestinal failure units, manage complex IBD cases that may be missed by standard coding.

Purpose of the Study:

• To develop and validate novel cohort retrieval methods for identifying the total Inflammatory Bowel Disease (IBD) patient population within a large university teaching hospital.
• To compare the efficacy of multiple clinical databases in capturing IBD patients.
• To quantify the underestimation of IBD cohorts by current standard methods.

Main Methods:

• Utilized 11 distinct clinical databases (including billing codes, registries, patient portals, prescriptions, and clinical notes) from 2007-2023.
• Employed a penalized logistic regression (LR) classifier, statistically comparing and validating its performance across databases.
• Integrated predictions from the LR model with manual review of IBD clinic letters for final cohort estimation.

Main Results:

• A validation cohort of 2800 patients confirmed the limitations of individual databases.
• The penalized LR model identified 8,159 IBD patients with an AUROC of 0.85.
• Combining the LR model with clinic letters yielded a total estimate of 13,048 IBD patients, significantly exceeding the 8,048 identified by ICD-10 codes and medication data alone (a 38.3% underestimation).

Conclusions:

• Diagnostic billing codes and medication data are insufficient for comprehensive IBD cohort identification in secondary care.
• A multimodal, cross-database approach significantly improves the capture of IBD patients.
• Future improvements necessitate advanced methods, such as natural language processing (NLP), for more robust patient identification.