Beyond ICD-10 Codes: A Multi-class Machine-Learning Approach to Identify True IBD Cases in Electronic Medical Records.

PURPOSE: The adoption of electronic medical records (EMRs) has expanded research opportunities. International Classification of Diseases (ICD) codes are used for case identification but lack sensitivity and specificity. Machine learning (ML) offers an alternative by integrating EMR data to identify cases. This study aimed to develop a multi-class ML model to accurately predict inflammatory bowel disease (IBD) among patients with ICD-10 codes for Crohn's disease (CD) or ulcerative colitis (UC).

METHODS

Patients with ≥1 ICD-10 code for CD or UC were identified from the EMR; a random cohort underwent manual validation. A total of 198 features were selected. Logistic regression (LR), random forest (RF), and XGBoost (XGB) models were trained and tested on the validated cohort, with and without sevenfold recursive feature elimination (RFECV). Model interpretability was assessed using SHapley Additive exPlanations. Models were then applied to the remaining ICD-10-based cohort.

RESULTS

Among 34,884 patients with ICD-10 codes, 1200 were validated manually (33% CD, 32% UC, 35% No IBD). The positive predictive value (PPV) of a single ICD-10 code was 65%. RFECV reduced features to 41 (XGB), 30 (LR), and 55 (RF). The XGB RFECV model achieved the best performance, with a PPV of 91.33%, specificity over 90% across all classes (up to 96.39%), and sensitivities ranging from 82.81 to 91.74%. Performance remained similar across models after RFECV.

CONCLUSION

Our multi-class ML approach improved IBD classification compared to ICD-10 codes, increasing the PPV from 65 to over 90%. This framework provides a scalable, accurate, and interpretable method for EMR-based research.