Development and validation of a prediction algorithm to identify birth in countries with high tuberculosis incidence...

Though targeted testing for latent tuberculosis infection ("LTBI") for persons born in countries with high tuberculosis incidence ("HTBIC") is recommended in health care settings, this information is not routinely recorded in the electronic health record ("EHR"). We develop and validate a prediction model for birth in a HTBIC using EHR data.

In a cohort of patients within Kaiser Permanente Southern California ("KPSC") and Kaiser Permanent Northern California ("KPNC") between January 1, 2008 and December 31, 2019, KPSC was used as the development dataset and KPNC was used for external validation using logistic regression. Model performance was evaluated using area under the receiver operator curve ("AUCROC") and area under the precision and recall curve ("AUPRC"). We explored various cut-points to improve screening for LTBI.

KPSC had 73% and KPNC had 54% of patients missing country-of-birth information in the EHR, leaving 2,036,400 and 2,880,570 patients with EHR-documented country-of-birth at KPSC and KPNC, respectively. The final model had an AUCROC of 0.85 and 0.87 on internal and external validation datasets, respectively. It had an AUPRC of 0.69 and 0.64 (compared to a baseline HTBIC-birth prevalence of 0.24 at KPSC and 0.19 at KPNC) on internal and external validation datasets, respectively. The cut-points explored resulted in a number needed to screen from 7.1-8.5 persons/positive LTBI diagnosis, compared to 4.2 and 16.8 persons/positive LTBI diagnosis from EHR-documented birth in a HTBIC and current screening criteria, respectively.

Using logistic regression with EHR data, we developed a simple yet useful model to predict birth in a HTBIC which decreased the number needed to screen compared to current LTBI screening criteria.

Our model improves the ability to screen for LTBI in health care settings based on birth in a HTBIC.