PLoS ONE (Jan 2021)

Radiologist observations of computed tomography (CT) images predict treatment outcome in TB Portals, a real-world database of tuberculosis (TB) cases.

  • Gabriel Rosenfeld,
  • Andrei Gabrielian,
  • Qinlu Wang,
  • Jingwen Gu,
  • Darrell E Hurt,
  • Alyssa Long,
  • Alex Rosenthal

DOI
https://doi.org/10.1371/journal.pone.0247906
Journal volume & issue
Vol. 16, no. 3
p. e0247906

Abstract

Read online

The TB Portals program provides a publicly accessible repository of TB case data containing multi-modal information such as case clinical characteristics, pathogen genomics, and radiomics. The real-world resource contains over 3400 TB cases, primarily drug resistant cases, and CT images with radiologist annotations are available for many of these cases. The breadth of data collected offers a patient-centric view into the etiology of the disease including the temporal context of the available imaging information. Here, we analyze a cohort of new TB cases with available radiologist observations of CTs taken around the time of initial registration of the case into the database and with available follow up to treatment outcome of cured or died. Follow up ranged from 5 weeks to a little over 2 years consistent with the longest treatment regimens for drug resistant TB and cases were registered within the years 2008 to 2019. The radiologist observations were incorporated into machine learning pipelines to test various class balancing strategies on the performance of predictive models. The modeling results support that the radiologist observations are predictive of treatment outcome. Moreover, inferential statistical analysis identifies markers of TB disease spread as having an association with poor treatment outcome including presence of radiologist observations in both lungs, swollen lymph nodes, multiple cavities, and large cavities. While the initial results are promising, further data collection is needed to incorporate methods to mitigate potential confounding such as including additional model covariates or matching cohorts on covariates of interest (e.g. demographics, BMI, comorbidity, TB subtype, etc.). Nonetheless, the preliminary results highlight the utility of the resource for hypothesis generation and exploration of potential biomarkers of TB disease severity and support these additional data collection efforts.