P005
Enhancing tuberculosis diagnosis with a mathematical model integrating molecular test histories and patient-specific factors
M O Boldi(2) N Abdulghafor(2) W Lugon(2) R Brouillet(1) C von Garnier(3) J Mazza-Stalder(3) G Greub(1,4) O Opota(1)
1:Institute of Microbiology, Lausanne University and University Hospital of Lausanne, Lausanne, Switzerland; 2:Faculty of Business and Economics, University of Lausanne, Lausanne, Switzerland; 3:Division of Pulmonology, Department of Medicine, Lausanne University Hospital, University of Lausanne, Lausanne, Switzerland; 4:Infectious Diseases Service, Lausanne University and University Hospital of Lausanne, Lausanne, Switzerland
Background
TB diagnosis still depends on a multi-test strategy that includes microscopy, PCR, and culture from successive clinical specimens. This approach faces challenges such as high costs and complex data interpretation, and there is a risk of over-relying on these results at the expense of clinical context and patient history. In this study, we aim to develop a model to predict TB, specifically designed to enhance the performance of multiple testing strategies.
Methods
We analyzed a database of 4,179 patients and 9,405 records from 2008 to 2018, assessing the diagnostic performance of individual TB tests across variables like test type, specimen type, and patient demographics. We applied a Hidden Markov Model (HMM) to generate scores indicating a patient's likelihood of TB based on sequential test results. Finally, we evaluated the performance of this predictive model against the historical data.
Results
We calculated TB prevalence, tests sensitivity and specificity across diverse categories, including gender, age, test types (microscopy, GeneXpert, in-house PCR, and culture), and specimen types (sputum, induced sputum, bronchoalveolar lavage and bronchial aspirate). We used PCR test sequences from our cohort to feed into an HMM, which generated a TB likelihood score from 0 to 1. Comparing to a gold standard of culture and clinical data, the HMM achieved 95.7% sensitivity and 97.9% specificity at an optimal score of 0.995.
Conclusions
This work illustrates that a mathematical model can enhance TB diagnosis by integrating molecular test histories with patient-specific factors, improving sensitivity, specificity, and supporting clinical decision-making across various testing stages.