P080
Predictive AI Models for Personalized Prognosis of Tuberculosis Treatment
A Valafar(2) S Valafar(1)
1:Chicago Medical School; 2:University of California, Riverside
Evolutionary trajectory of a pathogen partially depends on the immune response and drug pressure. This is also true for M. tuberculosis. The ability to predict this evolutionary trajectory, enables prognosis of the disease and potential course-change to avoid emergence of resistance. In this abstract we report a model that predicts the emergence of canonical isoniazid (INH) mechanisms of resistance (katG315, inhA-15, and inhA-8).
To develop such a predictive model, we extensively experimented with Logistic Regression (LR) and Deep Neural Models (DNMs) (multi-layered deep neural networks). All models were trained to takes an exhaustive combination of specific genotypes of the M. tuberculosis genomes (collected prior to the onset of resistance during the course of the treatment) as their input, and provide a prediction of the likelihood of emergence of one of the listed three canonical mutations. Data for training and testing of these models were downloaded from the CRyPTIC consortium and the TB Portals.
Our best performing model was a DNM with an estimated prognostic accuracy of 73% with a sensitivity of 55%, specificity of 84%, a PPV of 69%, and an NPV of 75% for correctly predicting the emergence of the three canonical INH resistance mutations. These results are high enough to demonstrate the potential of DNMs as a prognostic tool at least for TB. We expect that larger genotypic/phenotypic data sets allows the training of such models to achieve higher sensitivity, specificity, and prognostic accuracy.