PT-019 - APPLYING ARTIFICIAL INTELLIGENCE-BASED DIGITAL TWINS IN TUBERCULOSIS CLINICAL TRIALS

Z. Li¹, R. Savic^2,3, M. Morimoto², D. Purohit⁴, C. Silva⁵; ¹UCSF, San Francisco, CA, USA, ²University of California, San Francisco, San Francisco, CA, USA, ³UCSF Center for Tuberculosis, University of California, San Francisco, San Francisco, CA, USA, ⁴University of California, San Francsico, San Francisco, CA, USA, ⁵University of California San Francisco, San Francisco, CA, USA.

Background: In drug-susceptible tuberculosis clinical trials, all new regimens need to compare their effects to the 6-month standard-of-care treatment with rifampicin, isoniazid, pyrazinamide, and ethambutol. This study aims to develop artificial intelligence-based digital twins to replace patients in the control group with standard-of-care treatment.

Methods: A total of 6905 patients from 6 clinical trials were collected, in which patients from 5 clinical trials were used to develop the model and patients from study S31 clinical trial were used as external validation. All patients in the control group from 5 clinical trials (1784 patients) were combined and randomly split into a training dataset (70%) and a testing dataset (30%). 7 machine learning models were developed to predict the favorable outcome of patents and the model with the best prediction performance in the testing dataset was selected and updated with the whole dataset to get the final model. Important features from the final machine learning model were applied to match patient baseline demographics between the case and control groups in the study 31 trial. All matched patients in the control group were replaced by digital twins, and their outcomes were predicted by the final machine learning model. Non-inferiority analysis between the case and control group was then conducted to test the statistical power.

Results: The artificial neural network model showed the best performance of favorable outcome prediction in the testing datasets. After updating the whole dataset, the final artificial neural network model showed good prediction performance (AUROC of 0.785) in the external validation dataset. Sex, smear grade, cavitary status, HIV infection status, race, body mass index, and age were identified as important predictors and used to match patients between the case and control groups. After replacing matched patients in the control group with their digital twins, the non-inferiority analysis still showed an inferior effect on the case group compared to the control group.

Conclusion: Artificial intelligence-based digital twins can be established to replace the physical patients in the control group in drug-susceptible tuberculosis clinical trials and maintain statistical power.