Research Mentor(s): Sriram Chandrasekaran
Authors: Noah Black, Kirk Smith, Sriram Chandrasekaran, PhD
Until COVID-19, tuberculosis was the single largest cause of infectious death globally. The disease consistently registers more than ten million diagnoses annually and exceeds over one million deaths per year. Tuberculosis is primarily a lung disease caused by the Mycobacterium tuberculosis bacterium. Treating tuberculosis is complex: it requires patients to consistently take a slew of antibiotics over the course of many months. Tuberculosis treatment, diagnosis, and mortality has suffered a regression in recent years due to healthcare complexities caused by the COVID-19 pandemic and the increase of multi-drug resistant and extensively-drug resistant tuberculosis. Often, chest radiographs are used to diagnose and track the progression of the disease for patients. Additionally, a patient’s socio-demographic information is extremely relevant to tuberculosis’ trajectory. Poverty, malnourishment, and other social factors are the main drivers of tuberculosis and as such, they are extremely relevant in tuberculosis prognosis. Imaging and socio-demographic data are reported to and published by the National Institutes of Health through the TB Portals Program. We seek to combine these data modalities through deep and traditional machine learning techniques in order to prognose patient outcomes. We hope that clinicians are able to use our model to quickly assess the most vulnerable patients so that they can direct more resources towards those who need it. Our algorithm culminates in a simple and interpretable classification prediction notifying physicians whether their patient is expected to have a successful or unsuccessful treatment. First, we investigate chest radiograph segmentation with a custom UNET architecture called Minorly Residual UNET. We proceed to use established convolutional neural networks and transfer learning techniques to extract chest radiograph features. Finally, we concatenate this information with optimized clinical data before passing it into a final classification algorithm. This research represents the first attempt of automatic, multi-modal prognosis prediction for tuberculosis.