Can Artificial Intelligence based Automated CT Brain Interpretation Software help Early Clinical Decision Making for Stroke Patients in Real World Resource Limited Settings with Non-Specialist Physicians? An Interrupted Time Series Study from Tezpur, Assam, India

We evaluated the impact of an AI (Artificial Intelligence) based software for automated interpretation of CT brain through a retrospective interrupted time series study at a rural hospital in Tezpur, India, that is managed by non-specialist physicians and supported by teleradiology services. We compared the diagnostic accuracy of the software to detect an abnormality against the teleradiologist report and evaluated the impact in stroke patients by comparing the time from CT imaging to significant intervention from before the deployment to the timestamps after the deployment. The specificity and negative predictive value were remarkably high for most of the findings, but the sensitivity and positive predictive value were low for subdural and subarachnoid hemorrhage (n=531). The median time to intervention was significantly lower at 59 minutes in the post deployment phase (IQR: 30.5, 128) than 83 minutes before the deployment (IQR: 57, 144) for acute stroke patients (n=176). Our study showed that it is feasible and impactful to deploy AI based software in resource limited hospitals to help the physicians to make early decisions on life saving interventions in critically ill patients.