– Intracranial hemorrhage (ICH) is a frequent and critical clinical problem.

– Nearly half of the mortality occurs within the first 24 hours and earlier intervention leads to improved outcomes

– CT of the head is the most widely used method to diagnose ICH.

– The diagnosis is dependent on how quickly a CT scan is performed and interpreted by a physician.

– We developed a machine learning model and associated “back-end” software to automatically classify CTs as positive or negative for ICH. The model also further categorizes positives studies as subdural, epidural, intra-parenchymal, intra-ventricular, and/or subarachnoid hemorrhages.

– Radiologists are alerted of positive cases through Spok Mobile alarms and e-mail.

– An explainability module was developed to explain machine learning model inference.

Intracranial hemorrhage (ICH) is a frequent and critical clinical problem with significant morbidity and mortality. Nearly half of the mortality occurs within the first 24 hours and earlier intervention has been shown to improve patient outcomes. Computed tomography (CT) of the head is the most widely used method to diagnose ICH. Interpretation of head CTs is key to triage and management of patients with ICH. Ultimately, the diagnosis is dependent on how quickly a CT scan is performed and interpreted by a physician. This approach helps ensure that head CTs with ICH are promptly interpreted rather than waiting for a radiologist to sequentially review imaging studies in a queue. The goal of implementing such a model is to reduce time to diagnosis, facilitate early intervention, optimize patient outcomes, and improve patient flow through the hospital. Our intent is to reduce the time to CT interpretation from hours to minutes.

We have developed a machine learning model to analyze head CTs for ICH and triaging radiologist worklists accordingly. The model was validated on a dataset consisting of every unenhanced head CT scan (n = 5965) performed in our emergency department in 2019 without exclusion. The model demonstrated an AUC of 95.4%, sensitivity of 91.3%, and specificity of 94.1%. Radiologists are alerted to positive CT scans by Spok Mobile messages and email. In addition, an explainability module was developed to explain machine learning model inference to radiologists.

Salehinejad H, Kitamura J, Ditkofsky N, Lin A, Bharatha A, Suthiphosuwan S, Lin HM, Wilson JR, Mamdani M, Colak E. A real-world demonstration of machine learning generalizability in the detection of intracranial hemorrhage on head computerized tomography. Scientific Reports. 2021 Aug 23;11(1):17051.