In gratitude of the loyal support of our members, the CNS is offering complimentary 2021 Annual Meeting registration to all members! Learn more.

  • Machine Learning for Predicting Delayed Onset Trauma Following Ischemic Stroke

    Final Number:
    1275

    Authors:
    Anthony K Ma BS, MS; Charlie Weige Zhao; Charles Christian Matouk BSc MD

    Study Design:
    Other

    Subject Category:

    Meeting: Congress of Neurological Surgeons 2018 Annual Meeting

    Introduction: Stroke is currently the fifth leading cause of death in the United States. Interestingly, only 18% of stroke patients die from the initial trauma. Most patient deaths following an acute stroke resulted from complications weeks to months after the initial event. These complications include a recurrent stroke, MI, pneumonia, pulmonary embolism, etc. In this study, we apply machine learning techniques to effectively predict the most probable life-threatening risk that follows the initial event to optimize outcomes in stroke recovery.

    Methods: Patient profile data was obtained from the International Stroke Trial (IST) database with 6-years of accumulated data and over 19,000 patient cases. We trained a neural network on over 4364 training examples with 48 features such as age, gender, blood pressure, and presence or absence of infarct to predict the most likely cause of death out of 8 most common possibilities. Our trained model was evaluated with a testing data set of 500.

    Results: Initial profiling of the 19,000 patients in the IST database revealed that only 18% of patients died from initial stroke within 3 months, 24% faced serious complications, and 58% had only minor to no complications. The most common serious complications included 20% pneumonia, 12% heart failure, 12% gastrointestinal bleeding, and 12% cardiac arrest. After training the neural network model, we achieved an overall risk stratification testing accuracy of 78.6% (std err = 0.3%).

    Conclusions: Only a small percentage of stroke victims die from the initial trauma. Many patients have a prolonged time window before facing secondary causes of mortality. We applied principles in artificial intelligence to predict the most likely complications based on a patient’s profile. Our approach of using a combinatorial approach from machine learning for stroke-related mortality prediction has yet to be described in literature and may guide future post-stroke management.

    Patient Care: Many patients have a time window after their initial stroke attack before facing a secondary complication that ultimately causes death. Applying principles of artificial intelligence to help predict risk at the individual patient level can help tailor a physician’s treatment plan to optimize survival outcomes.

    Learning Objectives: 1) Describe the epidemiology of stroke patients. 2) Characterize the most likely causes of complications following ischemic stroke. 3) Predict the most likely secondary cause of mortality for an individual patient.

    References:

We use cookies to improve the performance of our site, to analyze the traffic to our site, and to personalize your experience of the site. You can control cookies through your browser settings. Please find more information on the cookies used on our site. Privacy Policy