{
  "abstract": "Introduction Nearly 800,000 people in the US have a stroke every year and it represents a leading cause of long-term disability. While there have been improvements in stroke surveillance, prevention, acute care, and post-acute rehabilitation, stroke burden has only increased over the past decades. Current stroke risk assessment tools exist but are limited and population-specific. With the growing movement towards personalized medicine, new data driven prediction models may provide more accurate patient-centered risk assessment. The Behavioral Risk Factor Surveillance System Survey (BRFSS) provides an attractive source of data as it includes self-reported health metrics on a broad number of variables. We aim to develop and validate a deep learning architecture for population-level stroke risk stratification, based on data from the BRFSS.Methods and Materials A binary classification model based on a deep neural network trained on select data from the 2022 Behavioral Risk Factor Surveillance System Survey was developed. Data was prepared in Python, and the model was built using the Keras API for TensorFlow. After reserving a test set to avoid data leaks, missing values were removed and data was scaled and encoded separately. Though the data set was significantly unbalanced, no attempts at rebalancing given model architecture and the goal of broad generalizability. The model was improved iteratively and hyperparameters tuned manually for best performance. The final model was a deep neural network with 16 total layers including dense, batch normalization, and dropout layers. Validation was achieved with K-fold cross-validated with a K=5. The model was designed to predict occurrence of stroke based on a number of risk factors including sex, age, smoking, prior cardiac events, BMI, social and demographic factors, and other health metrics. Presented metrics were calculated based on reserved training data. Results were compared to a baseline balanced random forest classifier with ADASYN oversampling of the minority class.Results The model was able to outperform a balanced random forest classifier (AUC = 0.679) and correctly predict stroke occurrence with an AUC of 0.818, making it competitive with similar models trained on other data sets and demonstrating potential clinical usefulness as a general screening tool.Conclusions The deep neural network-based model was able to effectively predict stroke status in patients based on Behavioral Risk Factor Surveillance System Survey data. These preliminary results show promise for this architecture and, with further refinement, may lead to insights into the importance of health and demographic factors, as well as provide a foundation for deployable models. This model, or one similar, could serve as a digital triaging tool to identify high-risk individuals in primary care settings using self-reported health metrics. Next steps include creating an ensemble model including a gradient boost component and a final decision tree to improve performance as well as performing a Shapley Additive Explanations analysis.Disclosures A. Amendolara: None. C. Reynoso: None. P. Colin: None.",
  "authors": [
    {
      "affiliations": [
        "Neurology, St. Luke’s University Health Network, Bethlehem, PA"
      ],
      "name": "A Amendolara"
    },
    {
      "affiliations": [
        "Biomedical Sciences, Noorda College of Osteopathic Medicine, Provo, UT"
      ],
      "name": "C Reynoso"
    },
    {
      "affiliations": [
        "Biomedical Sciences, Noorda College of Osteopathic Medicine, Provo, UT"
      ],
      "name": "P Colin"
    }
  ],
  "title": "E-167 A deep neural network-based classifier for predicting stroke occurrence",
  "uid": "53d54e89-1e96-56f1-a841-093c26b69062"
}