Poster Title: Social Determinant Predictors of Outcomes in Intracerebral Hemorrhage
Student: Tiffany Ong, Class of 2024
Faculty Mentor and Department: Aarti Sarwal, MD, Neurology
Funding Source: None
ABSTRACT
Background:
Intracerebral hemorrhage (ICH) is one of the most debilitating forms of stroke with significant morbidity and mortality, disproportionately affecting patients from rural areas more than their urban counterpart despite adjusting for comorbidities. These patients often experience delay in their care as they require inter-hospital transfers for specialized care. Using three machine learning prognostic models to predict 30-day mortality, modified Rankin scale upon discharge, and discharge disposition in ICH patients, we explored social determinants of health factors driving the disparities we see within our tertiary center.
Hypothesis:
Methods:
Preprocessing and model construction were performed on Python 3.8. Preprocessing functions and ML models were imported from the Python library scikit-learn. The dataset contained 138 samples with 13 useable features: 8 categorical features, 4 numeric features, and 1 ordinal feature. Missing values were imputed using the SimpleImputer library. All categorical variables were one-hot-encoded and ordinal variables were integer encoded. Three machine learning models were trained on the data provided to predict three different labels: 30 Day Mortality (Class 0: Alive, Class 1: Dead/Hospice), modified Rankin Score upon discharge (7 classes), and Discharge Disposition (Class 0: Home/Inpatient Rehabilitation Facility, Class 1: Hospice/Acute Care Facility Skilled Nursing Facility/Other Health Care Facility/Expired/Left Against Medical Advice/AMA/Long Term Care Hospital). Data was split 60/40 for training and testing sets respectively. Mean and standard deviation of F1 score, recall, and precision were calculated over 10 trials. Feature importance was determined using permutation feature importance.
Results:
138 ICH patients admitted in calendar in 2019 were used as initial sample. For all of these models the five most important features were the patient's GCS on admission, the patient's ICH Score, the patient's age, the patient's smoking status, and the patient's maximum distance from our hospital. The Multinomial Naive Bayes model showed F1 score of 0.8129 for 30-day mortality. Random forest performed lower but better than chance for mRS at discharge, F1 score 0.68 and discharge disposition 0.34.
Conclusions:
Distance travelled by an ICH patient from our tertiary medical center with shown to be indeterminant in patients’ outcome in our predictive model. Factors including rural urban status defined by rural-urban commuting area (RUCA) codes, time traveled to specialized care from point of origin, and payer status are currently being evaluated in this sample to create a prognostic model accounting for social determinants affecting ICH outcomes.
Source of mentor’s funding or other support that funded this research: None
Powered by Acadiate
© 2011-2024, Acadiate Inc. or its affiliates · Privacy