Abstract WP011: Monovariate Prediction of 90-Day Mortality in Acute Ischemic Stroke Using NIHSS: A Machine Learning, Ecological Analysis of 147,391 Patients

Background: Since its inception, the National Institutes of Health Stroke Scale (NIHSS) has defined severity for acute ischemic stroke (AIS). The simple risk calculator for arteriovenous malformation rupture (risk=105-age) has clear utility in patient planning and communication. However, only complex models have been developed in attempts to estimate the mortality risk in AIS. Methods: Patients from the TriNetX Research Network 18 years and older within the last 20 years were included based on the ICD-10 code I63, indicating AIS. Patients were excluded if they were missing NIHSS on the day of AIS. Propensity score matching (1:1) controlled for age, sex, race, ethnicity, and 93 comorbidities of the Charlson Comorbidity Index. The 90-day mortality of patients with each NIHSS level on admission was compared to patients with NIHSS=0 on admission after full matching. A linear model (probability of 90-day mortality) was trained where NIHSS was an odd number and tested where NIHSS was even. Results: Among 147,391AIS patients, the median admission NIHSS was 4. After matching, 90-day mortality rates were 10.1% in NIHSS=8, 20.8% in NIHSS=15, 30.3% in NIHSS=20, 40.4% in NIHSS=25, and 50.1% in NIHSS=28. The model was trained on 71,239 patients and showed that expected 90-day mortality was estimable by the equation: 1.7% x NIHSS (R 2 of 0.993). The model was validated on 81,242 patients, and true values for expected mortality were within 3% of predicted values. Overall observed 90-day mortality (12.6%) differed insignificantly (p=0.22) from predicted 90-day mortality was 13.3%. The observed-vs-predicted plot had the linear trendline predicted = 0.9678 * observed + 0.0237 (R 2 of 0.986). On residual analysis, predicted values were 0.0195 percentage points higher than observed values. Receiver-Operator Characteristic curve showed sensitivity of 93.5%, specificity of 99.1%, positive predictive value of 88.7%, negative predictive value of 99.1%, and area under the curve of 0.957. Conclusions: After controlling for demographics and comorbidities, the risk of 90-day mortality associated with AIS can be estimated by the equation 1.7% x NIHSS. The severity of the stroke explained the observed differences in mortality by NIHSS level to a high degree, as indicated by the R 2 of 0.993. Though subject to the ecological fallacy, this simple prediction model may have significant implications for communicating expectations with healthcare providers, patients, and families.