BACKGROUND: Clinical severity assessed by the National Institutes of Health Stroke Scale (NIHSS) score is not available in national hospital database. METHODS: Data from the French National Health Data System (SNDS) in the region of the Brest Stroke Registry (BSR), were used to calculate a probability of death among hospitalized ischemic strokes (IS) and intracerebral hemorrhages (ICH), relying on multivariable logistic regressions using available stroke/patient characteristics. Data from the BSR, have been used to derive equations putting in relation initial NIHSS and probability of death. These equations were used to estimate initial NIHSS based on probability of death. Data from the Dijon Stroke Registry were used to assess the concordance between predicted and observed initial stroke severity levels outside of the geographical area of its training dataset. RESULTS: In the years 2012-2019, the BSR reported 5883 IS and 816 ICH among people aged 16 or above, while 5623 IS and 787 ICH could be identified in the SNDS. Ten-day mortality was found to be the best proxy for initial stroke severity. In the Dijon Stroke Registry, among 1,254 IS, our algorithm predicted 53.0% events of minor severity (initial NIHSS≤4), 38.6% events of intermediate severity (5-20), and 8.5% events of high severity (≥ 21), compared to known prevalence of 53.0%, 38.0%, and 9.0%, respectively. No reliable predictions could be made for ICH. CONCLUSION: The possible estimation of prevalence of initial stroke severity levels with satisfying performances in healthcare database is likely to improve epidemiological surveillance of this disease at national and local level.