Population-based stroke registries can provide valid stroke incidence because they ensure exhaustiveness of case ascertainment. However, their results are difficult to extrapolate because they cover a small population. The French Hospital Discharge Database (FHDDB), which routinely collects administrative data, could be a useful tool for providing data on the nationwide burden of stroke. The aim of our pilot study was to assess the validity of stroke diagnosis reported in the FHDDB. All records of patients with a diagnosis of stroke between 2004 and 2008 were retrieved from the FHDDB of Dijon Teaching Hospital. The Dijon Stroke Registry was considered as the gold standard. The sensitivity, positive predictive value (PPV), and weighted kappa were calculated. The Dijon Stroke Registry identified 811 patients with a stroke, among whom 186 were missed by the FHDDB and thus considered false-negatives. The FHDDB identified 903 patients discharged following a stroke including 625 true-positives confirmed by the registry and 278 false-positives. The overall sensitivity and PPV of the FHDDB for the diagnosis of stroke were, respectively, 77.1 % (95 % CI 74.2–80) and 69.2 % (95 % CI 66.1–72.2). For cardioembolic and lacunar strokes, the FHDDB yielded higher PPVs (respectively 86.7 and 84.6 %; p < 0.0001) than those of other stroke subtypes. The PPV but not sensitivity significantly increased over the years (p < 0.0001). Agreement with the stroke registry was moderate (kappa 52.8; 95 % CI 46.8–58.9). The FHDDB-based stroke diagnosis showed moderate validity compared with the Dijon Stroke Registry as the gold standard. However, its accuracy (PPV) increased with time and was higher for some stroke subtypes.