Employing two search strategies for case identification—diagnostic code searches and radiology text string searches—we found an annual incidence of pediatric ischemic stroke of 2.4 per 100,000 person-years in a retrospective cohort of Northern Californian children. We found that radiology searches were more sensitive than diagnostic code searches, particularly for perinatal ischemic stroke, although both search strategies had limited accuracy, with low PPV’s.
Published pediatric stroke incidence rates are difficult to compare because of the multiple variables that could affect these estimates: population, demographics (higher rates in black children),4, 17
study time period (pre or post the MRI era), age range included, and multiple aspects of study design (e.g., retrospective vs. prospective; search strategies; methods of case confirmation). Our overall ischemic stroke incidence rate of 2.4 per 100,000 person-years is 2 to 4 fold higher than previously published estimates in U.S. children that also included perinatal strokes ().1, 2
Our estimate of 1.3 later childhood
(non-neonatal) ischemic strokes per 100,000 person-years also exceeds estimates from those U.S. studies that excluded neonates.3–5
None of these prior reports used radiology searches to identify cases, and the majority of them relied only on inpatient stroke diagnoses.1–5, 18–20
Our study demonstrates the low sensitivity of such diagnostic code searches, suggesting that our expanded case identification methods, with the addition of radiology searches, may explain the higher ischemic stroke rate we observed. Had we relied only on stroke and CP ICD-9 code searches, we would have obtained an overall ischemic stroke incidence rate of only 0.97 per 100,000 person-years (95% CI: 0.78–1.2), a rate comparable to those reported in prior retrospective studies in the U.S. that used ICD-9 code searches alone ().1, 3–5
Comparison of incidence rate estimates for childhood stroke from prior and the current population-based studies
The highest published rate of childhood ischemic stroke (including perinatal stroke) was from a population-based study in Dijon, France (study period 1988–1989): 7.9 per 100,000 person-years.21
This was the only prospective study (), and therefore did not rely on any retrospective search strategies for case identification. Rather, it was a registry where all stroke cases were identified prospectively in both the in-patient and out-patient setting as part of a larger epidemiologic stroke study that included adults. Although the higher rate may also reflect a different patient population, it does suggest that incidence rates from retrospective studies are likely underestimates.
In our study, we found that ICD-9 code searches are not only insensitive for pediatric stroke but also fairly inaccurate: the poor PPV of many stroke codes has been previously reported, and was confirmed in our study (Supplemental-Table
These data have unfortunate implications on pediatric stroke research which has depended largely on retrospective observational studies. Administrative datasets, for example, have been useful because they allow the identification of a large number of subjects with this relatively rare disease. However, these studies typically rely on ICD-9 codes alone for case identification and therefore are not only missing false negative cases, but including false positives. The latter issue can be overcome in studies where cases can be confirmed through chart review, but this is typically not an option for studies utilizing administrative datasets. When available, radiology text-string searches, although time consuming with a low yield, appear to be a better option for retrospectively detecting ischemic stroke cases. However, in evaluating and designing studies relying on diagnostic code searches alone, investigators should consider which subjects are most likely to be missed: among the later childhood stroke cases that did not receive an ICD-9 code, almost half had meningitis or sepsis as their stroke etiology, and therefore might actually be excluded from certain stroke studies.
Our study has several limitations. Our conclusions regarding the relative advantages of the two search strategies are limited in that we studied only a single large managed care program. Because coding practices may differ in different institutions, our findings may not be generalizable to other settings. This may explain differences in the PPV of some specific ICD-9 codes compared to a prior pediatric report;8
however, overall, PPV’s were remarkably consistent between studies (Supplemental-Table
). Another limitation is that we likely failed to detect some cases, despite our use of two search strategies. Of particular concern are KPMCP patients that present acutely to a non-Kaiser hospital; head imaging studies performed outside of the KMPCP system would not have been included in our radiology text-string search. However, diagnostic codes for all out-of-plan care are maintained in KPMCP databases and were searched for this study. In addition, these patients usually return to the KPMCP system for follow-up head imaging and follow-up clinical care. However, missed cases would impact both our stroke incidence estimates and estimates of the sensitivity of the different search strategies. For the latter, the “gold standard” was simply the combination of the different retrospective strategies. Had our gold standard been a thorough and comprehensive prospective registry, as done in Dijon, France, the sensitivity estimates would have been even lower.
Despite these limitations, our study is the first to employ two search strategies to estimate the incidence of pediatric ischemic stroke in the U.S., and found a rate 2 to 4 times higher than prior reports. Although radiology searches appear to have a greater sensitivity for pediatric stroke, both search strategies have relatively low yield. These challenges in retrospective case identification support recent calls for the field to move towards prospective multicenter studies of pediatric stroke.5, 22