H. Ellekjaer et al., Identification of incident stroke in Norway - Hospital discharge data compared with a population-based stroke register, STROKE, 30(1), 1999, pp. 56-60
Background and Purpose-The validity of hospital discharge diagnoses is esse
ntial in improving stroke surveillance and estimating healthcare costs of s
troke. The aim of this study was to assess sensitivity, positive predictive
value, and accuracy of discharge diagnoses compared with a stroke register
.
Methods-A record linkage was made between a population-based stroke registe
r and the discharge records of the hospital serving the population of the s
troke register (n = 70 000). The stroke register (including patients aged 1
5 and older and with no upper age limit), applied here as a "gold standard,
" was used to estimate sensitivity, positive predictive value, and accuracy
of the discharge diagnoses classification. The length of stay in hospital
by stroke patients was measured.
Results-Identifying cerebrovascular diseases by hospital discharge diagnose
s (International Classification of Diseases, 9th Revision [ICD-9], codes 43
0 to 438.9, first admission) lead to a substantial overestimation of stroke
in the target population. Restricting the retrieval to acute stroke diagno
ses (ICD-9 codes 430, 431, 434, and 436) gave an incidence estimate closer
to the "true" incidence rate in the stroke register. Selecting ICD-9 codes
430 to 438 of cerebrovascular diseases gave the highest sensitivity (86%).
The highest positive predictive value (68%) was achieved by selecting acute
stroke diagnoses (ICD-9 codes 430, 431, 434, and 436), at the expense of a
lower sensitivity (81%). Accuracy of ICD codes 430 to 438.9 (n = 678) reve
aled the highest proportion of incident strokes identified by the acute str
oke diagnoses (ICD-9 codes 430, 431, 434, and 436). Seventy-four percent of
hospital discharge diagnoses classified as first-ever stroke kept the orig
inal diagnosis. Only 4.6% of the discharge diagnoses were classified as non
stroke diagnoses after validation. The estimation of length of stay in the
hospital was improved by selection of acute stroke diagnoses from hospital
discharge data (ICD-9 codes 430, 431, 434, and 436), which gave the same es
timate of length of stay, a median of 8 days (2.5 percentile = 0 and 97.5 p
ercentile = 56), compared with a median of 8 days (2.5 percentile = 0 and 9
7.5 percentile = 51) based on the stroke register.
Conclusions-Hospital discharge data may overestimate stroke incidence and u
nderestimate the length of stay in the hospital, unless selection routines
of hospital discharge diagnoses are restricted to acute stroke diagnoses (I
CD-9 codes 430, 431, 434, and 436). If supplemented by a validation procedu
re, including estimates of sensitivity, positive predictive value, and accu
racy, hospital discharge data may provide valid information on hospital-bas
ed stroke incidence and lead to better allocation of health resources. Dist
inguishing subtypes of stroke from hospital discharge diagnoses should not
be performed unless coding practices are improved.