Using hospital discharge data to identify incident pregnancy-associated cancers: a validation study

The study population comprised 470,277 women who gave birth in New South Wales (NSW) in the period 2001 to 2008, which corresponded to 679,736 maternities. NSW is the most populous state of Australia with a resident population of approximately 7 million people. Approximately one-third of all Australian births occur in NSW public or private hospitals.

Data were obtained from three linked NSW population databases: the Perinatal Data Collection (PDC), Admitted Patient Data Collection (APDC) and Central Cancer Registry (CCR). The PDC is a statutory surveillance system that includes births of at least 20 weeks gestation or at least 400 grams birth weight. The APDC (referred to as “hospital data”) is a census of all inpatient hospitalisations to NSW public and private hospitals. The CCR (referred to as “registry data”) is a statutory case-based registry of all newly diagnosed cancers in NSW since 1972 with the exception of non-melanoma skin cancers. Cancer diagnoses in the CCR are coded according to the 3^rd edition of the International Classification of Diseases for Oncology [18]. The registry data are validated by a rigorous procedure to ensure the completeness and quality of data are high. Over 90% of cancers are verified by pathology and are confirmed as the primary cancer diagnosis. Less than 1% of cancers are ascertained from the death certificate [19]. Based on treatment categories, cancers were categorised into 13 clinical groupings [19]. For validation purposes, we used the registry data as the ‘gold standard’ of identification of incident cancers associated with pregnancy.

Record linkage of the three databases was carried out by the NSW Centre for Health Record Linkage using probabilistic record linkage methods [20, 21]. This involves a process of blocking and matching combinations of selected variables with identifying information such as name, date of birth, address and hospital. Each match was assigned with a probability weight [22]. The probabilistic record linkage for this study is highly reliable with less than three in 1,000 false positive links and less than five in 1,000 missed links [20]. The researchers were provided anonymised data. The study was approved by the NSW population and Health Services Research Ethics Committee.

The PDC birth records were used to identify women who gave birth from 2001 to 2008 and to obtain duration of pregnancy in completed weeks of gestation. The birth records were then linked to the cancer registry and hospitalisation records to identify newly diagnosed cancers during pregnancy or within 12 months of delivery [1, 2]. For identification of pregnancy-associated cancers from hospital data we searched up to 22 diagnosis fields associated with each pregnancy and postpartum hospital admission for cancer diagnoses. The diagnosis fields were coded according to the ICD-10-AM coded (10^th revision of the International Classification of Disease, Australian Modification ) with cancer diagnoses indicated by codes of C00–C96 [23]. As our aim was to identify incident cancers, any hospitalisation record coded as secondary, in-situ or benign cancer was excluded from the analysis. Hospital-identified pregnancy-associated cancers were examined based on two groups: i) “index cancer hospitalisation”, identified by taking the record with the earliest admission date for each cancer type. This replicates a scenario where individuals can be identified in the hospital data and the first record is considered to be an incident cancer; ii) ‘all cancer hospitalisations’, refers to the total number of hospitalisations with a cancer diagnosis and women could have multiple hospitalisations during pregnancy. This replicates a scenario where identification of individuals is not possible and hospitalisations (records) are used as the unit of analysis.

The incidence estimated from ‘all cancer hospitalisations’ and the ‘index cancer hospitalisation’ were compared with the incidence estimated from registry data (both overall and by cancer type). The ratios of the hospital cancer incidence to the registry cancer incidence (denoted incidence rate ratio, IRR) were calculated. To assess the validity of the hospital identification of incident pregnancy-associated cancers only ‘index cancer hospitalisation’ data were compared with registry data. The following reporting characteristics with 95% exact binomial confidence interval (95%CI) were computed: sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Sensitivity (sometimes referred to as true positive fraction) is the proportion of maternities with an associated cancer, as ascertained by the registry gold standard, that were also identified in the hospital data, thus measuring completeness of identification. Specificity (equivalent to one minus false positive fraction) is the proportion of maternities without an associated cancer that were also not identified in the hospital data. PPV is the proportion of maternities with an associated cancer identified in the hospital data that in fact had an associated cancer. NPV is the proportion of maternities without an associated cancer identified in the hospital data that in fact did not have an associated cancer. We compared admission date and registry date among true-positive pregnancy-associated cancers to quantify the discrepancy between hospital and registry data regarding the timing of cancer incidence. In addition, we examined the false-positive and false-negative pregnancy-associated cancers in the hospital data (compared with registry data) to determine if some of these were prevalent cancers or had an inconsistent cancer type. To differentiate truly incident from false-positive cancers, we compared the hospital-identified cancers with registry data from 1994 onwards and regarded those previously notified cancers of the same type as prevalent cancers. Analysis was carried out in SAS, Version 9.2 (SAS Institute, Cary NC, USA) [24].

Between 2001 and 2008, a total of 988 pregnancy-associated cancers were identified from the registry data, corresponding to an overall crude incidence of 145.4 per 100,000 maternities (Table 1). The resultant incidence estimates of hospital-identified pregnancy-associated cancer were notably higher among “all cancer hospitalisations” (252.7 per 100,000 maternities, IRR = 1.7) and lower among the “index cancer hospitalisations” (113.3 per 100,000 maternities, IRR = 0.8) (Table 1). Breast and lymphohaematopoeitic cancers resulted in large number of hospitalisations. The most common cancer in the registry data was melanoma with an incidence of 47.4 per 100,000 maternities, which was two times higher than that reported in the “index cancer hospitalisation” (20.9 per 100,000 maternities, IRR = 0.4). The next most common cancers were breast (32.1 per 100,000 maternities in registry data) and thyroid and other endocrine cancers (19.6 per 100,000 maternities), and these rates were also comparatively higher than that reported in the “index cancer hospitalisation” (22.4 and 17.5 per 100,000 maternities, and IRR = 0.8 and 0.9 respectively) (Table 1).

Overall, the sensitivity of hospital record identification of cancer was 60.4% and PPV was 77.7% (Table 2), and excluding melanoma were 72.5% and 77.2% respectively. The sensitivities for the majority of pregnancy-associated cancers were over 70.0%, except for melanoma (36.1%), breast (62.9%), colorectal (64.3%) and head and neck cancers (69.2%) (Table 2). Respiratory and urogenital cancers had 100% sensitivities, although their sample sizes were small. There was some variation in positive predictive values by cancer type. Breast, thyroid and other endocrine, gynaecological, neurological and head and neck cancers had PPVs ranging from 81.8% to 87.3%. For the remaining cancers, PPV ranged from 56.4% for lymphohaematopoeitic cancer to 79.9% for melanoma (Table 2). Due to the rarity of pregnancy-associated cancer, the specificities and negative predictive values ranged from 99.9% to 100% (data not shown).

Of the 580 true-positive pregnancy-associated cancers, 66% had a corresponding cancer-associated admission date in the same month as the registry, 28% had an admission date 1–2 months prior to registration, 5% had 3–5 months prior to registration and only 1% had ≥ 6 months prior to registration.

Compared with registry data, there were 166 false-positive and 380 false-negative pregnancy-associated cancers among “index cancer hospitalisations” (Table 2). False-positive cancers were predominately lymphohaematopoeitic cancer (56.4%, n = 48) (Table 2). Of the 166 false positives, registry data indicated that 86 (51.8%) were prevalent cancers and 22 (13.3%) were inaccurate/inconsistent identification of the cancer type. The cancers that were most frequently misclassified in the hospital data were colorectal cancer (n = 7, misclassified as upper gastrointestinal cancer), followed by melanoma (n = 3, misclassified as breast or bone and other connective tissue cancers) and gynecological cancer (n = 4, misclassified as upper gastrointestinal cancer). False-positive prevalent cancers were primarily lymphohaematopoeitic (38 out of 86, 44.2%), melanoma (13, 15.1%), breast (14, 16.3%) and thyroid and other endocrine cancers (10, 11.6%). When examining the false-positive cancers by type and year of admission, there was no systematic variation in the distributions of false positives.

False-negative cancers were predominantly melanoma (53.7%, n = 204) (Table 2). Of the 380 false negatives, 60.3% were in-situ/localised melanoma and 24.7% were regionalised/distant cancers of other types.