Summary of findings
In this cross sectional observational study, we explored temporal trends in first-ever cancer diagnoses by emergency admission in England between 1999 and 2008. Adjusted analyses highlighted patient groups at greater risk of diagnosis by this route. These groups included older patients and those living in the most deprived areas, as found in other studies [9, 29]. Even though the number of patients who had at least one previous all-cause emergency admission was small, this characteristic was associated with lower risk of subsequent cancer diagnosis by the emergency route. Unlike previous studies, we did not find regional variation in diagnoses by emergency admission once adjusted for other factors . Encouragingly, we found a reduction in the rate of first cancer diagnoses by the emergency route over time, which suggests that initiatives to improve cancer awareness in primary care and to improve access to cancer services may be having positive effects.
Comparison with other studies
Our finding that 13.9% of diagnoses were made during an emergency admission is similar to the estimate of 12.9% from the National Audit of Cancer Diagnosis in Primary Care which also included emergency referrals, and 16.4% of oesophagogastric cancers by Palser et al., 2013 [9, 29, 30]. However, our finding is lower than the estimate of 24% from NCIN, which was calculated from a combination of HES, cancer registry, and screening data as well as Cancer Waiting Times, and our earlier study which used HES data [9–11]. The discrepancy between earlier estimates and ours here may be due to differences in case assignment (diagnosis during two or three years compared with first-ever diagnosis, respectively) . Compared to other sources such as cancer registries, GPRD data have produced a lower incidence of cancer but the GPRD also appears to contain valid and reliable records of cancer diagnoses [19, 20].
Previous studies have found regional variations in the incidence of cancer and patient outcomes but in this study we found no statistically significant difference by practice region in the risk of diagnosis by emergency admission [4, 9, 12]. Our finding may be due to the relatively small sample size and more research is required for validation. Similarly, sex was also no longer statistically significant in adjusted regression analyses. This finding may again be due to the sample size or the cancer types included in the study. With known sex-differences in risk and outcomes by different types of cancers, further research and especially studies by cancer type should continue to consider sex as a potential predictor [8, 11, 16, 20].
Poor access to primary care services is one explanation for delayed diagnoses including those made by the emergency route, although the relationship is likely to be complex [4, 11, 32]. While the number of consultations with GPs or nurses in the 30 days immediately before diagnosis was not a statistically significant risk factor once all other factors had been considered, a much greater proportion of patients diagnosed by emergency admission also consulted with a GP or nurse in the preceding 30 days compared with patients diagnosed by non-emergency routes (13.8% compared with 0.46%, respectively). Lyratzopoulos et al. (2012) also found variation in pre-referral consultations by age, sex, ethnicity, and cancer type . Further exploration of this phenomenon in patients with potentially delayed or untimely and/or late diagnosis is warranted to determine if “alarm” symptoms are being overlooked .
Strengths and limitations
Our study benefitted from the use of linked primary and secondary care data combined with mortality data from the GPRD. Cancer registry data are typically considered the gold standard for research on cancer incidence and patient outcomes. However, both GPRD and HES data have shown comparable completeness to cancer registries [19, 33]. The integrated data are likely to improve the accuracy of patient identification beyond that achieved in previous studies [11, 13, 14, 29, 34]. We applied rigorous sampling criteria, tracking back to patients’ first records and cross-referenced between data from primary and secondary care, to capture patients’ first-ever diagnoses. Our sampling method is therefore more specific than other studies that used shorter timeframes for case ascertainment, such as Raine et al. (2010) and Bottle et al. (2012) who used one-year and three-year look-back periods respectively [11, 13]. Furthermore, the United Kingdom offers universal health coverage to its population, including registration with a GP, and people can only be registered with one general practice at a time . This makes the data collected by GPs in their electronic patient records a very valuable resource for research .
Limitations of this study include the lack of analysis by cancer type, which would be useful to inform clinical practice and management, especially given known differences in symptoms, treatments and prognoses between cancers. We performed our analyses at the patient level and not by cancer type because a minority of patients had dual first-ever diagnoses. Another reason for this approach was the lack of precise cross-mapping between Read codes and ICD-10 codes for cancer diagnoses that are required for analyses by cancer type. A further methodological consideration is that despite facilitating the accurate capture of first-ever diagnoses, our sampling method prevented the inclusion of patients diagnosed with secondary or recurrent cancers and patients not diagnosed with cancer for the first-ever time. However, the service-seeking behaviours and care pathways of patients defined as such may differ from patients with a first-ever diagnosis of cancer. Thus, it was appropriate for this study to exclude these patients, but these patient groups may be of interest for further study. Our findings may also be affected by missing data for some variables, such as ethnicity and social deprivation, although the risk factors identified in this study are supported by empirical evidence. As the modified Charlson Index was mapped using an ICD-9-CM based scoring method, there may have been inaccurate and incomplete comorbidity matching to Read and ICD-10 diagnosis codes in this study. This may be one explanation for the relatively low number of patients with recorded comorbidity.
We applied stepwise selection despite known limitations of this procedure, such as sensitivity to variable ordering and combinations. To select the final regression model, we considered a full model and those using forward, backward and interactive-forward approaches. There was little difference between the models in fit or the variables retained, in support of the appropriate use of stepwise selection in this study. The interval between initial primary care presentation and referral for the four main cancers (breast, colorectal, lung and prostate) is longer than 30 days for a large proportion of diagnosed patients [29, 37]. For this reason, a longer time interval, such as a window of the prior 3 months, may be more suitable when examining patterns in consultations preceding cancer diagnosis [1, 37]. Referral rates for suspected cancer differ across the country [12, 29, 38, 39]. Unfortunately, the quality of related data was poor in the study dataset and little interpretation of the results for referrals can be made without validation using other data sources.