Trends in time to cancer diagnosis around the period of changing national guidance on referral of symptomatic patients: A serial cross-sectional study using UK electronic healthcare records from 2006–17

Highlights • Revised UK suspected-cancer guidance liberalised investigation of patients.• Diagnostic interval was longer for patients with newly introduced referral criteria.• Scope remains to reduce diagnostic interval further.


Introduction
Early cancer detection is central to improving outcomes [1]. Most early-detection strategies focus on the timely recognition and investigation of people likely to have undiagnosed cancer [2][3][4]. As screening detects <6 % of cancer [5], UK strategies focus on promptly recognising the symptoms, signs or test results associated with undiagnosed cancer ("features of possible cancer", or simply "features") [6]. In 2005, UK suspected-cancer guidance was published, listing features warranting cancer testing or investigation [7].
The guidance was revised in 2011 for ovarian cancer [8], and in 2015 for remaining cancers [2]. The aim was to expedite cancer diagnosis by lowering the risk of undiagnosed cancer warranting clinical action from ≥5 % to 3 % [2], which was achieved by introducing more vague features into the guidance [2,9]. The revised guidance is officially applicable in England, and endorsed in Wales and Northern Ireland [10].
Our objective was to explore the timeliness of cancer diagnosis in England, Wales and Northern Ireland in 2006-2017 for 11 common internal cancers. We compared time from first feature to diagnosis between two groups: "Old-NICE" (with features of possible cancer in the original 2005 guidance) and "New-NICE" (only participants with features introduced during guidance revision). We hypothesised that times to diagnosis would be longer for New-NICE than for Old-NICE participants, because diagnosing cancer is more challenging and may take longer when symptoms are vague [9,[11][12][13][14]. We also hypothesised that the difference in time to diagnosis between New-NICE and Old-NICE groups would reduce over time, as evidence on vague cancer features emerged and was translated into practice by guidance revision [2,15].

Study setting and design
This serial, cross-sectional, primary-care study used UK Clinical Practice Research Datalink (CPRD GOLD) with linked National Cancer Registration and Analysis Service (NCRAS, Set 15) data. CPRD GOLD comprises prospective, coded, and anonymised medical records from >600 UK general practices, with 389 having NCRAS linkage [16]. The study examined participants in the year before their cancer diagnosis between 2006 and 2017.
• Cancer typical of the opposite sex; e.g. male breast cancer.
• Screen-detected cancer, identified from NCRAS or by CPRD screening codes in the year before diagnosis. • No primary care attendance or no recorded feature of the participant's cancer in the year before diagnosis.

Variables and outcome measures
2.3.1. Features of possible cancer CPRD codes for features of possible cancer were collated [18], based on the symptoms, signs or blood test results in the original or revised guidance (Table 1) [2,7,8]. Occurrences of these codes, restricted to the relevant cancer site, identified participants presenting with these features in the year before diagnosis. Separate generic "suspected-cancer" codes were identified to explore for changing recording practices.

Milestone dates and diagnostic interval
The cancer diagnosis date was the earliest CPRD or NCRAS diagnostic code. The first recorded feature of possible cancer (index feature) was identified, along with the index date. Our outcome variable was "diagnostic interval": days from index date to diagnosis [19].

NICE grouping
Participants were grouped by their index feature(s) (Fig. 1, Table 1): • Old-NICE: participants with ≥1 index feature from the 2005 guidance [7]. • New-NICE: limited to participants who only had index feature(s) introduced during guidance revision [2,8].
Participants whose only index feature was a generic "suspectedcancer" code were omitted from analyses.

Other variables
Age and sex were identified from the CPRD year of birth, assigning a birthday of 1st July.

Analyses
Simple descriptive statistics summarised age (mean and standard deviation), sex (male, n, %), NICE grouping (New-NICE group, n, %), and the index feature(s) (n, % of all index features). We summarised diagnostic interval using mean (standard deviation) and the 25th, 50th, 75th, and 90th centiles. Diagnostic interval has a skewed distribution and was log-transformed for analyses [13].
Semiparametric varying-coefficient methods estimated coefficients representing the percentage difference in mean log-transformed diagnostic interval between New-NICE and Old-NICE groups (see accompanying methodological paper [20]). A coefficient of 0 represents no difference between the NICE groups. Positive coefficients indicate that diagnostic intervals are longer for the New-NICE than the Old-NICE group; negative coefficients, that they are shorter. The coefficients are estimated on a daily basis, so cannot be reported using a single summary statistic, and are plotted (with 95 % confidence intervals, using bootstrapping, n = 1000 replications [21]) to allow visualisation over 2006-17. The models adjusted for age and sex. Analyses examined each cancer site separately, sample size permitting (package "np" in R) [22].

Table 2
Numbers of potential inclusions (individual diagnoses), with Cancer Registry linkage, and exclusions, to give final sample sizes by cancer site. The final sample is described in terms of size (N), age (mean, SD), number (%) who are male, and number (%) with an index cancer feature introduced during guidance revision.

Study size
For the descriptive statistics, we included all CPRD participants meeting our inclusion criteria. Semiparametric varying-coefficient analyses were limited to cancer sites with participant numbers providing ≥90 % power at the 5 % level to detect a 14-day difference in diagnostic interval between New-NICE and Old-NICE groups. Assuming mean diagnostic intervals of 114 and 100 days, respectively, for the Old-NICE and New-NICE groups, a common standard deviation of 100 days and 10 % of participants classified as New-NICE requires 5980 total participants. An effect size of 14 days matches the two-week-wait target for urgent investigation. We assessed uncertainty in the estimates by confidence interval width.

Missing data and bias
To explore for potential bias associated with changing coding practice, we identified, for annual cohorts: (a) the percentages of participants excluded for having no coded features or only suspected-cancer codes; (b) the proportions of Old-NICE and New-NICE participants; (c) demographic characteristics of participants excluded because they lacked coded features.
Median (interquartile range) diagnostic intervals by year and by NICE grouping are plotted in Fig. 3 Fig. 3). Median diagnostic intervals were longer for New-NICE than for Old-NICE participants for colorectal (70 vs 51 days), oesophageal (77 vs 55 days), and lung (139⋅5 vs 124 days) cancers; however, this difference tended to decrease or disappear over time (Fig. 3). In ovarian cancer, diagnostic intervals were shorter in the New-NICE than in the Old-NICE group overall (56 vs 72 days), notably in 2010-16 (Fig. 3).

Semiparametric varying-coefficient analyses
Semiparametric varying-coefficient analyses were powered for bladder, breast, colorectal, lung, prostate and uterine cancers. The percentage differences (with 95 % confidence intervals) in mean logtransformed diagnostic interval between New-NICE and Old-NICE groups over time are plotted in Fig. 4.
After guidance revision on 23rd June 2015, New-NICE diagnostic intervals tended to shorten relative to those of the Old-NICE group in prostate (Fig. 4e) and uterine (Fig. 4f) cancers (note the downward trajectory towards the horizontal dashed line).
For colorectal cancer, the difference in diagnostic interval between the New-NICE and Old-NICE groups reduced over time. After guidance revision, New-NICE diagnostic intervals were shorter than Old-NICE intervals, as indicated by the trend dropping below the horizontal dashed line (Fig. 4c).
For lung cancer, New-NICE were longer than Old-NICE diagnostic intervals in the years 2006-10. In 2010-15, there was no difference between the groups. In 2016 (post guidance revision), New-NICE diagnostic intervals shortened relative to Old-NICE diagnostic intervals, but this was not sustained into 2017-18 (Fig. 4d).

Missing data and bias
The proportions of eligible participants excluded for lack of coded features increased over time for bladder, colorectal, lung, oesophageal, ovarian, pancreatic, stomach, and uterine cancers. This coincided with increased use of suspected-cancer codes (Fig. S1). The demographic details of excluded and included participants were similar (Table S1 and Table 2). The proportions of Old-NICE and New-NICE participants were largely similar across time within cancer sites (Fig. S2).

Findings
This study examined diagnostic intervals for 11 cancers in England, Table 4 Diagnostic interval (25th, 50th, 75th, and 90th centiles, mean and standard deviation) by cancer site. Wales and Northern Ireland over 2006-2017, a period including major revision of national suspected-cancer referral guidance. As hypothesised, times to diagnosis were generally longer for "New-NICE" participants (with index feature(s) of cancer introduced during guidance revision) than for "Old-NICE" participants (with feature(s) in the original guidance). Importantly, for colorectal cancer, New-NICE diagnostic intervals were shorter than Old-NICE diagnostic intervals after guidance revision. The gap between New-and Old-NICE groups decreased for prostate and uterine cancers over time, consistent with decreasing New-NICE diagnostic intervals aided by increasing Old-NICE diagnostic intervals for prostate cancer. The revised national guidance and GP responses to its preceding evidence base may have contributed to these changes, along with other early-diagnosis initiatives. In conclusion, scope remains to reduce time to diagnosis for symptomatic cancers in England, Wales and Northern Ireland.

Strengths and limitations
A considerable strength is the study's primary-care setting, where suspected-cancer guidance is implemented. The CPRD is the largest primary-care database worldwide and is recognised for its high-quality data [23]. We used established methods for case identification [18], with validation of cancer diagnosis by NCRAS where linkage was available. NCRAS data completeness improved in 2013 [24]. Pre-2013 studies report a concordance rate of 83⋅3% between CPRD and cancer registry information [25]. The CPRD diagnosis date was a median of 11 days (interquartile range -6 to 30 days) later than the registry date pre-2013 for colorectal, lung, gastrointestinal, and urological cancers [26]. Thus pre-2013 diagnostic intervals may be overestimated compared with post-2013 values. Reassuringly, no step-change in Newor Old-NICE diagnostic intervals were observed around 2013, suggesting that any associated bias is small.
We studied diagnostic interval rather than the primary care (time from index date to referral) or secondary care (time from referral to treatment) interval to avoid restricting analyses to participants referred to secondary care [19]. A limitation was the inability to analyse diagnostic intervals separately for participants referred via the two-week-wait pathway [27] because robust data sources for identifying them were unavailable to us. We found conflicting evidence of changes in GP recording practice over time. The proportion excluded for lack of coded features increased over time for some cancers, often coinciding with increased use of "suspected-cancer" codes. The proportions of Old-and New-NICE groups over time were constant and the similar demographic details for included and excluded participants suggests no marked selection bias. We excluded approximately 26 % of participants for lack of coded features, a proportion consistent with evidence that coded CPRD data identifies 80 % of visible haematuria or jaundice events, and 60-70 % of abdominal pain in patients with pancreatic or bladder cancers [28]. Of participants without recorded features, some will have presented at Emergency Departments without prior primary-care consultations [5,29,30], some will had the information recorded in "free text" [28], and others may have presented with features outside NICE guidance. Such features were deliberately omitted from our study, as irrelevant to our focus on guidance revision.
Our analytical method allowed us to explore trends in the difference in diagnostic interval between groups aligned by their index feature(s) to the revised (New-NICE) or original (Old-NICE) guidance [20]. The method was derived to explore the time-varying and gradual impact of emerging clinical evidence that is legitimised into clinical practice by official guidance revision and implementation [20].

Comparison with existing literature
Our findings build on previous analysis of the original 2005 NICE guideline's impact on diagnostic interval [13]. Mean diagnostic interval for 15 UK cancers reduced between 2001-2 and 2007-8 by 5⋅4 days (95 % CI: 2⋅4-8⋅5 days) from an initial value of 125⋅8 days. Similar to our study, median diagnostic intervals were shortest for cancers commonly presenting with lumps/masses (e.g. 26 days for breast) and longest for cancers often presenting with symptoms shared with other diseases (e.g. 112 days in lung cancer) [13]. Our estimates of diagnostic interval for colorectal cancer are similar to those obtained by the International Cancer Benchmarking Partnership using different data sources [31]. Our findings are consistent with the taxonomy of cancer symptom "signatures" and diagnostic difficulty [9]. Breast cancer had a narrow signature of a single alarm feature (breast lump) highly predictive of undiagnosed cancer plus the shortest diagnostic interval. In contrast, lung cancer had a very broad signature and the longest diagnostic interval.
Jensen et al. [27] investigated the impact of implementing a standardised cancer patient pathway in Denmark in 2007-2009. Post-implementation diagnostic intervals were 15 (12)(13)(14)(15)(16)(17) days shorter than peri-implementation values for the 37 % of patients actually referred via a cancer pathway, but were 4 (1-7) days longer for the 63 % of patients diagnosed via other routes. The authors concluded that the cancer pathways expedited diagnosis for a minority of patients.

Clinical interpretation and policy implications of the findings
The relationship between diagnostic interval and mortality (and stage) is U-shaped, reflecting confounding by indication [32][33][34]. Patients with advanced tumours generally receive an expedited diagnosis (possibly as an emergency) and have poor outcomes because of their high inherent mortality: the so-called "sick-quick". Conversely, patients presenting with vague symptoms usually have longer diagnostic intervals, and higher mortalitythought to reflect the impact of diagnostic delay, particularly between referral and diagnosis [32][33][34][35]. The revised guidance aimed to benefit patients by legitimising doctors to investigate at a lower risk of undiagnosed cancer. This change can reduce both diagnostic delay and emergency presentation. In this study, for colorectal cancer, New-NICE diagnostic intervals reduced relative to Old-NICE interval after guidance revision. This is consistent with general practitioners acting on the vague ("New-NICE") features introduced during guidance revision. Indeed, the proportion diagnosed via the urgent cancer referral pathway increased from 30 % (95 %CI 29 %-30 %) in 2013 to 33 % (33 %-34 %) in 2016, spanning the period of guidance revision [36].
Our findings of increasing Old-NICE diagnostic intervals over time may reflect growing strain on NHS diagnostic-endoscopy and imaging services [37], as demand for all indications (not just cancer) rises [38], particularly if CT-based targeted screening for lung cancer is introduced [39]. In 2018, inadequate diagnostic capacity was considered a rate-limiting step in the diagnostic pathway [40], and a negative impact of Covid-19 on diagnostic services is already becoming apparent [41].

Conclusions
We conclude that scope remains to reduce time to cancer diagnosis. The revised colorectal cancer diagnostic guidance may be exerting a downward pressure on time to diagnosis of this cancer, through impacts on the vague features of cancer introduced during guidance revision. Future studies using causal analysis should examine the impact of guidance revision on staging at diagnosis and survival for all cancers, and the possible downstream effects on investigative services. Policymakers are urged to enhance cancer diagnostic services so that they do not pose a rate-limiting step in the diagnostic pathway, and to protect them from the pressures of Covid-19.

Funding
This study was funded by Cancer Research UK [C56843/A21550], who were not involved in any aspect of the conduct of the study, in writing the manuscript or in the decision to submit for publication. This research is also linked to the CanTest Collaborative, which is funded by Cancer Research UK [C8640/A23385], of which WH is co-Director, GL is Associate Director, AS is Senior Faculty, and SP is an affiliated Research

Declaration of Competing Interest
WH was clinical lead of the guideline development group which formulated the revised NICE suspected-cancer guidelines (NG12). This paper is written in a personal capacity and is not to be interpreted as representing the views of the Group or of NICE. The remaining authors report no declarations of interest.