The effect of time before diagnosis and treatment on colorectal cancer outcomes: systematic review and dose–response meta-analysis

Abstract

Background

This systematic review and meta-analysis aimed to evaluate existing evidence on the relationship between diagnostic and treatment intervals and outcomes for colorectal cancer.

Methods

Four databases were searched for English language articles assessing the role of time before initial treatment in colorectal cancer on any outcome, including stage and survival. Two reviewers independently screened articles for inclusion and data were synthesised narratively. A dose–response meta-analysis was performed to examine the association between treatment interval and survival.

Results

One hundred and thirty papers were included in the systematic review, eight were included in the meta-analysis. Forty-five different intervals were considered in the time from first symptom to treatment. The most common finding was of no association between the length of intervals on any outcome. The dose–response meta-analysis showed a U-shaped association between the treatment interval and overall survival with the nadir at 45 days.

Conclusion

The review found inconsistent, but mostly a lack of, association between interval length and colorectal cancer outcomes, but study design and quality were heterogeneous. Meta-analysis suggests survival becomes increasingly poorer for those commencing treatment more than 45 days after diagnosis.

Registration

This review was registered, and the protocol is available, in PROSPERO, the international database of systematic reviews, with the registration ID CRD42021255864.

Introduction

Colorectal cancer is the third most common cancer diagnosed worldwide and the second most common cause of cancer death [1]. The incidence of colorectal cancer is expected to increase due to aging populations, changing patterns of risk factors and screening [2]. While population-based screening programmes exist in many countries, most colorectal cancers are diagnosed symptomatically after engagement with the healthcare system [3]. Consequently, there has been much focus on understanding and intervening during the time before diagnosis and treatment, and measures to reduce these times often form part of cancer control strategies or recommendations [4–6]. However, the evidence suggesting that reducing the length of the diagnostic and treatment pathway will lead to improved cancer related outcomes is mixed and many policies are based on expert consensus rather than empirical research.

The pathway to cancer diagnosis and treatment represents a diverse period of activities, often involving more than one healthcare service. The full length of time before initial treatment can be divided up into discrete periods, hereafter referred to as ‘intervals’, usually involving a specific health care service or diagnostic activity [7, 8]. The Aarhus statement made recommendations about research on diagnostic and treatment intervals to promote consistency across study design and reporting [5].

Eight systematic reviews have examined the effect of the length of the diagnostic and treatment pathway on different outcomes in colorectal cancer and have found mixed evidence. Ramos et al. conducted two reviews assessing the interval between symptom onset and either diagnosis or treatment and found no association between interval length and either survival [9] or stage [10]. Four reviews assessed the treatment interval only (time between diagnosis and treatment) with mixed findings: two found a negative association, such that in one review an interval longer than 4 weeks was associated with poorer survival [11] and in the other an interval of longer than 3 months was associated with poorer survival [12]; one found a combination of negative and non-linear associations with survival for colon cancer only [13]; and the final review found no association [14]. One review assessed multiple intervals on different outcomes and reported no association to be the predominant finding [15]. The final review specifically examined the role of time to diagnosis and treatment in people diagnosed under 50 years of age, and found mixed evidence of the length of intervals on stage and survival [16] These previous reviews also raised common issues across the literature including the heterogeneity of interval definitions and diverse interpretations of what constitutes delayed cancer care.

Of the previous systematic reviews, four performed a meta-analysis [9–12], with the remaining four noting that the heterogeneity in interval definition and use, and outcome selection, made pooling results difficult [13–16]. All the meta-analyses used a form of random effects model to pool results, with different approaches taken to standardise the intervals used in the original research. Two approaches converted all existing time categories to dichotomous intervals by keeping the shortest group and combining all other categories. However, the diversity in categorisation methods meant the shortest category ranged from up to 1 month to up to 6 months [9, 10]. The findings of the meta-analyses using this method were of no association between the interval and stage or survival. One meta-analysis only included studies that reported categorisation that could be converted to more or less than 1 month, with a secondary analysis for those that also reported a category around 12 weeks [12]. This meta-analysis found that treatment intervals of longer than 12 weeks were associated with poorer survival. The most complex method converted all reported hazard ratios to represent a 4-week ‘delay’, which could then be multiplied over the entire length of an interval [11]. The findings of this meta-analysis were that treatment intervals of longer than 4 weeks were associated with increased risk of death.

Previous research has indicated that the association between discrete intervals of time and outcomes, such as stage and survival, is non-linear, reflecting the role of confounding by indication, or the ‘waiting-time paradox’ [17–19]. This describes the phenomenon in which those who present with severe symptoms, and thus likely associated with later stage disease, may be treated more quickly. None of the techniques used in previous meta-analyses have the capacity to assess this non-linear association. Analysis using a dose–response technique is able to assess non-linear relationships. It can also accommodate intervals that have been split into different numbers of categories. Dose response meta-analysis allows for the fitting of a dose–response curve using only aggregate data from published studies [20, 21], thus enabling one to estimate a non-linear association. While dose–response meta-analyses have been used in cancer research previously, for instance in the relationship between physical activity and cancer mortality [22] and night shift work and breast cancer risk [23], this analytic approach has not been applied to studies of diagnostic and treatment intervals for cancer.

We aimed to systematically review and synthesise the evidence on the effects of time to diagnosis and treatment in colorectal cancer. This is the second reporting of a systematic review of the literature, with the first focusing on the methodological characteristics of the included studies and the specific epidemiological and statistical issues presented [24]. This current paper aims to assess the evidence, specifically, to update previous reviews, address shortcomings associated with heterogeneous interval definitions, and conduct a dose–response meta-analysis to examine the relationship between interval duration and colorectal cancer survival.

Methods

Design

A systematic review and meta-analysis of the literature was conducted to assess the role of any discrete interval of time before first treatment in people with colorectal cancer, on any outcome. This review was conducted according to the PRISMA guidelines [25], and the protocol was registered in the international prospective register of systematic reviews (PROSPERO, registration number CRD42021255864). Descriptive synthesis was used to summarise the evidence for all included papers, and a subset were included in the dose–response meta-analysis.

Search strategy and eligibility criteria

A more detailed description is provided elsewhere [24]. Briefly, four databases (Ovid Medline, EMBASE, EMCARE and PsycInfo) were searched from database inception to 2 November 2020. The search was subsequently updated to include articles published up to 11 November 2021 (search strategy for all databases shown in Supplementary Fig. 1).

Inclusion criteria

The inclusion criteria for studies were original research, published in English, and reporting a quantitative association between a length of time and any outcome. Eligible articles were also sought from previous reviews [9–15] and the reference lists of included papers. Additional inclusion criteria for the meta-analysis were assessment of the treatment interval (time between diagnosis and first treatment) on overall survival.

Article screening, data extraction and quality assessment

Each article was screened twice by two of four reviewers (AD, RJB, LB, KM), first by title and abstract and then by the full text. Any disputes were resolved by discussion or by a third reviewer. Data for the descriptive synthesis were extracted by one reviewer (AD), with a 10% check by a second reviewer (RJB). Data for the meta-analysis were double extracted by two reviewers (AD, RJB). Data were extracted using a study specific form. Extracted items included data on the study design, methodology and findings including sample characteristics. Attention was paid to how articles dealt with confounding by indication/the waiting-time paradox, through either the use of non-linear analysis techniques or controlling for factors such as emergency presentation and alarm symptoms, which are known to shorten the length of an interval [15]. Where possible, the interval definitions outlined in the Aarhus statement were used to describe and group intervals [5]. The Covidence program was used to manage screening and data extraction [26].

Risk of bias

The risk of bias of included studies was assessed using design-specific checklists from the Joanna Briggs Institute [27] (cohort or cross-sectional as appropriate). Each study was categorised as being at high or low risk, with high-risk studies also having the specific check list item that led to this decision being recorded.

Synthesis

The characteristics of each included article were summarised using descriptive statistics. The reported direction of effect for each interval and outcome analysis included in each article were tabulated based on the methods used in Neal et al., where the possible categories for each finding were longer interval worse, longer interval better, no association and non-linear association [15].

Meta-analysis

To be included in the dose–response meta-analysis studies had to assess the role of the treatment interval (between diagnosis and first treatment) on overall survival. This interval and outcome pair was selected as it was the interval most commonly considered across studies in the systematic review. In addition, the association between this interval and survival is prone to confounding by indication, and therefore is best represented by a non-linear association. Included articles had categorised this time interval, but any type or number of categories were able to be included. Additionally, to be included, articles had to report the data needed to perform a dose–response meta-analysis. This included, for example, a hazard ratio and confidence interval, the number of people included in each level of the categorised interval, and the number of deaths in each level of the categorised interval. The corresponding author of each eligible article was contacted to provide required data for the meta-analysis when these were not available in the published report. Where necessary, the number of deaths was extracted from the provided Kaplan–Meier plot and life tables using published formulae [28, 29] and using PlotDigitizer software [30].

To fit a dose–response meta-analysis, a ‘dose’ is required for each level of the included interval. The dose for each level of the reported interval was calculated using the method of Il’Yasova et al. [31]. Using this method, the midpoint of the range of values in the time interval category was used, and the lower margin plus half the range of the neighbouring category was used for intervals with no upper boundary (for example, the dose for two subsequent intervals of 15–30 days and >30 days would be set as 22.5 and 37.5 days, respectively). When required, the published regression estimates were converted to have the lowest valued category as the reference group [32].

A one-step dose response meta-analysis was performed using the dosresmeta package in R [33, 34]. The one-step technique was used to allow the inclusion of studies that divided the interval into only two categories (for example, dichotomised around the median value). Based on the most up to date evidence, the association between the treatment interval and survival is non-linear [35–37]. Therefore, a restricted cubic spline model was fitted, with three knots at the 10th, 50th and 90th percentiles [38]. This assumption of non-linearity was tested using the Wald test where the null hypothesis is that the slope of the second of the two splines is equal to zero [21]. Study heterogeneity was not assessed as this cannot be estimated when a dose–response method has been used. Publication bias was not assessed due to the small number of included studies.

Due to the observational study designs used, studies presented an adjusted analysis as their primary analysis. Because intervals are lengths of time, care must be taken to only control for factors that consider the temporal sequence of events. The main concern was adjustment for factors that occur after treatment, such as length of stay, surgical margins, post operative complications and adjuvant chemotherapy. Additionally, stage was considered an intermediary if it was noted to be recorded at treatment as opposed to diagnosis. Two analyses were therefore conducted: one including studies that controlled for intermediaries on the causal pathway, the other that excluded these.

Results

Study selection

Figure 1 outlines the selection of studies into the systematic review and meta-analysis. After two rounds of screening, 130 papers were eligible for inclusion in the descriptive synthesis [6, 17, 18, 35–37, 39–162], and eight in the meta-analysis [36, 37, 67, 97, 102, 138, 142, 157].

Fig. 1. PRISMA diagram.

Flowchart of included articles in the systematic review and dose-response meta-analysis.

Study characteristics

Table 1 outlines the characteristics of the included studies, which are described individually in Supplementary Table 1. Full details of the included studies are published elsewhere [24]. Predominately, the included studies used retrospective cohorts (n = 90, 69%), from previously collected data such as medical records (n = 89, 69%) and registries (n = 39, 30%). Half the research came from three countries (United Kingdom n = 28, 22%; United States of America n = 21, 19% and Spain n = 12, 9%) with only one included study conducted in a lower-income country (Ahmed et al. [40], from Pakistan). The most common outcomes assessed were survival/mortality (n = 80, 62%) and stage (n = 78, 60%), with only one article assessing each of the following outcomes: quality of life, patient satisfaction, healthcare costs, readmission rates and presentation types. Overall, 45 different discrete intervals were assessed in the included studies. Only 13 intervals (29%) were used more than once; these are depicted in Fig. 2. Forty-five (35%) studies were found to be of low research quality (described in Supplementary Table 1), most commonly because of insufficient reporting of details about the interval used or inappropriate statistical methods.

Table 1.

Characteristics of the included studies.

	All studies (n = 130)		Meta-analysis (n = 8)
	n	%	n	%
Study designa
Retrospective cohort	90	69.2	8	100.0
Cross-sectional	35	26.9	0	0
Case–control	3	2.3	0	0
Prospective cohort	3	2.3	0	0
Pre-test and post-test	1	0.8	0	0
Data sourcea
Medical records	89	68.5	2	25.0
Registry	39	30.0	4	50.0
Patient interview	23	17.7	0	0
Administrative data	12	9.2	3	37.5
Clinical registry	10	7.7	1	12.5
Patient survey	8	6.2	0	0
Clinician survey	5	3.8	0	0
Previous research	3	2.3	0	0
Population
Single hospital	66	50.8	2	25.0
Population based	43	33.1	5	62.5
Multiple hospitals	19	14.6	1	12.5
Primary care	2	1.5	0	0
Diagnosis included
Colon	20	15.4	4	50.0
Rectal	9	6.9	1	12.5
Both	101	77.7	3	37.5
Sample sizeb
<100	7	5.4	0	0
100–499	70	53.8	0	0
500–999	24	18.5	2	25.0
1000–4999	19	14.6	2	25.0
5000–9999	6	4.6	0	0
10,000–99,999	10	7.7	2	25.0
>100,000	7	5.4	3	37.5

^aMultiple options allowed, will add to more than 100%.

^bStudies with more than one cohort included separately.

Fig. 2. Intervals assessed more than once.

Names in brackets refer to intervals defined by the Aarhus statement [5].

Descriptive synthesis

The findings of the articles reporting on the most frequently assessed intervals are summarised in Table 2. For the most common outcomes, stage and survival, the most frequent finding across all intervals was of no association between the length of the interval and the outcome. This finding also holds true when studies appraised to have low quality were excluded. Where a directional effect was found, there tended to be an even split of articles finding longer intervals better and worse, with the exception of the secondary care interval (between referral to secondary care and start of treatment) on stage and survival, where the most common finding after no effect was a longer interval length associated with better outcomes. A non-linear association, specifically U-shaped, was observed in three intervals (treatment interval, diagnosis interval, and symptom onset to diagnosis interval).

Table 2.

Summary of the direction of effect for the six most common intervals.

Total interval	Longer interval better	No association	Longer interval worse	Non-linear associationb
Survival	Jullumstro (2009) (colon)	Barillari (1989)a	Iversen (2009) (rectal)
	McDermott (1981)	Bharucha (2005)	Terhaar sive Droste (2010)
	Mulcahy (1997)	Fegiz (1989)a
		Gonzalez-Hermoso (2004)a
		Goodman (1993)
		Iversen (2009) (colon)
		Jullumstro (2009) (rectal)
		McDermott (1981)a
		Olsson (2004)
		Roncoroni (1999)
		Thompson (2011)
Stage	Jullumstro (2009) (colon)	Barillari (1989)a	Korsgaard (2006) (rectal)
	Mulcahy (1997)	Bharucha (2005)	Langenbach (2003)
		Cerdan-Santacruz (2011)a	Langenbach (2010)
		Clarke (1970)a	MacArthur (1984)a
		Fegiz (1989)a	Roncoroni (1999)
		Gonzalez-Hermoso (2004)a
		Graffner (1986)
		Holliday (1979)
		Jarvinen (1984)
		Jullumstro (2009) (rectal)
		Kiran (2002)
		Korsgaard (2006) (colon)
		McDermott (1981) (colon)a
		Olsson (2004)
		Ratcliffe (1989)a
		Rubin (1980)a
		Terhaar sive Droste (2010)
		Thompson (2011)
		Turunen (1982)
		Wheeler (1999)a
		Young (2000)

Treatment interval	Longer interval better	No association	Longer interval worse	Non-linear associationb
Survival	Khorana (2019)	Abdulaal (2020)	Bagaria (2018)	Gleason (2021)
	Roder (2019)	Amri (2014)	Davidson (2021)	Kaltenmeier (2019) (W shaped)
		Auvinen (1992)	Gort (2010)	Kucejko (2020)
		Curtis (2018)	Grass (2020)	Lo (2021)
		Edwards (2021)	Lee (2019)	Shin (2013)
		Garcia-Botello (2021)	Sandar (2015)
		Pruitt (2013)	Yun (2012) (rectal)
		Roland (2013)
		Strous (2019)a
		Trepanier (2020)
		Wanis (2017)
		Yun (2012) (colon)
Stage		Abdulaal (2020)
		Van Hout (2011)a
Patient satisfaction			Porter (2005)
Readmission rates				Gleason (2021)

Diagnostic interval	Longer interval better	No association	Longer interval worse	Non-linear associationb
Survival		Cubiella (2021)	Bako (1998) (colon)	Torring (2013)
		Dregan (2013)	Pruitt (2013) (colon)	Torring (2012)
		Fernandez-de Castro (2019)		Torring (2011) (alarm symptoms)
		Pruitt (2013) (rectal)
		Redaniel (2015)
		Singh (2012)
		Stapley (2006)
		Torring (2011) (vague symptoms)
Stage	Pearson (2019)a	Chen (2017)a	Ortiz-Ortiz (2016)
		Fernandez-deCastro (2019)
		Fisher (2010)
		Majumdar (1999)
		Ozturk (2019)a
		Singh (2012)
		Stapley (2006)
Patient satisfaction			Porter (2005)
Quality of life		Walming (2018)

Symptom onset—diagnosis interval	Longer interval better	No association	Longer interval worse	Non-linear associationb
Survival	Porta (1991)	deSousa (2014)	Kim (2016)	Pita-Fernandez (2016) (rectal)
		Maguire (1994)
		Pita-Fernandez (2016) (colon)
		Polissar (1981)
Stage	Arbman (1996) (rectal)	Arbman (1996) (colon)	Kim (2016)
	Auvinen (1992)	Deng (2012)a	Pahlman (1985)a
	Leiva (2017)	deSousa (2014)
	Stubbs (1986)a	Gomez-Dominguez (2006)a
		Maguire (1994)
		Majumdar (1999)
		Pita-Fernandez (2016)
		Porta (1991)
		Ristvedt (2005)a

Patient interval	Longer interval better	No association	Longer interval worse	Non-linear associationb
Survival		Lim (1974)a
		Zhang (2015)
Stage		Ahmed (2021)
		Cerdan-Santacruz (2011)a
		Chen (2017)a
		Gomez-Dominguez (2011)a
		Majumdar (1999)
		Ozturk (2019)a
		Van Hout (2011)a
		Khattak (2006)a
Patient satisfaction		Porter (2005)
Quality of life		Walming (2018)
Type of presentation		Khattak (2006)a

Secondary care interval	Longer interval better	No association	Longer interval worse	Non-linear associationb
Survival/mortality	Comber (2005)a	Abdulaal (2020)
	Di Girolamo (2018)	Aslam (2017)
	Patel (2018)	Iversen (2009)
Stage	Comber (2005)a	Abdulaal (2020)
	Patel (2018)	Guzman Laura (2011)
		Ratcliffe (1989)a
		Tiong (2017)

^aIndicates studies with risk of low quality.

^bU-shaped except where noted.

The only interval that takes place outside of the healthcare system (the patient interval or time from first symptom to first presentation at a health service) was notable in that all of the included studies found no association with any outcome, which included survival, stage, patient satisfaction, quality of life or type of presentation to health care. More than half of the included studies of the patient interval were of poor quality (7/12, 58%), the highest proportion of any of the commonly used intervals, as detailed in Supplementary Table 1.

Only a few studies assessed the role of interval length on patient reported outcomes with mixed findings. Porter et al. [125] found a longer time was associated with poorer patient satisfaction for both the diagnostic and treatment intervals, but not the patient interval, and Walming et al. [156] found no association between the length of the patient or diagnostic intervals with quality of life.

Regarding confounding by indication, many of the included studies controlled for factors that could be acting as the ‘indication’, such as emergency presentations or alarm symptoms at presentation. When examining studies assessing the role of the treatment interval on survival, that considered the effect of confounding by indication (e.g., by fitting a regression model with restricted cubic splines), the most common findings were of a U- or W-shaped association [35–37, 97] or a longer time associated with poorer survival [45, 79]; one study found the opposite of a longer time associated with better survival [138].

Dose–response meta-analysis

The eight papers included in the dose–response meta-analysis are described in Table 3. Each article considered in the dose–response meta-analysis individually found mixed evidence on the direction of the effect between treatment interval and overall survival, with four finding a U-shaped association [36, 37, 97, 138], three finding no association [67, 142, 157] and one finding longer intervals associated with poorer survival [102].

Table 3.

Characteristics of studies included in the dose–response meta-analysis (n = 8).

Article	Disease site	Categories	n per category	n deaths per categorya	HR (95% CI)	Key inclusion criteria	Model adjusted forb
Edwards (2021)	Rectal	≤60 days	201	110	1.03 (0.63, 1.66)	Surgically resected, stage II–IV	Age, distance to treatment centre, congestive heart failure, prior cardiac arrest, diabetes, pathologic stage, treatment sequence, adjuvant chemotherapy, treatment centre
		>60 days	830	38	Reference
Kaltenmeier (2019)	Colon	<7 days	173,578	86,736	1.56 (1.45, 1.68)	Surgically resected, stage I–III	Age, education, income, distance from treating hospital, race, sex, comorbidity, insurance, tumour stage, hospital type, no. of lymph nodes examined, geographic location
		7–30 days	218,320	78,670	Reference
		31–60 days	72,082	26,828	1.13 (1.02, 1.25)
		61–90 days	13,882	5745	1.49 (1.19, 1.85)
		91–120 days	4707	2149	2.28 (1.61, 3.23)
		121–180 days	2819	1333	2.46 (1.48, 4.09)
Kucejko (2020)	Colon	1–2 weeks	53,436	37,971	1.424 (1.391, 1.458)	Stage I–III	Age, sex, race, stage, tumour size, tumour grade, nodes positive, nodes examined. Propensity scores calculated with age, sex, race, stage, colon site, Charlson Comorbidity Index, tumour size, tumour grade, presence of negative margins
(NCDB, >65 years)		3–4 weeks	30,042	17,309	Reference
		5–6 weeks	16,854	8398	1.021 (0.987, 1.056)
		7–12 weeks	11,453	3179	1.116 (1.124, 1.209)
		>12 weeks	2220	472	1.345 (1.256, 1.439)
Kucejko (2020)	Colon	1–2 weeks	33,079	21,092	1.379 (1.319, 1.442)	Stage I–III	Age, sex, race, stage, tumour size, tumour grade, nodes positive, nodes examined. Propensity scores calculated with age, sex, race, stage, colon site, Charlson Comorbidity Index, tumour size, tumour grade, presence of negative margins
(NCDB, <=65 years)		3–4 weeks	21,791	8211	Reference
		5–6 weeks	10,289	3384	0.985 (0.922, 1.052)
		7–12 weeks	6518	1441	1.217 (1.13, 1.31)
		>12 weeks	1106	241	1.68 (1.459, 1.933)
Kucejko (2020)	Colon	1–2 weeks	27,534	16,114	1.155 (1.111, 1.202)	Stage I–III	Age, sex, race, stage, tumour size, tumour grade, nodes positive, nodes examined. Propensity scores calculated with age, sex, race, stage, colon site, Charlson Comorbidity Index, tumour size, tumour grade, presence of negative margins
(SEER/Medicare)		3–4 weeks	11,343	5832	Reference
		5–6 weeks	4673	2419	1.047 (0.987, 1.244)
		7–12 weeks	2685	912	1.159 (1.079, 1.244)
		>12 weeks	414	131	1.255 (1.075, 1.466)
Lee (2019)	Both	≤30 days	36,115	14,773	Reference		Patient profile, economic status, environmental factors, health conditions, hospital attributes, multidisciplinary team involvement
		31–150 days	2553	1553	1.51 (1.43, 1.59)
		≥151 days	1252	960	1.64 (1.54, 1.76)
Lo (2021)	Colon	<1 week	45,436	15,527	1.71 (1.66, 1.75)	Surgically resected, stage I–III	Age, sex, race, ‘high school failure rate’, median household income, insurance type, metro/urban/rural, distance to treating facility, Charlson Comorbidity Index/Deyo comorbidity, pathological stage, adjuvant chemotherapy, differentiation, nodes, margins, tumour size, facility type
		1–3 weeks	57,055	11,502	1.12 (1.09, 1.15)
		3–6 weeks	54,297	9513	Reference
		6–9 weeks	18,246	3358	1.1 (1.06, 1.15)
		9–12 weeks	6239	1299	1.25 (1.18, 1.33)
		>12 weeks	6121	1482	1.43 (1.35, 1.52)
Shin (2013)	Both	≤1 week	994	240	1.68 (1.31, 2.16)	Surgically resected, local or regional stage	Age at diagnosis, sex, SEER stage, income tertile, place of residence
		1–4 weeks	626	92	Reference
		4–8 weeks	228	38	1.07 (0.73, 1.57)
		8–12 weeks	60	7	0.84 (0.39, 1.82)
		>12 weeks	38	14	2.65 (1.5, 4.7)
Strous (2019)	Both	≤35 days	474	77	Reference	Non-emergency	Age, gender, age, comorbidities, site, stage, differentiation, post-operative complications
		>35 days	316	61	1.2 (0.88, 1.64)
Wanis (2017)	Colon	≤30 days	368	97	Reference	Surgically resected, stage I–III Non-emergency	Age, pathologic stage, tumour grade, margin status, lymphovascular invasion, adjuvant chemotherapy
		>30 days	540	130	0.9 (0.48, 1.7)

SEER Surveillance, Epidemiology and End Results, NCDB National Cancer Database.

^aItalics in this column indicates values estimated from reported Kaplan–Meier plots and life tables.

^bItalics in this column indicate covariates that may lie on the causal pathway.

Results of the dose–response meta-analysis are shown in Fig. 3a. We found a U-shaped association between the treatment interval and overall survival with the nadir of the curve at approximately 45 days (test for non-linearity, p = 0.004).

Fig. 3. Dose-response meta-analysis results.

Results of the dose–response meta-analysis for all studies (a) and subset of studies (b).

Removing the four studies that adjusted for factors potentially on the causal pathway did not change the shape of the dose–response curve (test for non-linearity, p = 0.010) (see Fig. 3b), but the confidence intervals became wider. This also removed the only study that was deemed of low study quality [142].

Discussion

This systematic review and meta-analysis aimed to assess the association between time before initial treatment of colorectal cancer on any cancer-related outcomes. For all commonly used intervals and outcomes, the main finding across the literature was of no association between the length of an interval and poorer outcomes. However, several features may have limited the ability of studies to demonstrate an association, if it exists, including poor study quality, and methodological choices such as dividing the interval into a limited number of categories. Our review used a dose–response meta-analysis technique for the first time in this research field, finding a U-shaped association between the treatment interval and overall survival with a nadir around 45 days.

The predominant finding of no association in the descriptive synthesis is comparable to previous systematic reviews in this area. This may reflect methodological heterogeneity, but it could reflect the nature of most colorectal cancers as a slowly progressing disease, whose life-course is much longer than the time periods measured in the included studies. Evidence from tumour growth models and studies examining tumour doubling times suggest that the median time for a colorectal cancer to double in size is 211 days (range 112–404 days) [163], and the risk of metastasis within two months is 2.7%, increasing to 24.7% at 6 months [164]. Given the relatively short intervals assessed in the review, especially those occurring once a patient presents to a healthcare provider, there may be insufficient time for disease progression that would affect clinical outcomes such as survival and stage.

Even accounting for this possibility, the design of many studies prohibited the finding of an association if one did exist. This reflects methodological limitations such as arbitrary classification of time periods, the use of a small number of time categories, and assumptions about linear associations between time and outcomes [24]. For example, there were a number of articles that dichotomised the interval, limiting their ability to observe any non-linear effects, and substantially decreasing the statistical power of the analyses [165, 166]. While design and reporting guidelines exist for this research area [5], recommendations regarding the specific, and sometimes unique, methodological considerations involved in the analysis of interval research studies have only recently been reported [24].

The use of non-linear models allows for a more sophisticated depiction of potential associations between time and outcomes such as survival. The left-side slope of a U-shaped graph representing poorer outcomes for shorter intervals may indicate quicker clinical responses to patients with more severe symptoms. Studies that attempted to control for this waiting-time paradox do not fully resolve this issue. Many of the included studies controlled for factors such as emergency presentations, either by exclusion or by controlling for presentation type in statistical analysis, but the evidence was still mixed and tended towards a finding of no association. It is likely that factors that cause the shorter time/poorer outcome relationship are more complex than can be controlled for in statistical models.

The dose–response meta-analysis found a U-shaped association between the treatment interval and overall survival with a nadir around 45 days. This finding is comparable to the only article included in the meta-analysis that performed a restricted cubic spline regression in their original research which found a W-shaped association with the nadir at around 35 days [36]. The results of the dose–response meta-analysis provide stronger evidence that, at least for some sub-groups of patients with colorectal cancer, time to treatment may be associated with poorer outcomes. While this method has advantages over previous meta-analytic methods, one must be cautious in the interpretation of a U-shaped association between time and survival. The findings of the dose–response meta-analysis do not mean that the optimal time between diagnosis and treatment is defined by the nadir of around 45 days. The period before this nadir on the curve reflects confounding by indication, and the nature of the interventions to expedite treatment or triage patients in higher need, will vary by setting, and between individual patients and clinicians. A more appropriate conclusion of the U-shaped dose–response curve is that 45 days represents a potential upper limit of the length of the treatment interval, above which outcomes may become poorer, especially for those not presenting through emergency routes, or with alarm symptoms, two factors commonly associated with confounding by indication.

Overall, while there were few studies in the descriptive synthesis that assessed a non-linear relationship, the evidence around confounding by indication and the results of the meta-analysis suggest that there is likely an association between interval length and cancer outcomes. Most of the non-linear associations were U-shaped, showing increased risk of poor outcomes with longer intervals. An alternative method of determining a maximum acceptable delay, given the difficulty in interpreting U-shaped findings, is to nominate an unacceptable level of increased risk and determine the length of time this relates to.

Strengths and limitations

This systematic review of 130 articles is the largest conducted that is focused on time to diagnosis and treatment of colorectal cancer. Our systematic approach, following the PRISMA guidelines, increased the likelihood that all relevant studies have been included but we were limited to studies reported in English. Given the size of the review, only a 10% sample was double extracted for the descriptive review. However, all studies included in the meta-analysis were double extracted.

The selection of intervals and available data sources in included studies limited our ability to summarise and synthesise findings across intervals before diagnosis and treatment. Despite the number of included articles, more than half assessed a unique interval, specific only to that article. The frequent assessment of intervals such as the treatment interval reflects the increasing utilisation of population-based databases, allowing for larger and more representative samples, but limited to start and end points that are routinely collected in the healthcare system. Intervals earlier along the diagnostic and treatment pathway involving patient-reported events and primary care are less well documented in population datasets and subsequently were less likely to have evidence from larger studies. Similarly, none of the intervals found more than once examined the timing of investigations as either a start or an end point, despite the importance of this step in the diagnostic pathway. Consequently, we were limited in our ability to synthesise data on pre-diagnosis intervals. The specific intervals assessed across the entire body of literature were more diverse than those defined in the Aarhus statement. The most frequently used intervals aligned with points along the diagnostic and treatment pathway which reflect the role of specific healthcare providers, which can provide evidence for specific situations, and may be important for local service planning. However, even with the more commonly used intervals, we encountered heterogeneity of methods and poor reporting which meant only the treatment interval could be assessed in a meta-analysis. The overwhelming majority of studies included in this review were conducted in higher-income countries (United Kingdom, United States, Spain), with only one [40] from a lower-income country (Pakistan), limiting the generalisability of our findings. A recent systematic review that assessed the lengths of multiple intervals also found fewer studies conducted in lower-income countries, especially for diagnostic and treatment intervals. However, the evidence suggests that in many cases the reported intervals were longer for lower-income countries [167]. In addition to this lack of generalisability, with the exception of the United States, most of the evidence comes from countries with some form of universal health coverage, which utilise a variety of initiatives to reduce times, and will limit the interpretability of findings [168].

The use of a dose response analytical technique has many benefits over traditional techniques, with previous meta-analyses limited to assessing intervals as dichotomous categories and restricted to only those categories reported in the original research [9, 10, 12]. However, it requires data that were not readily available in published reports. Almost half of the eligible papers in this meta-analysis could not be included due to insufficient information, either in the articles or after contacting the corresponding author. In addition, while meta-regression techniques using individual patient data would be an improvement over the method used here, many of the larger studies used data from large databases that do not allow for the easy sharing of data (such as SEER, NCDB) leaving us having to synthesise reported summary statistics.

Translation into policy and practice

The uncertain evidence around the effect of time on outcomes is concerning given that reducing the time to diagnose and treat cancer is a common international strategy in cancer control. However, due to the non-linear associations between time and cancer outcomes, and the effects of confounding by indication, there are challenges in determining a specific time point beyond which outcomes become poorer, or even if a time point of this nature exists. Maximum acceptable intervals are important to formulate recommendations and policies to govern the health care system. The dose–response method used here provides another source of evidence that can help inform decisions about acceptable timing of events on the diagnosis and treatment pathway. We only had sufficient data to conduct a meta-analysis examining the treatment interval on survival. Currently, many of the existing recommended times along the diagnostic pathway are based on expert consensus alone [24, 169], although some have attempted to make greater use of evidence from studies that have analysed time as a non-linear association [3, 169]. Beyond expert consensus, other methods such as socially acceptable waiting times, for example the population consensus method used in the Netherlands (treek norm), may assist in ensuring recommendations are agreeable to everyone in the absence of stronger evidence about effects on outcomes [168].

The inclusion of social acceptability in addition to the clinical effect of waiting times is important because waiting times can influence patient emotional wellbeing and quality of life. Few studies assessed these outcomes, with mixed findings of either no association or longer times associated with poorer psychosocial outcomes. Interestingly, the only intervals where associations between interval lengths and clinical outcomes were found were those within the health care system, as no evidence of an association was found in the patient interval where perhaps more factors are within the patient’s control. Porter et al. [125] found the patient interval had the greatest variation in length, but noted it had the least effect on patient experience.

Conclusion and recommendations

Despite a significant volume of research in this area it remains uncertain how important time to diagnosis of colorectal cancer is on stage or survival. Our meta-analysis suggests an upper limit of 45 days for the treatment interval. However, we cannot provide further guidance on maximum limits for intervals leading to cancer diagnosis. The high frequency of studies finding no association could represent the natural progression of many colorectal cancers being relatively slow growing or it may reflect the sub-optimal quality and methodological approaches of many studies to date. Further well-designed studies of pre-diagnosis intervals, including between first symptom and first presentation (patient interval) and between first presentation in primary care and first referral to secondary care (primary care interval), are required that would enable similar use of dose–response meta-analytical analyses. This technique could also be applied more broadly to examine the effects of time to diagnosis and treatment on outcomes for other cancer types.

Author contributions

All authors have significantly contributed to the conceptualisation, design, conduct and reporting (writing and reviewing) of this systematic review.

Funding

AD is supported by a NHMRC Postgraduate Scholarship. RJB is supported by a Victorian Cancer Agency, Early Career Research Fellowship (ECRF20015).

Data availability

Data used in this review is provided in Supplementary Appendices; any additional data are available upon request to the corresponding author.

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

Ethics approval was not required for this review.

Supplementary information

The online version contains supplementary material available at 10.1038/s41416-023-02377-w.