Risk stratification of childhood cancer survivors necessary for evidence-based clinical long-term follow-up

Background: Reorganisation of clinical follow-up care in England was proposed by the National Cancer Survivorship Initiative (NCSI), based on cancer type and treatment, ranging from Level 1 (supported self-management) to Level 3 (consultant-led care). The objective of this study was to provide an investigation of the risks of serious adverse health-outcomes associated with NCSI Levels of clinical care using a large population-based cohort of childhood cancer survivors. Methods: The British Childhood Cancer Survivor Study (BCCSS) was used to investigate risks of specific causes of death, subsequent primary neoplasms (SPNs) and non-fatal non-neoplastic outcomes by NCSI Level. Results: Cumulative (excess) risks of specified adverse outcomes by 45 years from diagnosis among non-leukaemic survivors assigned to NCSI Levels 1, 2 and 3 were for: SPNs—5% (two-fold expected), 14% (four-fold expected) and 21% (eight-fold expected); non-neoplastic death—2% (two-fold expected), 4% (three-fold expected) and 8% (seven-fold expected); non-fatal non-neoplastic condition—14%, 27% and 40%, respectively. Consequently overall cumulative risks of any adverse health outcome were 21%, 45% and 69%, respectively. Conclusions: Despite its simplicity the risk stratification tool provides clear and strong discrimination between survivors assigned to different NCSI Levels in terms of long-term cumulative and excess risks of serious adverse outcomes.

based and correspond to their risk of serious adverse health outcomes (NHS Improvement, 2011a, b;Glaser et al, 2013;Independent Cancer Taskforce, 2015). The NCSI proposed a system of risk stratification principally based on cancer type and treatment received and ranges from Level 1 (supported selfmanagement) to Level 3 (multidisciplinary consultant-led clinical care) (NHS Improvement, 2011a, b;Glaser et al, 2013). It is a development of our previously proposed methodology for stratifying survivors into three levels of clinical follow-up care (Wallace et al, 2001). So far, there have only been a few small-scale studies assessing the risks of adverse health outcomes associated with using this risk stratification tool, which in aggregate relate to just over 900 survivors Eiser et al, 2006;Michel et al, 2009;Edgar et al, 2013).
This NCSI risk stratification tool presented here is simple and does not require detailed information on cumulative doses of individual cytotoxic drugs or radiation doses to major organs. This has considerable advantages as such details may not be readily available, particularly if treatments were given decades ago.
Levels 1 and 2 of clinical care proposed by the NCSI have substantially lower frequencies of hospital attendance compared with Level 3. It is essential to investigate whether individuals proposed to be followed-up at Levels 1 or 2 experience an increased risk of serious adverse health-outcomes, which could potentially be avoided, or ameliorated, with appropriate hospital follow-up.
There is a pressing need to reliably quantify the overall risks of adverse health-outcomes associated with the NCSI system of risk stratification, formally introduced in 2011 (NHS Improvement, 2011a, b), because it is in increasingly widespread use in the NHS for childhood cancer survivors. Such usage results from the widespread participation of CCLG Centres in the development work of the NCSI. Data from the British Childhood Cancer Survivor Study (BCCSS) provide an unrivalled opportunity to reliably estimate these risks, as there has been no previous largescale investigation of the serious health risks associated with the NCSI Levels of care in the long-term, and to the best of our knowledge this is the first risk stratification tool that has been proposed for general clinical use for childhood cancer survivors. We provide a large-scale population-based investigation of the risks of serious adverse health-outcomes associated with the NCSI Levels of clinical care up to 45 years from diagnosis using data from the BCCSS.

MATERIALS AND METHODS
Sources of data 1. The BCCSS population-based cohort was used to assess the risks of specific causes of death and subsequent primary neoplasms (SPNs). The BCCSS cohort consists of 17 981 individuals who were diagnosed with cancer, when aged 0-14 years inclusive, between 1940 and 1991, in Britain, and who survived at least 5 years. Information on exposure to initial radiotherapy and chemotherapy was available for the entire cohort in the form YES/NO/NO RECORD. Further details including objectives, methods and response rates are available elsewhere (Hawkins et al, 2008 were treated within one of the national Medical Research Council (MRC) randomised trials into treatment of acute lymphoblastic leukaemia (ALL) were used to investigate risks of adverse health outcomes in relation to treatment received, based on the assumption that treatment was delivered, as prescribed in protocol.
Risk stratification. The NCSI clinical Levels of care are described in Figure 1A (further details in Supplementary Appendix 1). For the purposes of the present investigation risk stratification, based on childhood cancer type and its treatment, was undertaken as described in Figure 1B for all childhood cancers except leukaemia and for individuals who were diagnosed with ALL and treated within an MRC trial, as described in Figure 1C.
Ascertainment and grading of adverse health events. Ascertainment of causes of deaths and incident SPNs within the BCCSS was entirely population-based and achieved by individual patient electronic record linkage via the NHS Information Centre. For each death, we obtained the death certificate and underlying cause of death as coded by the Office for National Statistics. Potential SPNs were confirmed by reviewing relevant diagnostic reports, particularly histopathology reports. Further details are available in our most recent detailed investigations of causes of death (Reulen et al, 2010;Fidler et al, 2016) and SPNs (Reulen et al, 2011) in our cohort. Occurrence of non-fatal non-neoplastic adverse health-outcomes were ascertained through the BCCSS questionnaire. Each non-fatal non-neoplastic condition was graded using the CTCAE version 3 (Cancer Therapy Evaluation Programme, 2006) by three authors ERL and DLW in collaboration with KCO of the North American Childhood Cancer Survivor Study (CCSS) to ensure comparability with the CCSS. This version of the CTCAE (Cancer Therapy Evaluation Programme, 2006) was used because all related published studies used this version and such standardisation facilitates making satisfactory comparisons. Only events graded 3 or 4 were included here. In addition to indicating the condition, the survivor provided the diagnosis date/age. Potentially treatmentrelated non-fatal non-neoplastic events were grouped into 10 specific categories as defined by the CTCAE (vision; hearing; speech; circulatory; pulmonary; gastrointestinal; renal; musculoskeletal; neurological; endocrine), plus an overall category for any of these events.
Statistical methods. Risks of three outcomes were investigated: SPNs; fatal non-neoplastic conditions; non-fatal non-neoplastic conditions. Time at risk for each of these outcomes began at five years subsequent to first primary neoplasm (FPN) diagnosis.
End of period of risk depended on the adverse outcome being analysed. For SPNs/specific fatal non-neoplastic conditions, exit from risk was the date associated with the first of the following events: diagnosis of a SPN/death from specific cause, loss to follow-up, death from other cause or study end-date (the median questionnaire completion date). For non-fatal non-neoplastic conditions, exit from risk was the date of diagnosis of the condition, otherwise the questionnaire completion date.
Survivors were stratified according to their NCSI Levels of clinical care and the cumulative incidence by period of follow-up from diagnosis for each adverse health outcome was estimated. For the estimation of cumulative incidence of a SPN and specific fatal non-neoplastic conditions, other deaths were treated as competing risks (Gooley et al, 1999;Coviello and Bogges, 2004). For non-fatal non-neoplastic conditions, the complement of the Kaplan-Meier estimate (1-KM) was used to estimate the cumulative risk. Log rank tests were used to investigate for heterogeneity and trend in cumulative risk.
Standardised incidence ratios (SIRs) and standardised mortality ratios (SMRs) were calculated as the ratio of observed (O) to expected (E) numbers of relevant events (O/E). Expected numbers were estimated by accumulating person years at risk within specific gender and five-year age and oneyear calendar period strata and multiplying by gender, age and calendar period specific neoplasm and death rates in the general population of England and Wales. Poisson regression was used to test for heterogeneity and linear trend in SIRs and SMRs across NCSI Levels.
All analyses were carried out using Stata statistical software (version 13; Stata Corp., College Station, TX, USA). Statistical significance was taken at the 5% level, with two-sided tests.
Among NCSI Level 1 survivors restricting attention to those causes with more than five observed deaths, there was a five-fold excess of deaths from pulmonary causes. Among NCSI Level 2 survivors, restricting attention to those causes with more than five observed deaths, there was a four-fold excess in the number of deaths observed from cardiac, cerebrovascular, pulmonary and neurological causes. Classification to NCSI Level ( Figure 1B and C) used only limited information in the BCCSS computer record, before accessing available medical records. Detailed examination of the medical history of these cases, reviewed by physicians, suggested that a greater degree of routine follow-up would not have been helpful in preventing or delaying the excess deaths (data not published due to the potentially individually identifiable nature of the data).
Table 2 (lower half) gives the cumulative risks of specific potentially treatment-related non-fatal non-neoplastic conditions among survivors by NCSI Level. The cumulative risk of each specific condition increased with increased NCSI Level (all Po0.01), with the exception of gastrointestinal and musculoskeletal conditions.
Risk of SPNs, fatal and non-fatal non-neoplastic conditions among survivors of specific types of cancer assigned to particular NCSI Levels. For evidence-based long-term clinical follow-up the most practically useful risk stratification information relates to the risk of specific adverse outcomes among individuals with a specified type of cancer and NCSI Level. Supplementary Appendix 2 (Supplementary Table 1  with each specific cancer type (except ALL) and assigned to a particular NCSI Level for which there were at least 100 survivors still at risk at 25 years after diagnosis. After ALL the corresponding interval was 20 years.
SPNs. Among survivors assigned to NCSI Levels 1, 2 and 3 the cumulative risks by 25 years from diagnosis (20 years for ALL) were o1%, between 1% and 4%, and between 2% and 13%, respectively (Supplementary Appendix 2). The corresponding SIRs were o2-fold expected, between 2 and 5-fold expected, and between 3 and 17-fold expected, respectively. Among survivors of each specific cancer, both cumulative and excess risks increased with increased NCSI Level, but the risks varied strongly with cancer type.
Fatal potentially treatment related non-neoplastic conditions. Among survivors assigned to NCSI Levels 1, 2 and 3 the cumulative risks by 25 years from diagnosis (20 years for ALL) were o1%, between o1% and 2%, and between 1% and 3%, respectively. The corresponding SMRs were as expected, between 1 and 6-fold expected, and between 2 and 9-fold expected, respectively. Again, both cumulative and excess risks increased with increased NCSI Level after specific cancers, but varied importantly by cancer type.
Non-fatal potentially treatment related non-neoplastic conditions. The cumulative risks were 5%, between 5% and 12%, and between 7% and 19% by 25 years from diagnosis (20 years for ALL) among survivors assigned to NCSI Levels 1, 2 and 3, respectively.
Potential impact of missing treatment information. To assign individuals to a NCSI Level we needed the treatment information specified in Figure 1B and C. We have the advantage of the outcomes relating to SPN and death for the entire cohort irrespective of whether treatment information is missing or available. Therefore we can investigate cumulative risks of such outcomes among those with/without sufficient treatment information for risk stratification for evidence of heterogeneity. Similarly among those who returned questionnaires we can compare the risks of non-fatal non-neoplastic outcomes between those with/ without sufficient treatment information for risk stratification. Among the entire cohort of 13 130, 5-year survivors of all childhood cancer except leukaemia, 8675 (66%) had sufficient treatment available to assign to Levels 1, 2 or 3. In Supplementary Appendix 3 we investigate cumulative risk of SPN, death and nonfatal non-neoplastic outcomes between those with/without sufficient treatment information. In Supplementary Appendix 4 we undertake a similar investigation in relation to survivors of ALL. There was no evidence of important impact, except that the cumulative risk of SPNs among non-leukaemic survivors was higher among those with sufficient treatment information.

DISCUSSION
This study provides the first large-scale investigation of the longterm (up to 45 years from diagnosis) risks of serious adverse health-outcomes associated with a simple risk stratification tool which is already being used on an increasingly widespread basis within CCLG Centres throughout the UK. The study has the additional advantage of being population-based. The levels of clinical follow-up care proposed for childhood cancer survivors result from the NCSI (NHS Improvement, 2011a, b) and are a development of a previously proposed risk stratification tool (Wallace et al, 2001).
By 45 years from diagnosis, cumulative risks of developing any SPN, dying of any non-neoplastic cause or being diagnosed with any potentially treatment-related non-fatal non-neoplastic condition among survivors assigned to Levels 1, 2 or 3 were 21%, 45% and 69% (Table 3). Excess risks also increased with increasing NCSI Levels 1, 2 and 3: for SPNs two-fold, four-fold and eight-fold expected, respectively; for non-neoplastic deaths two-fold, threefold and seven-fold expected, respectively.
The risk stratification tool presented above, or an earlier version (Wallace et al, 2001), has been investigated in terms of the risks of adverse health-outcomes associated with Levels 1, 2 and 3 in only three previous studies, which included in aggregate just over 900 survivors Eiser et al, 2006;Michel et al, 2009;Edgar et al, 2013). The largest of these studies was undertaken by Edgar et al (2013) based on 607 (5-year) survivors originally diagnosed with cancer before aged 19 years between 1971 and 2004 at a single institution in Scotland. These investigators reported that the prevalence of adverse health-outcomes increased from 12% to 36% to 65% for Levels 1, 2 and 3, respectively. Restricted to grades X3 in CTCAE version 3 (Cancer Therapy Evaluation Programme, 2006), the corresponding prevalences were 1%, 11% and 39%, respectively. This is much lower than observed in our study and reflects the fact that the percentage of 5-year survivors with a current age beyond 25 years for Levels 1, 2 and 3 were 7%, 27% and 29%, respectively. The BCCSS is a mature cohort and the age distributions between the NCSI Levels do not vary much. From the analysis of all non-fatal non-neoplastic outcomes, 81%, 79% and 80% of the survivors were aged beyond 25 years for Levels 1, 2 and 3. Childhood cancer survivors experience elevated risks of adverse health-outcomes into middle age and beyond, and have an accelerated risk of events over that normally seen with ageing (Oeffinger et al, 2006;Reulen et al, 2010Reulen et al, , 2011Armstrong et al, 2014;Fidler et al, 2016). In the current study those assigned to NCSI Levels 1, 2 and 3 were found to experience an increasing risk of a severe adverse healthoutcome by 45 years from FPN diagnosis-21%, 45% and 69%, respectively.
The second largest study was based on 198 survivors recruited from one paediatric and one adult follow-up clinic in England, they found that Level 3 survivors, as defined by Wallace et al (2001) reported more short-term symptoms attributable to cancer treatment (for example, pain, fatigue, breathlessness) than Level 2 survivors, and Level 1 reported none Eiser et al, 2006). Also Level 3 survivors reported more late effects (for example, infertility, cardiac dysfunction, second cancers) than Level 1 or 2 survivors. However only eight survivors were classified to Level 1 and the survivors were aged only 16-39 years at survey Eiser et al, 2006).
The final of the three previous studies, including 112, 5-year childhood cancer survivors aged 18-45 years who were recruited from a late effects clinic, found that Level 3 survivors, reported more late effects than survivors classified to Level 2. However the Table 2. Observed and expected numbers, SMRs and cumulative risks of specific non-neoplastic causes of death (upper half). Cumulative risks for corresponding specific non-fatal non-neoplastic conditions (lower half) for survivors of non-leukaemic childhood cancers assigned to NCSI Levels 1, 2 and 3 Level 3 survivors were older than the Level 2 survivors and there was only one survivor classified to Level 1 (Michel et al, 2009).
Follow-up care proposed by the NCSI for Levels 1 and 2 have substantially reduced frequencies of hospital attendance compared with Level 3 (NHS Improvement, 2011a, b; Glaser et al, 2013). Therefore it was essential to investigate whether those individuals proposed to be followed-up at Levels 1 or 2 experience an increased risk of any serious adverse health-outcomes which could potentially be avoided, or at least diagnosed at an earlier stage, with appropriate hospital follow-up. Detailed investigation for such evidence revealed that the NCSI flexible system of adjusting the level of care depending on existing morbidities is unlikely to lead to deaths which might be prevented or delayed (data not published due to the potentially individually identifiable nature of the data).
Strengths of our study include its large-scale and populationbased design; also it benefits from substantially longer follow-up than available to any related previous study. In addition we have taken into account all serious adverse health-outcomes, including both those which were fatal or non-fatal, and so have avoided the limitation of previous related studies which were based on survivors alone. The population-based design of our study ensures that the observed and expected numbers, underlying the SMRs and SIRs, both relate to the entire population of Britain and avoid the potential biases relating to hospital-based studies. An additional strength of our study was that we had sufficient numbers to explore risk of adverse health outcomes by specific types of childhood cancer stratified by NCSI Levels of care. Such risk should be of practical benefit in clinics and for further developing standardised clinical follow-up guidelines (UK CCSG, 2005;SIGN, 2013;COG, 2013).
The principal weakness of our study relates to the crudeness of the cancer treatment information and the substantial fraction for whom treatment information was missing. However when sufficient numbers were available we have explored risks for specific FPN types who received cancer treatments of increasing levels of long-term toxicity by NCSI Levels 1, 2 and 3. Also our investigation of the potential impact of missing treatment on our risk estimates was mostly reassuring (Supplementary Appendices 3 and 4) and therefore it is unlikely that missing treatment has impacted the generalisability of our findings to the British population of childhood cancer survivors. Classification to NCSI Level used only limited information in the BCCSS computer record, before assessing available medical records. On subsequent detailed examination of their medical records we found that we inevitably classified some individuals initially to Level 1 or 2, when they should have been Level 3. However, since we reviewed the medical records of individuals contributing to each cause of death which was in excess of expected among Levels 1 and 2 survivors much of this misclassification was identified by the physicians. General population rates for England and Wales were used to generate expected numbers for SIRs and SMRs but excluding Scotland is unlikely to have important impact.

CONCLUSION
The proposed NCSI risk stratification tool is simple, giving a considerable advantage in terms of clinical application. However, it provides clear and strong discrimination between survivors assigned to the three different levels in terms of their long-term cumulative risk of serious adverse health-outcomes in an appropriate rank order. It also provides clear and strong discrimination between survivors assigned to the three levels in relation to excess risks. This is reassuring as this tool is already in Log-rank test for equality of observed risks comparing Levels 2 and 3 yields P ¼ 0.2052. (B) Observed and expected risks of any fatal nonneoplastic event after acute lymphoblastic leukaemia. Log-rank test for equality of observed risks comparing Levels 2 and 3 yields P ¼ 0.0018. (C) Observed risks of any non-fatal non-neoplastic condition after acute lymphoblastic leukaemia. Log-rank test for equality of observed risks comparing Levels 2 and 3 yields P ¼ 0.0002. Any subsequent primary neoplasm 5% (2-fold) 14% (4-fold) 21% (8-fold) Any potentially treatment related non-neoplastic death 2% (2-fold) 4% (3-fold) 8% (7-fold) Any potentially treatment related non-fatal non-neoplastic condition 14% 27% 40% Overall cumulative risk 21% 45% 69% increasingly widespread use within the NHS for childhood cancer survivors.
As the risk stratification tool provides strong discrimination between groups of survivors, in terms of their long-term risk of adverse health outcomes, it is likely to be useful internationally. Furthermore with such strong discrimination, the survivor strata identified provide a basis for intervention studies of the various elements which comprise models of care, with a full economic evaluation, within a wide variety of health care systems whether privately or state organised.