influence - individualized follow-up for breast cancer
Annemieke Jeeninga-Witteveen is a PhD student in the research Group Health Technology and Services Research (HTSR). Her supervisors are prof.dr. M.J. IJzerman and prof.dr. S. Siesling from the Faculty of Behavioural, Management and Social sciences (BMS).
The incidence of breast cancer is rising. Fortunately, because of earlier detection and better treatment modalities, mortality is decreasing.This leads to an increasing number of breast cancer patients at risk for recurrence and a potential shortage in health care capacity. In contrast to the primary treatment, cancer follow-up is still consensus-based and not tailored to particular (high-risk) subgroups. After curative primary treatment, patients are followed clinically for at least five years with annual follow-up visits, after which the frequency is determined based on age. The aim of follow-up is threefold: 1) psychosocial support, 2) monitoring of treatment effects and, 3) early detection of locoregional recurrence (LRR) or second primary (SP) breast cancer. The first two aims are mostly addressed in the first year. After the first year the main objective is early detection of recurrent disease aiming for survival benefit. Early detection of distant metastases (DM) is not part of regular follow- up care, as earlier detection will not (yet) result in better prognosis. Of the patients curatively treated for primary invasive breast cancer, 4% will develop a local (LR) or regional (RR) recurrence (LRR) and almost 5% will be diagnosed with a SP breast cancer in the ten years following the primary diagnosis. Models for individualizing follow-up based on recurrence risk would enable clinicians to make informed decisions and to focus resources on patients with higher risk, while avoiding unnecessary and potentially harmful follow-up visits for women with very low risks.
To obtain more information on expected prognosis after recurrence, the association between the disease-free interval (DFI) and survival after a LRR or SP breast cancer was studied in Chapter 2. A total of 37,278 women first diagnosed with early breast cancer between 2003-2006 were selected from the Netherlands Cancer Registry (NCR). The NCR is a nationwide population-based registry, which records all newly diagnosed tumors since 1989. The information on patient, tumor and treatment characteristics, as well as data concerning recurrences within the first five years following primary breast cancer are recorded from the patient files by specially trained registration clerks. LRRs and SP tumors within five years of first diagnosis were examined and the five-year period was divided into three equal intervals. LRRs or SP tumors were diagnosed in 890 (2.4%) and 897 (2.4%) women respectively. Longer DFI was strongly and independently related to an improved survival after a LRR (long versus short: Hazard Rate (HR) 0.65, 95% Confidence Interval (CI) 0.48-0.88; medium versus short HR 0.81, 95% CI 0.65-1.01). Other factors related to improved survival after LRR were younger age and surgical removal of the recurrence. No significant association was found between DFI and survival after SP tumors. This study showed that survival after SP breast cancer and later LRR was relatively high, indicating the benefit of providing follow-up for early detection of recurrence.
For individualized follow-up based on recurrence risk, accurate prediction of LRR and SP breast cancer risk is required. Current prediction models mostly employ regression, but with large datasets machine learning techniques such as Bayesian Networks (BNs) may be better alternatives. Advantages of using BNs are the ease of interpretation due to the graphical representation, simple validation, possibility to include prior information, their flexibility of including both observational and causal inference, flexibility in outcome parameter within the model, and how they handle missing data. Using data from 37,320 women with early breast cancer from the NCR diagnosed between 2003-2006, different BNs were compared with models based on logistic regression in Chapter 3. BN structures were developed using 1) Bayesian network classifiers, 2) correlation coefficients with different cut-offs, 3) constraint-based learning algorithms, and 4) score-based learning algorithms. The different models were compared with logistic regression using the area under the Receiver Operating Characteristic (ROC) curve, an external validation set obtained from the NCR from 2007-2008 (N=12,308) and subgroup analyses for a high and a low risk group. The BNs with the most connections between variables showed the best performance in both LRR and SP prediction (c-statistic of 0.76 for LRR and 0.69 for SP). In the external validation, logistic regression generally outperformed the BNs in both SP and LRR (c-statistic of 0.71 for LRR and 0.64 for SP). The differences were nonetheless small. Although logistic regression performed best on most parts of the subgroup analysis, BNs outperformed regression with respect to average risk in SP cases. Although estimates of regression coefficients depend on other independent variables, there is no assumed dependence relationship between coefficient estimators and the change in value of other variables as in the case of BNs, this analysis suggests that regression is still more accurate for risk estimation for both LRRs and SP tumors.
In Chapter 4, the development and validation of the first-ever time dependent logistic regression model for the prediction of the annual risk of LRR of breast cancer is discussed, developed based on data from 37,230 patients. Data on primary tumors diagnosed in 2007 and 2008 from 43 Dutch hospitals was used for external validation of the performance of the prediction model (n=12,308). The resulting model was implemented in the online nomogram INFLUENCE (www.utwente.nl/influence). Risk factors were determined using logistic regression and the risks were calculated per year, conditional on not being diagnosed with recurrence in the previous year. The model takes into account the age of the patient, grade of differentiation, size, multifocality, receptor status and nodal involvement of the primary tumor, and whether patients were treated with radio-, chemo- or endocrine therapy. The risk factors used in our model are filtered from the population based registry and are readily available in (Dutch) clinical practice, which makes the use of the INFLUENCE nomogram possible without extra efforts or data gathering. The index cohort showed c-statistics of 0.84, 0.77, 0.70, 0.73 and 0.62 respectively per subsequent year after primary treatment. Model predictions were well calibrated. Bootstrapping was used for internal validation and displayed only a small overestimation of the risk of developing a LRR. Estimates in the validation cohort did not differ significantly from the index cohort. This time-dependent logistic regression model and nomogram for the prediction of the annual risk of LRR of breast cancer is simple to use and shows a good predictive ability in the Dutch population. It can be used as an instrument to identify patients with a low or high risk of LRR who might benefit from a less or more intensive follow-up after breast cancer. It can support shared decision-making on personalized follow-up schedules.
Women remain at risk for recurrence after five years and after a first recurrence. The pattern of recurrence risk and prognostic factors for the development of subsequent breast cancer recurrences can provide valuable information for informed decision-making and patient-centered follow-up. We therefore investigated the pattern of site-specific recurrence and identified prognostic factors for first and subsequent recurrences during a follow-up of ten years using data from 9,342 women treated for primary invasive breast cancer in Chapter 5. Extended Cox regression was used to model the hazard of recurrence over ten years of follow-up for not only site-specific first, but also subsequent recurrences after LR or RR. In total, 362 patients had LR, 148 RR and 1,343 DM as first recurrence. The risk of first recurrence was highest during the second year post-diagnosis (3.9%; 95% CI 3.5-4.3) with similar patterns for LR, RR and DM. The pattern, as well as identified prognostic factors for overall first recurrence, seemed to be dominated by the high percentage of DM. When recurrence risk was analyzed according to site, a difference in identified prognostic factors was present. The hazard of subsequent recurrences after LR and RR both declined towards the end of follow-up. The risk of developing a second recurrence was significantly higher after RR than after LR (50% vs 29%; p<0.001). And after a second recurrence, more than half of the women were diagnosed with a third recurrence. Young age (<40 years), primary tumor size >2 cm, tumor grade II/III, positive lymph nodes, multifocality and no chemotherapy were prognostic factors for first recurrence. As most risk factors only have modest effects, multiple risk factors need to be taken into account for risk prediction and subsequent follow-up decisions. Although the percentage of patients with first recurrence that develop a second recurrence is high, the percentage with a second recurrence among all breast cancer patients remains very low (1.9% during ten years). And as almost 50% of the second recurrences will be diagnosed in the first follow-up year after the diagnosis and treatment of the first recurrence, combined with the low absolute number of second recurrences, more intensive follow-up for detection of subsequent recurrence is not likely to be (cost-)effective.
The guideline for breast cancer follow-up was revised in 2012 in the Netherlands. Changes were the frequency of physical examinations in the first three years and the recommendation to provide a personalized aftercare plan. In Chapter 6 we determined the guideline adherence, differences in follow-up before and after the revision and the extent of personalization based on known risk factors. Patients curatively treated for primary invasive breast cancer in 2010 or 2012/2013 from five Dutch hospitals were selected from NCR (n=250). The number of follow-up consultations and mammograms were compared using a one-way analysis of variance. Poisson regression was used to determine the influence of known risk factors of recurrence on the number of follow- up consultations. The 2010 sample received significantly more consultations than the 2012/2013 sample in the first year (2.4 vs. 2.0). Follow-up before and after publication of the 2012 guideline was almost similar after the first year, despite differences between the recommendations. Also an underuse of mammograms was observed in both samples, especially in the first year. The number of consultations was increased for patients diagnosed in 2012/2013 when radiotherapy had been part of their treatment and those with a smaller tumor size, compared to patients diagnosed in 2010. The recurrence risk was determined using the INFLUENCE nomogram. Unexpectedly, higher recurrence risk resulted in less visits. The recommendation to develop individual aftercare plans did not lead to actual personalization of follow-up. Aftercare plans were only found for 7% of the patients and in only one of the five hospitals. They were not personalized but uniform and implemented as part of an ongoing study. Despite providing different recommendations, follow-up after breast cancer before and after publication of the 2012 guideline was similar. In fact, before 2012 follow-up already resembled the 2012 guideline, suggesting that the new guideline might have been a formalization the current practice at that time. Besides the poor adherence in general, the significant decline in visits in the first year could be an indication of change in practice because of the new guideline. More evidence, incorporated in a shared decision process together with all clinicians involved and the patient is warranted in order to get towards personalized follow-up.
After five years of annual follow-up following breast cancer treatment, Dutch guidelines are age-based: annual follow-up for women <60 years, 60-75 years biennial and none for >75 years. With these age-based recommendations, implicit assumptions are made as to what level of risk is appropriate for which frequency of follow-up visits. Ideally, risk groups should clearly differentiate in risk and the frequency of follow-up visits should match the risk of recurrence, with higher risk patients receiving more follow-up visits. In Chapter 7, we analyzed the long-term breast cancer recurrence patterns and determined how the current age based recommendations on the follow-up schedules after five years correspond to the actual risk of LRR and SP. Moreover, alternative age cuts-off were proposed and risk of LRR or SP was compared to the risk of breast cancer in the general screening population aged 50-74 years. Women with early stage breast cancer in 2003/2005 were selected from the NCR (N=18,568). When comparing with the five-year risk of primary breast cancer in different age groups of women without a history of breast cancer that are invited to the biennial national screening program, the risk of LRR and SP after five years of clinical follow up was at least twice as high. Using competing risk regression, we found that the current age-based recommendations for breast cancer patients after five year of follow-up do not match well with the actual risk of LRR and SP: the risk was lower for women aged <60 years after five years of follow-up who received annual follow-up, compared to women aged 60-74 receiving less intensive biennial follow-up. This contradiction was caused by the relatively low risk of the 50-60 year old group, which lowered the risk of the complete group, including the age group of younger women (<50 after five years of follow-up, <45 at diagnosis) with higher risks. If follow-up were to be based on age, alternative cut-offs for the age (<50, 50-69, >69), could separate this higher risk group from the women with lower risks and a more logical relation between the risk and the recommendations could be achieved, as well as more differentiation in risk. However, the differences are not large. This raises the question whether there should be a difference in recommendation at all and exemplifies that age, or any single risk factor for that matter, is not able to capture the risk differences in women at risk for recurrence and is not sufficient for determining follow-up. Other factors than age were of greater influence on recurrence risk and should also be taken into account for individualizing follow- up based on risk for recurrence.
Getting towards more personalized follow-up asks for specific modelling requirements, because of the sequential decisions and individual risks and effects. In Chapter 8 we gained insight in how to model optimal allocation of resources for personalized follow-up. We formulated a discrete-time Partially Observable Markov Decision Process (POMDP) which is a generalization of a Markov decision process. It allows us to model the sequential decision making process in which the information about the true state of the system is incomplete. Because the true health state of a patient, i.e., whether a patient is disease-free or not, is only partially observable, a POMDP is ideally suited to this problem. A finite horizon was used. The aim was to maximize the total expected number of quality-adjusted life years (QALYs). Transition probabilities were obtained from data from the NCR. The optimal policies were determined for three risk categories based on differentiation of the primary tumor, but the model could be used for any possible patient profile. Our results show that, indeed, the optimal follow-up policy depends heavily on the personal risk characteristics of a patient: a slightly more intensive follow-up for patients with a high risk and poorly differentiated tumor was suggested, and a less intensive schedule for the other risk groups. Secondly, we found that the outcomes of the POMDP model were quite sensitive to certain input parameters, mostly to the growth rate of an LRR and the life expectancy of a patient after (early detected) breast cancer. Better estimates of these parameters are necessary in order to apply this model in practice.
As age is currently used to determine the follow-up policy after the standard five years of follow-up, stratification by age was chosen to introduce personalized follow-up schemes with the POMDP model in Chapter 9. Using data from the NCR of 37,230 patients with early breast cancer between 2003-2006, the risk of recurrence was determined for four age groups (<50, 50-59, 60-69, >70). Optimal schedules were derived considering the risk of recurrence, benefit of early detection and also disutility of (false positive) mammography and biopsies. To investigate the potential benefit of providing more intensive follow-up, the decision to perform a mammography or to wait was made twice a year. Recurrent disease could be detected by both mammography and women themselves (self-detection). It was optimal to have more intensive follow-up around the peak in recurrence in the second year after diagnosis. A slightly more intensive follow-up was proposed for the first age group with the highest risk (five visits for <50 years), the other age groups were recommended less visits: four for ages 50-59, three for 60-69 and three for ≥70. We also found that the test history is of great influence on the optimal schedule, as such the number of visits provide an upper bound. This risk-based follow-up would lead to a small increase in the total QALYs and a cost savings of over €2,800,000 per cohort starting follow-up every year. We demonstrated how follow-up could be personalized based on the risk of recurrence for different age categories, also taking into account the benefits and harms of mammography. This enables clinicians to make informed decisions and focus resources on patients with higher risk, while avoiding unnecessary and potentially harmful follow-up visits for women with very low risks. Within the age groups there will still be heterogeneity in risk. To get towards truly personalized follow-up models, more characteristics than only age need to be accounted for. Fortunately the model can easily be extended to take into account more risk factors and provide even more personalized follow-up schedules.
We can conclude that the current consensus-based recommendation for follow-up do not reflect the actual risk of recurrence. Personalized follow-up based on individual risk can reduce the tremendous burden on the health care capacity due to the resulting higher patient volume requiring follow-up. Using the INFLUENCE nomogram, the risk of recurrence can be taken into account when deciding on follow-up after curative treatment. Furthermore, our POMDP models could be employed to optimize the schedules, with balanced benefits and harms. However, before actual use in practice, more precise input values with regard to tumor growth are necessary.
