Optimal management of colorectal liver metastases in older patients: a decision analysis

Abstract

Background

Comparative trials evaluating management strategies for colorectal cancer liver metastases (CLM) are lacking, especially for older patients. This study developed a decision-analytic model to quantify outcomes associated with treatment strategies for CLM in older patients.

Methods

A Markov-decision model was built to examine the effect on life expectancy (LE) and quality-adjusted life expectancy (QALE) for best supportive care (BSC), systemic chemotherapy (SC), radiofrequency ablation (RFA) and hepatic resection (HR). The baseline patient cohort assumptions included healthy 70-year-old CLM patients after a primary cancer resection. Event and transition probabilities and utilities were derived from a literature review. Deterministic and probabilistic sensitivity analyses were performed on all study parameters.

Results

In base case analysis, BSC, SC, RFA and HR yielded LEs of 11.9, 23.1, 34.8 and 37.0 months, and QALEs of 7.8, 13.2, 22.0 and 25.0 months, respectively. Model results were sensitive to age, comorbidity, length of model simulation and utility after HR. Probabilistic sensitivity analysis showed increasing preference for RFA over HR with increasing patient age.

Conclusions

HR may be optimal for healthy 70-year-old patients with CLM. In older patients with comorbidities, RFA may provide better LE and QALE. Treatment decisions in older cancer patients should account for patient age, comorbidities, local expertise and individual values.

Introduction

There are over 39 million people in the US over the age of 65 years, an increase of 13.2% since 2000.¹ Furthermore, 54.7% of cancer cases are diagnosed in patients over 65 years.² Age-related increases in cancer incidence and the growing geriatric population is leading to increased numbers of older patients with cancer, the second most deadly being colorectal carcinoma (CRC).

Colorectal liver metastases (CLM) develop in 50–60% of CRC patients and are responsible for two-thirds of mortalities.^3,4 Metachronous CLM represent approximately 71% of disease recurrence in patients who underwent CRC resection, and are the most frequent initial recurrence site.^5,6 Survival with untreated CLM is dismal and most patients die within a year after diagnosis.^7–10 Surgery offers the highest cure rate, approximately 40% at 5 years;^11–15 however, novel strategies such as ablative therapies and evolving chemotherapy agents are also effective.¹⁶

Primary modalities for CLM management include best supportive care (BSC), systemic chemotherapy (SC), radiofrequency ablation (RFA) and hepatic resection (HR). For older patients, non-surgical therapies are often favoured, with the assumption that surgical morbidity and mortality are unacceptable owing to comorbidities or advanced age. Safety and success of HR has improved in the past two decades with careful patient selection, advances in anaesthesia and better post-operative care,¹⁷ prompting investigations into HR in older patients;^18–31 however, all studies are single-centre retrospective studies with limited numbers.

Management decisions for older patients with CLM is further complicated by balancing comorbidities, which increased treatment-related toxicity and competing causes of mortality, quality of life (QoL) and risks of therapy. As there are no clear guidelines and a randomized trial to examine this issue is unlikely, we chose to assess the relative efficacies utilizing a Markov decision analysis (DA) methodology. The objective of this study was to determine, from a patient perspective, the optimal strategy for the management of older patients (age ≥ 70 years) who present with liver metastases after primary CRC surgery. This study evaluated commonly used strategies for treating older patients with CLM using a decision-analytic model to determine gains in life expectancy (LE) and quality-adjusted life expectancy (QALE). No other study has focused on exploring treatment strategies specifically for the elderly. As this question may never be answered with randomized trials, this DA serves as a comprehensive synthesis of the current available evidence.

Methods

Model design

A Markov state transition model was developed using TreeAge Pro software (v2009; TreeAge Software, Inc., Williamstown, MA, USA) to evaluate the effectiveness of BSC, SC, RFA and HR for treating CLM in older patients. A Markov DA allows modelling outcomes for clinical problems associated with continuous (e.g. risk of recurrence/progression) as opposed to a one-time risk (e.g. risk of peri-operative mortality). Furthermore, it allows for modelling outcomes when the timing of events is important and when these events may happen multiple times.³² The Markov DA assumes that a patient is always in one of a finite number of health states and that events are represented as transitions from one state to another. A utility value, which is a QOL value on a 0–1 scale, is assigned to each health state. The overall QALE is the sum of time spent in each health state multiplied by the utility assigned to that health state.³²

In our model, it is assumed that all patients present with CLM after a resection of the primary colorectal lesion. This represents the majority of patients who present with resectable CLMs.^4,33 We also assumed that patients entering the analysis have CLMs that are amenable to all treatment options. This assumption allows fair comparison between strategies as the invasive strategies each have their own limitations such as the size of lesions for RFA and the distribution of lesions for HR. In reality, patients who are not amenable to all treatment options at presentation represent a heterogeneous group for whom therapeutic choices are often limited. A simulated patient is randomly allocated to one of four different treatment options: (1) BSC; (2) SC; (3) RFA; or (4) HR. Within a given treatment option, patients can only transition from one health state to another once per cycle. The model was simulated for 5 years using one-month cycle lengths with the assumption that tumour recurrence will not be detected before 1 month.³⁴

For the Markov DA model, several health states were defined for the four different treatment modalities. All events of interest were modelled as transitions between health states (Fig. 1). In the BSC arm, all CRC liver metastases are left untreated and patients are provided supportive care only. SC was defined as 5-FU, leucovorin and irinotecan, currently the standard regimen for CLM, which has established safety and efficacy in older patients.^35–38 Severe toxicity was defined as grade ≥ 3 toxicity according to the Common Terminology Criteria for Adverse Events scale.^39,40 An assumption of our model is that chemotherapy is halted if a patient develops severe toxicity and therefore there is no transition from severe to no/mild toxicity. With each cycle, there is also a probability that a patient transitions from no/mild toxicity to severe toxicity. RFA and HR were built with a similar structure to ensure model balance according to good modelling practice.^41–45 Peri-procedural chemotherapy was not modelled as separate strategies and was included in the analysis as part of RFA or HR. Recurrence was defined as any local or metastatic disease progression and resulted in patients receiving non-invasive treatments (i.e. BSC or SC) without repeat ablation or surgery. The choice of entering the BSC or SC treatment arms after either RFA or HR was related to the patient’s baseline co-morbidity (see below). A number of patients entering our model were defined to have RFA that was ineffective (inability to obtain complete oncologic clearance and therefore was not performed) and thus proceeded to receive BSC or SC. The effective and ineffective arms for RFA were to reflect clinical practice and create a balance with the HR arms of resectable and unresectable, respectively. Similarly, patients determined at the time of HR to have unresectable disease also preceded to BSC or SC. Treatment complications could impact outcomes in two ways: they impart a disutility (i.e. loss of health-related QoL) in the short term (3 months) and increase baseline mortality in the long term (up to 24 months). At the start of each simulated cycle, the patient can stay in their current treatment modality or move to another, depending on transitional probability (derived from probability of recurrence of 50% over 10 months), age and comorbidity.

Figure 1.

Schematic of decision analysis model with Markov health states. *At the start of each simulated month, the patient can stay in their current state or move to another. The model is run until one of four termination criteria is met. This process is repeated ∼10,000 times to determine the most likely outcome for a given group of patients. **Patient enters SC or BSC treatment strategy depending on comorbidity level; all downstream health states are identical to BSC or SC arm. CRC, colorectal cancer; LM, liver metastasis

The model was run until one of four conditions was met: (1) all simulated patients have died; (2) all simulated patients reach 100 years of age, (3) incremental benefits gained per cycle have become <0.001/cycle, or (4) 60 cycles (5 years) have passed. The upper age limit of 100 years was necessary given the paucity of reliable mortality data for patients older than 100 years. An incremental utility gain of less than 0.001 per cycle was defined as negligible in order to improve model efficiency. The 60-cycle (5-year) limit was placed on our model in order to enhance clinical relevance, as the lack of recurrence by 5 years after treatment generally defines a cure.⁴⁶ Another reason for this limit is the scarcity of data on survival and recurrence rates after 5 years after RFA or HR. The half-cycle correction was applied to all health states, after good modelling practice, in order to improve the accuracy of expected value estimates.⁴⁷

Age and comorbidities

Age ≥ 70 years was used to define older patients. This chronological definition was chosen owing to its common use in geriatric oncology studies.^48,49 Patient comorbidities were defined using the Charlson score, a weighted scoring system validated for predicting morbidity and mortality in CRC patients with liver metastases.⁵⁰ According to the overall score, patients were identified as having: low (0), medium (1), or high (≥2) comorbidity. Our model accounts for the effect of comorbidities in several ways. First, the level of comorbidity alters the probability of dying from an invasive procedure (RFA or HR). Previous studies in surgical oncology have demonstrated increased post-procedural mortality associated with increasing comorbidity.^51–53 Second, patients with high comorbidity (Charlson score ≥2) were excluded from receiving any systemic chemotherapy. Finally, our model accounts for the effect of comorbidity burden on LE by applying a comorbidity multiplier to standard life table mortality estimates.

We used a method first described by Welch et al., which used the declining exponential approximation for life expectancy (DEALE) model to quantify the effect of comorbidities on annual mortality.⁵⁴ The DEALE approximation assumes that, across any short period of time, LE is equal to the inverse of the annual mortality rate.⁵⁵ By extension, one can calculate the average annual mortality rate given any specific starting age by referring to standard life tables (Table 1). Welch’s method then assumed that healthier individuals have a longer LE and lower annual mortality rates than the average person. Conversely, sicker individuals have a shorter LE and thus higher mortality rates. For our model, we assumed that our defined categories of low, medium and high comorbidity translate to 75th, 50th and 25^th percentiles of LE from census data. We then calculated annual mortality using the DEALE model described earlier. Life expectancy values were derived from US census data specific to a 70-year-old individual.⁵⁶ Finally, a comorbidity multiplier was derived using the medium comorbidity level (and thus median life expectancy) as the reference (Table 1). This application of the Welch method has been used in other studies to approximate the impact of comorbidities on life expectancies.^57,58

Table 1.

Calculation of comorbidity multiplier

Life expectancya		Annual mortality	Charlson’s index	Level of comorbidity	Comorbidity multiplierb
Percentile	Years
75th	19.65	0.051	0	Low	0.715
50th	14.05	0.071	1	Medium	1
25th	8.1	0.123	≥2	High	1.735

^aPercentiles and remaining life expectancy derived from Life Tables of the United States.⁵⁶

^bThis modifies the annual probability of dying from other causes (Annual Mortality column in the table).

Outcomes were measured in LE and QALE. Health-related QoL was modelsed as utility values. In our model, utility was generally defined by the individual’s state of well-being derived from the preference one placed on being in that particular health state.⁵⁹ Typically, utilities range from 0 (equivalent to death) to 1 (representing perfect health). The base-case for primary analysis was a 70-year-old patient with a Charlson’s comorbidity index of 0 (low comorbidity), presenting with CLM after a CRC resection. As tumour-specific criteria for RFA and HR are different and often poorly supported by evidence,¹⁶ only patients with tumours potentially treatable by either strategy were included.

Model data

Probabilities were obtained from literature reviews by searching MEDLINE and EMBASE databases from 1950–2010 (Table 2). Studies with the highest-grade evidence were used to derive baseline probabilities. All sources were used to establish ranges for sensitivity analysis. In cases of equal grade evidence, probabilities were derived using an inverse variance-weighted pooling method.⁶⁰ This method is commonly employed in meta-analyses where the weight given to each study is the inverse of the variance of the effect estimate from each.⁶⁰ This method gives heavier weight to larger, as well as more precise, studies (i.e. with smaller standard errors). This generally produces more accurate pooled estimates than weighting by sample size alone. If parameter estimates were obtained from one source, 95% confidence intervals were used as ranges for sensitivity analyses. Where possible, data specific to patients over 70 years were used. Age-adjusted mortality probabilities were derived from actuarial life tables to model death from other causes.⁶¹ Hazard ratios for post-procedure morbidity and mortality were obtained from studies where multivariate regression was performed with both age and comorbidity as significant covariates. This allowed us to separately estimate risks for different age and comorbidity values.

Table 2.

Model parameters – baseline estimates, ranges and one-way sensitivity analysis

Variable	Baselinea	Rangea	Threshold within (outside of) rangeb	Strategy preferred below/above thresholdc	Reference
Global Variables
Age at diagnosis (years)	70	70–90	79.4	HR/RFA	–
Comorbidity by Charlson’s Index	0	0–3	>2	HR/RFA	–
Length of simulation (months)	60	12–120	39.3	RFA/HR	–
Probabilities
Procedure Success
RFA (Effectiveness)d	0.929	0.920–0.938	NT (NT)	–	64,67,68,70,71,73
HR (Resectability)d	0.908	0.900–0.916	NT (0.790)	(RFA/HR)	85–92
Death
Non-cancer deatha	Life Tablese	0.5–1.5 x Table	NT (2.7 x Table)	(HR/RFA)	61
On BSCa	0.0803	0.0719–0.0887	NT (0.0178)	(BSC/HR)	7
On SCa	0.0386	0.0346–0.0437	NT (0.0270)	(RFA/HR)	93
RFA (operative)	0.003	0–0.01	NT (NT)	–	63–73
HR (operative)	0.041	0.034–0.048	NT (0.104)	(HR/RFA)	18,19,21–28,30,31,74
Laparotomy (operative)	0.037	0.019–0.056	NT (NT)	–	94
Recurrence
RFAa	0.0642	0.0518–0.0766	NT (0.0496)	(RFA/HR)	63,68,70,71,73
HRa	0.0248	0.0192–0.0304	NT (0.0318)	(HR/RFA)	18,19,21,23,25–28,74
Complications
RFA	0.036	0.030–0.042	NT (NT)	–	64–66,68–73
HR	0.350	0.333–0.367	NT (0.553)	(HR/RFA)	18,19,21–24,26–28,31,74
Laparotomy	0.224	0.112–0.448	NT (NT)	–	94
Severe toxicity on SC	0.25	0.125–0.375	NT (NT)	–	95
Modifying ratiosb
Non-cancer death (by comorbidity)	Tablee	0.5–1.5 x Table	NT (2.7 x Table)	(HR/RFA)	Assumption
Procedural mortality (by comorbidity)	Tablee	0.5–1.5 x Table	NT (2.8 x Table)	(HR/RFA)	96
Mortality after complications (by time)	Tablee	0.5–1.5 x Table	NT (2.6 x Table)	(HR/RFA)	97
Utilitiesf
BSC	0.66	0.53–0.79	NT (NT)	–	Assumption
SC (no/mild toxicity)	0.60	0.48–0.72	NT (0.90)	(HR/RFA)	62
SC (mod/severe toxicity)g	0.48g	–	–	–	Assumption
RFA	0.72	0.58–0.86	NT (0.90)	(HR/RFA)	62
HR	0.72	0.58–0.86	0.60	RFA/HR	62
Complications HR/RFA/laparotomy	0.60	0.48–0.72	NT (NT)	–	62
Death	0	–	–	–	–

^aAll transitional probabilities listed as per-month equivalents.

^bProbabilities and utilities were varied within the plausible range as well as full range of 0 to 1.

^cStrategies within parentheses are for situations where the preferred treatment switch occurs at a threshold that is outside the plausible range.

^dProbability listed as an absolute value.

^eAll variables derived from tables were varied at ±50% of the values retrieved from tables.

^fRanges for all utilities are given as ±20% of baseline. See text for further details.

^gThis utility is always linked to the utility value for ‘chemotherapy with no/mild toxicity’ by a factor of 0.8X.

NT, no threshold; HR, hepatic resection; RFA, radiofrequency ablation; BSC, best supportive care; SC, systemic chemotherapy.

Utility values were derived from a prospective study by Ruers et al., who obtained health-related QoL data from 109 patients with CLM before and after treatment with HR, ablation and SC.⁶² As a result of the paucity of utility data, ranges used in sensitivity analysis were set as ±20% of the point estimates.

Model analysis

Only recurrence data were derived from our literature review; survival statistics were not pre-specified in the model. Instead, overall survival (OS) after RFA or HR reported in our literature review served as data for external validation by comparing reported survival-to-survival predicted by our model.

Base-case expected value calculations were performed in deterministic (using fixed values from best evidence) and stochastic (probability-based estimates using pre-specified distributions for each variable) manners. Significant differences between strategies were defined as incremental gains in LE or QALE of >1 month.

Sensitivity analysis

For one-way deterministic sensitivity analyses, all parameters were tested within the plausible range, and the entire range (for probabilities and utilities). Variables are considered sensitive if the preferred strategy (the one with maximum LE or QALE) changes when the variable is changed within the specified range. For sensitive variables in one-way analyses, threshold values are calculated to represent the value above or below which a different strategy would be preferred. Ranges for sensitivity analysis were derived from published literature as mentioned in previous sections and are presented in Table 2. In order to vary parameters in the form of tables such as the probability of non-cancer death, a multiplier set to a baseline of 1 was applied to such parameters to be used in the sensitivity analysis.

Two-way sensitivity analysis follows the same principles as one-way analyses except that two parameters are varied simultaneously through their specified ranges whereas all others remain constant. These provide additional information regarding the effects of uncertainty on the model outputs. However, owing to the sheer number of combinatorial pairs of variables available for analysis, two-way sensitivity analyses were reserved for continuous variables that were sensitive in the one-way analyses as well as variables hypothesized to be clinically important.

Therefore, for all parameters in a deterministic sensitivity analysis, except utilities, one-way analyses were performed on the LE model. Analyses for utilities used the QALE model. Two-way sensitivity analyses were performed on all clinically meaningful variables and those sensitive in one-way analyses.

Probabilistic sensitivity analysis was carried out on stochastic versions of our model to simultaneously evaluate the combined uncertainty of all variables in the model. This produces point estimates and standard deviations for the base case and estimations of strategy selection frequency. In cases where competing strategies provided <1 month of incremental gain, the result was defined as ‘indifferent.’ Uncertainty was assessed using 10 000 s-order Monte Carlo simulations sampling all parameter distributions. This allows us to generate 95% credible intervals for LE and QALE estimates for each strategy.

Internal and external model validation

Once our final model was constructed, we performed rigorous internal testing to ensure its integrity and internal validity. We began by examining each branch of the model for syntax errors. After this, all variables were tested using null to extreme inputs (0 to 1 for probabilities and utilities) to ensure that obvious and predictable outputs were produced.

Our model was built on the assumption that recurrences after invasive treatments precede and account for all cancer-related deaths. As we also accounted for non-cancer deaths explicitly, combining these two causes of death allowed us to estimate overall survival. External validation was performed by comparing OS predicted by our model with OS reported in our literature review.

Using baseline cohort analysis, survival curves for each strategy of our model were generated using the Declining Exponential Approximation of Life Expectancy (DEALE) technique.⁵⁴ The resulting median OS data produced by the model were compared with published OS data from our literature review. This validation was not performed with the SC and BSC strategies as OS data from the literature were used in our modelling of those strategies.

Results

The LE model was used to predict median, 3-, and 5-year OS for RFA (33 months, 45% and 20%, respectively) and HR (37 months, 51% and 33%, respectively, for our base-case cohort of patients). These predicted values agree with ranges of published data,^{18,19,21–28,30,31,63–74} supporting external validity of our model.

Deterministic base case-analysis resulted in LE of 11.9, 23.1, 34.8 and 37.0 months, and QALE of 7.8, 13.2, 22.0 and 25.0 months for BSC, SC, RFA, and HR strategies, respectively. Based on a 70-year-old patient with no comorbidities and base-case estimations of each parameter, HR offered the most LE and QALE over a 5-year period. All incremental gains between strategies were >1 month, and thus clinically significant. HR provided incremental benefits of 2.3 months of LE and 3 months of QALE over RFA. The ranking of preferred strategies did not change between LE and QALE analyses.

Deterministic sensitivity analysis

Our model was sensitive to four parameters within plausible ranges: age at diagnosis, comorbidity, length of simulation and utility after HR (Table 2). Four additional variables, recurrence post-RFA, recurrence post-HR, utility of SC and utility of well-state post-RFA, produced threshold values close to, but outside plausible ranges.

HR was also preferred for patients up to 79 years old, after which RFA was preferred (Fig. 2). Differences in LE offered by HR versus RFA ranged from a gain of 2.3 months for 70 year olds to a loss of 3.7 months for 90 year olds.

Figure 2.

Effect of age at diagnosis and length of simulation on life expectancy. HR, hepatic resection; RFA, radiofrequency ablation; BSC, best supportive care; SC, systemic chemotherapy

HR was preferred in healthier patients. Patients with high comorbidity benefited more from RFA than HR in both LE and QALE gains. However, the benefit of RFA over HR in patients with high comorbidity was only 0.8 months of LE and 0.2 months of QALE, and thus not clinically significant. In contrast, based on a 70-year-old patient with no/mild comorbidities, HR provided an incremental benefit of 3 months of QALE over RFA over 5 years of follow-up. In addition, the ranking of preferred strategies did not change between LE and QALE analyses.

By varying the simulation length from 12 to 120 months, a threshold value was identified at 39.3 months, after which HR was preferred over RFA (Fig. 3). LE gains from RFA and HR were equivalent under 39.3 months. Beyond 50 months of simulation, LE gains from HR over RFA became significant. At 120 months, a benefit of 8.3 months from HR over RFA was predicted.

Figure 3.

Stochastic expected outcomes of life expectancy and quality-adjusted life expectancy for all four treatment strategies. HR, hepatic resection; RFA, radiofrequency ablation; BSC, best supportive care; SC, systemic chemotherapy

Although probabilities of recurrence after RFA and HR were not sensitive variables in initial sensitivity analyses using published ranges, there was a trend towards both parameters having an impact on model outcomes (Table 2). Our analysis demonstrated that at a monthly probability of recurrence after RFA of 0.0496, RFA and HR provided equivalent outcomes. This threshold value is within 4.2% of the lowest published post-RFA recurrence data. Likewise, the threshold value associated with the monthly probability of recurrence after HR was 0.0318, 1.3% greater than the highest published post-HR recurrence value.

The only sensitive utility variable was utility after HR. In spite of a calculated threshold-value of 0.60, the maximum benefit of RFA over HR was 0.5 months and thus not clinically significant. In extended sensitivity analysis allowing ranges of 0 to 1, incremental benefits of one strategy over another did not reach significance. When tested in two-way sensitivity analysis with age, utility after HR and RFA became increasingly sensitive parameters with increasing age (Fig. 4).

Figure 4.

Impact of quality of life after a hepatic resection and age at diagnosis on preferred treatment strategy. HR, hepatic resection; RFA, radiofrequency ablation

Probabilistic sensitivity analysis

Compared with deterministic analysis, while the ranking of preferred strategies did not change in probabilistic analysis, there was no significant difference in outcomes between RFA and HR. However, both RFA and HR were significantly better than SC and BSC (Fig. 4).

At age 70 years, HR and RFA were preferred in 62.7% and 15.1% of simulations when assessing LE, and 64.6% and 21.4% when assessing QALE, respectively. At age 80 years, preferences for HR and RFA were 25.8% and 46.3%, respectively, in the LE model; 32.0% and 50.5%, respectively, in the QALE model. At age 90 years, preferences for HR and RFA were 1.6% and 91.3%, respectively, in the LE model; 23.9% and 55.0%, respectively, in the QALE model (Fig. 5). SC and BSC were never the preferred strategy.

Figure 5.

Proportion of preferred treatment strategies by age at diagnosis in simulated patients. HR, hepatic resection; RFA, radiofrequency ablation

Discussion

Older patients diagnosed with CLM after a resection of primary CRC pose a difficult clinical scenario for patients and treating clinicians. While increasing evidence demonstrates that patients should not be denied aggressive treatments based on chronological age alone,^75–78 clinical judgements for older patients remain complex and must consider published evidence, physiological changes of aging, comorbidities and patient values. Traditional therapies of HR and 5-FU-based SC have improved, and are safe and effective in older patients.⁷⁹ The emergence of ablative therapies such as RFA has been associated with minimal morbidity and mortality; however, long-term effectiveness as monotherapy for CLM is unclear.⁸⁰ Finally, BSC may provide good QoL without exposing patients to toxicities of cancer treatment. In the absence of good quality trial data to guide clinical decision-making, we built a Markov transition-state DA model to estimate outcomes of BSC, SC, RFA and HR when used to treat CLM in older patients.

Overall, this study demonstrates superiority of RFA and HR over non-invasive therapies in our mathematical model analyses. However, patient age was crucial to selecting the optimal treatment approach. Specifically, HR was significantly better than RFA for a 70-year-old patient, providing gains of 2.3 months of LE and 3.0 months of QALE. However, the 2.3-month advantage of HR over RFA at age 70 years became a 3.7-month disadvantage at age 90 years. This 6-month change in incremental LE suggests the significant impact of age as a key factor in decision-making, representing a contradiction to the notion that, in the absence of comorbidity differences, patients should receive identical treatments regardless of age.

In our sensitivity analyses, it is worthy to note that BSC and SC were never the preferred options. For RFA versus HR, there was an increasing preference for RFA with older age, especially in the LE model versus the QALE model (Fig. 5).

Furthermore, our decision model focuses on disentangling the effects of age and comorbidity on short- and long-term outcomes. With respect to age, our key assumption is that, independent of comorbidity, increasing age (within the range of 70–90 years) is not associated with increased short-term treatment-related complications. Instead, these complications are a function of technique, operator and comorbidity. Age increases the risk of long-term death from non-cancer causes, which we have modelled explicitly using life tables (Table 1). Increasing comorbidity is assumed to be associated with increased short-term treatment-related complications as well as long-term death from non-cancer causes. In addition, high comorbidity is associated with avoidance of systemic chemotherapy.

Compared with HR, RFA remains superior in terms of safety, as evidenced by the minimal procedural mortality observed in studies.⁶⁷ Although HR still provided better efficacy, the combined safety and efficacy of RFA has made it an attractive alternative to HR in our model (in spite of a higher recurrence risk). Additionally, our results suggest that while HR is preferred over RFA in healthy older patients, these benefits decrease with increasing age and comorbidities, which is suggested by a more gradual preference for RFA with increasing age and comorbidity. It is conceivable that the reduced recurrence rates associated with HR do not result in significant long-term benefits in these oldest patients because of competing mortality risks from non-cancer causes. Similar impacts of comorbidities on treatment have been demonstrated in other studies.^{51–53,81,82} Furthermore, as independent variables, utilities do not significantly alter strategy selection. Nonetheless, when we analysed the interaction of utility variables with age, the resulting trends indicate that utility estimates (i.e. patient preferences) become more important for older patients. Practically, what this suggests is that QoL differences between treatment options generally do not alter decision-making in this disease. However, clinicians managing older patients should be aware of patients’ values and expectations, particularly among the very elderly. Our finding that significant survival differences between RFA and HR only become evident after 50 months of simulation may have implications on research. Future comparative studies of RFA and HR efficacy should ensure sufficient follow-up times of >4 years to show meaningful results, as shorter comparative trials with insufficient follow-up may unfairly favour RFA over HR.

This study has several limitations. First, there is a paucity of high-quality literature regarding treatment of CLM in older patients. Although randomized control trials represent the highest level of evidence, very few have been performed comparing the various therapies for CLM in older patients, making it a limitation for such a decision analysis. While careful selection and summary of all available evidence was performed, for these reasons, the inclusion of non-randomized trials was an appropriate strategy to extend the source of evidence. Aware of potential data inaccuracy, we performed extensive sensitivity analyses around variable estimates. Second, only four mutually exclusive strategies were included in our model. In reality, there has been increasing focus on multi-modal management of CLM; however, these approaches are case-specific and have not been systemically studied among older patients. Furthermore, rather than modelling every possible treatment combination, our objective was to evaluate the most commonly used modalities and develop a structural framework upon which future models can be built. Patients with severe comorbidity (Charlson’s scores ≥2) were excluded from receiving chemotherapy in the model. This may have impacted sensitivity analyses; however, it should not affect LE or QALE for HR and RFA. Additionally, SC was never a preferred strategy in our analyses. Utility data were derived from one study⁶² where all patients underwent a laparotomy followed by HR, ablation or SC. This is not the clinical scenario described in this model; however, these data are a valuable starting point for linking LE with QoL in older patients. Clearly further studies of patient utilities in the CLM setting would be welcomed. Finally, on review of the literature, there are very few studies that examine outcomes after more than two attempts at repeat curative intent HR/RFA for colorectal cancer liver metastasis. This reflects that repeat attempts at curative surgery remains a novel approach at managing colorectal liver metastases and is certainly not the standard of care for recurrent disease at this point, particularly for very elderly patients. Therefore, repeat HR/RFA were not incorporated in our analysis but may need to be considered in the future as the evidence base matures.

The significance of our base-case results of HR over RFA in LE (2.3 months) and QALE models (3.0 months) must be interpreted in the correct context. Such values are often misinterpreted as minimal increases in life span.⁸³ In reality, LE and QALE represent shifts of entire survival curves of a cohort of individuals and incremental gains represent an increase in the probability of survival at any time point. Additionally, gains in LE and QALE observed in our study are similar in magnitude to OS benefits reported in SC trials for CLM.^35,36,84

DA is useful in this field owing to its ability to explicitly incorporate factors such as age, comorbidity, treatment-specific factors and patient values into a comprehensive model. It is often the lack of high-quality, randomized evidence that necessitates the use of DA as a synthesis of available evidence in a rational, quantitative structure. Using a detailed decision analytic model populated with current clinical data, our study suggests that RFA and HR are superior to BSC and SC in terms of LE and QALE in older adults. Moreover, these results highlight the importance of considering both age and comorbidity when choosing between RFA and HR. HR may provide a marginal benefit over RFA in healthy patients aged 70–79 years whereas RFA may be more appropriate for older patients with comorbidities and those aged 80 years or older. This may potentially improve patient-physician communication and informed decision-making in the clinical setting.

Conflicts of interest

None declared.