Stereotactic ablative radiotherapy versus video-assisted lobectomy for operable stage I non-small-cell lung cancer: study protocol for an emulated target trial

Abstract

Introduction

Video-assisted thoracoscopic surgery (VATS) lobectomy is a commonly employed surgical technique for the management of operable early stage non-small cell lung cancer (NSCLC). This procedure, however, is dependent on the patient’s ability to tolerate surgery. In light of this, stereotactic ablative radiotherapy (SABR) has emerged as a viable alternative treatment strategy for patients who are inoperable or who refuse surgery. Considering the lack of randomised controlled trials and the increased risk of bias in observational cohort studies, this study protocol proposes an emulated target trial design to investigate the causal effect of SABR, in comparison to VATS, on overall survival in operable early stage NSCLC patients.

Methods and analysis

Data on NSCLC patients will be collected from routinely collected university hospital records linked with German cancer registry data. This study protocol was developed using the target trial methodology outlined by Hernan et al. The protocol establishes specific parameters for key trial components in order to mitigate bias in the analysis of observational data and to facilitate the calculation of causal estimands. The target trial design that would be emulated is a multicentre open-label two-parallel arm superiority randomised trial. Mediators and confounding variables were determined through the use of a directed acyclic graph. The statistical analysis aims to measure the per-protocol and intention to treat effect of SABR versus VATS within 3 months of diagnosis, on survival, through the difference in restricted mean survival times, using weighted non-parametric Kaplan-Meier curves.

Ethics and dissemination

The Ethics Committee of the Medical Faculty of Martin Luther University Halle-Wittenberg with an approved addendum with Dnr 2023–112 has approved this study. The study uses anonymised routinely collected hospital and cancer registry data in accordance with applicable data protection regulations. Results will be disseminated through peer-reviewed publications and presentations at scientific conferences.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• Target trial framework: The protocol applies a target trial emulation framework to align eligibility, treatment strategies and follow-up using observational data.

• Routine data over time: Routinely collected hospital data from 2016 to 2023 are used to construct the study cohort and define exposures and covariates.

• Data linkage: Linked hospital and cancer registry data enable inclusion of detailed baseline clinical covariates and complete ascertainment of vital status.

• Bias control: Cloning with inverse probability of censoring weights is used to address immortal time bias and informative censoring introduced by the emulation approach.

• Residual confounding: Residual confounding may persist due to unmeasured or imperfectly measured determinants of treatment selection, and quantitative bias assessment will examine the robustness of the estimates to such bias.

Introduction

The preferred surgical treatment for early stage non-small cell lung cancer (NSCLC) has been gradually shifting from traditional open thoracotomy to video-assisted thoracoscopic surgery (VATS) lobectomy.¹ This shift is driven by the numerous benefits of VATS, including reduced postoperative complications, shorter hospital stays, lower pain levels and the potential for improved long-term survival outcomes.2,4 VATS, however, is typically reserved for younger, fitter patients who can tolerate the significant resection involved.⁵

On the other hand, medically inoperable patients or patients who decline surgical treatment have demonstrated noticeable benefits from stereotactic ablative radiotherapy (SABR).^{6 7} SABR refers to a highly precise, non-invasive form of radiation therapy that uses advanced imaging techniques to deliver a high dose of radiation to the tumour while minimising exposure to surrounding healthy tissue.⁸ SABR has been shown to exhibit low levels of toxicity and high rates of local control reaching 98% after 3 years.^{9 10} While considering the higher levels of comorbidities and poor performance status among SABR recipients compared with patient treated with surgery, a systematic review based on 45 studies has found 2-year overall survival (OS) rates to range from 35% to 96%, with a weighted average of 70%.¹¹ These results, coupled with its non-invasive nature, have led SABR to be widely accepted as the standard of care for treating medically inoperable patients with curative intent.

Several attempts to conduct randomised controlled trials (RCTs) to directly compare SABR with VATS in treating operable early stage NSCLC patients have been unsuccessful. The UK-based SABRTooth trial, for example, was found to be infeasible, while slow accrual hindered the progress of the STARS and ROSEL trials (STARS: NCT00840749; ROSEL: NCT00687986).¹² Despite these challenges, a pooled analysis of the 58 patients recruited from the STARS and ROSEL trials was conducted.¹³ The results published in 2016, suggested that SABR was better tolerated and may lead to better OS than surgery for operable stage I NSCLC patients. These findings incited significant debate in the lung cancer community and the consensus was that larger RCTs were required to validate these results. In 2021, Chang et al published the findings from the revised STARS trial.¹⁴ The study used propensity scores to match 80 newly recruited SABR patients with a prospective surgical cohort of 352 patients who underwent VATS L-MLND (Video-assisted thoracoscopic surgical lobectomy with mediastinal lymph node dissection) during the period of enrolment. The study highlighted the low toxicity rates reported in the SABR group and reported an OS rate of 87% after 5 years for patients receiving SABR versus 84% for those receiving VATS L-MLND.

These findings, however, contradict results reported by two major meta-analysis studies. In 2016, Deng et al reported the pooled results from 12 cohort studies that included 13 598 patients.¹⁵ The meta-analysis found operable stage I NSCLC patients who received SABR to have a significantly shorter OS compared with patients receiving lobectomy (HR = 1.68, 95% CI 1.09 to 2.06). The authors concluded that while lobectomy is still the recommended treatment for early stage NSCLC, large well-designed RCTs were still needed. In 2018, Chen et al confirmed those results using data from 16 propensity score-based studies.¹⁶ The authors found that in operable early stage NSCLC patients, all-cause mortality was higher for SABR patients compared with surgery, while lung cancer-specific mortality did not differ significantly between the two groups.

The validity of a meta-analysis, however, relies on the validity of the underlying studies. Consequently, these meta-analyses were limited by the quality of the retrospective data involved and, even with propensity score matching, selection bias and other significant factors (eg, specific comorbidity, smoking history and socioeconomic factors) were not always accounted for.

Given the lack of RCTs available and the risk of biases in observational studies, we propose an emulated target trial design to estimate the causal effect of SABR on survival in comparison to VATS. Target trial emulation, as described by Hernán and Robins, applies the principles of randomised trial design to the analysis of observational data to mitigate bias and facilitate the calculation of causal estimands.¹⁷ This approach has been used in several studies to improve causal effect estimates by avoiding common biases.^{18 19} Additionally, as clinicians are familiar with the design principles of RCTs, this approach makes it easier to interpret and communicate the results.

Methods and analysis

The target trial that we plan to emulate in this study is a multicentre open-label two-parallel arm superiority randomised trial. The intervention of interest is SABR within 3 months of a stage IA NSCLC diagnosis. The control group is the standard surgical procedure of VATS within the same time period after diagnosis. The primary outcome will be OS 1 year and 5 years after diagnosis. The research question is outlined using the PICO (Population, Intervention, Comparison and Outcome) structure (table 1). The components of the hypothetical target trial and how we emulate them in our real-world study are summarised in table 2.

Table 1. Key components of the research question addressed in the protocol.

Component	Description
Population	NSCLC patients 18 years old and older diagnosed at stage IA
Intervention	SABR radiotherapy treatment within 3 months after diagnosis
Comparator	Video-assisted lobectomy within 3 months after diagnosis
Outcomes	All-cause mortality (after 1 and 5 years)

NSCLC, non-small cell lung cancer; SABR, stereotactic ablative radiotherapy.

Table 2. Specification and emulation of a target trial of stereotactic ablative radiotherapy (SABR) versus video-assisted thoracoscopic surgical lobectomy (VATS) surgery within 3 months of diagnosis, lung cancer patients stage IA diagnosed during 2016–2023 in Germany.

Component	Target trial	Emulated trial using real-world data
Design	Multicentre open-label two-parallel arm superiority randomised trial
Aim	Estimate the effect of receiving SABR or VATS within 3 months of a non-small cell lung cancer (NSCLC) diagnosis on 1-year and 5-year overall and cause-specific survival	Same
Eligibility	STARS and ROSEL trial criteria 13	See eligibility criteria in text
Exclusions	STARS and ROSEL trial criteria 13	See eligibility criteria in text
Treatment strategies	1. SABR within 3 months of diagnosis 2. VATS surgery within 3 months of diagnosis	Same
Treatment assignment	Patients are randomly assigned to either strategy	Patients are non-randomly assigned to a treatment strategy. Randomisation is emulated via cloning of patients in both arms.
Treatment implementation	None	3 months grace period
Outcome	Death from all causes and cause-specific death within a year and within 5 years of diagnosis	Same
Type of outcome	Failure time	Same
Follow-up	Follow-up starts at diagnosis, equivalent to treatment assignment	Follow-up starts at diagnosis, which does not correspond to treatment assignment
Censoring	Loss to follow-up, administrative censoring	Loss to follow-up, administrative censoring
Adjustment variables	Age at diagnosis (≥18 years), sex (1,2), (ECOG 0,1,2), stage at diagnosis (1A), comorbidities, heart and lung function, hospital volume and socioeconomic deprivation.	Same
Causal contrast	Per-protocol	Intention-to-treat analysis: Patients that initiated SABR treatment compared with patients that initiated and completed VATS within 3 months after time zero. Per-protocol analysis: Patients who completed the full course of SABR compared with patients who received VATS after being initiated within 3 months after time zero.
Estimands	Differences in 1-year and 5-year survival and restricted mean survival time at 1 year and at 5 years between arms	Same

Data collection

Necessary data for our emulated target trial design will be determined through a non-parametric graphical model, also known as a DAG (directed acyclic graph). Using this graphical tool, we will be able to visualise the causal relationship between the exposure (SABR vs VATS) and the outcome (survival), by identifying potential confounders and mediators of the causal effect to be measured. The DAG was created through several discussions among our team of experienced radiation oncologists, surgeons, epidemiologists and biostatisticians (Figure 1). The associations demonstrated in our graph were also based on published literature, current treatment guidelines and previous clinical studies like the STARS and ROSEL trials.

Figure 1. A directed acyclic graph that illustrates the causal relationship between the treatment decision (SABR vs VATS) and the outcome (overall survival). EBUS, endobronchial ultrasound bronchoscopy; SABR, stereotactic ablative radiotherapy; VATS, video-assisted thoracoscopic surgery.

According to our DAG, patient characteristics such as age, comorbidities, lung function, performance status, hospital volume and the level of socioeconomic deprivation at the time of diagnosis were all factors that affected both the treatment to be received by the patient and their chances of survival. Therefore, these factors were regarded as our baseline confounders.

Hospital volume was considered a confounder due to the influence of VATS’s steep learning curve on the quality of treatment being delivered. Studies have reported that up to 60 operations are necessary for optimal performance of VATS resections.^{20 21} VATS performed in low-volume hospitals were also found to be associated with statistically significant postoperative morbidity and 90-day mortality.^{22 23}

On the other hand, chances of receiving adjuvant treatment or experiencing postoperative toxicities are influenced by the treatment strategy the patient has received. Surgery in stage I–II NSCLC has a major advantage, over SABR, in its ability to invasively stage lymph nodes, thereby allowing adjuvant chemotherapy to be administered if nodal metastasis was discovered.²⁴ In comparison, SABR-treated patients often have to rely on additional diagnostic tools, such as Endobronchial ultrasound bronchoscopy (EBUS), to evaluate the level of nodal involvement after an initial clinical assessment. Lacking the accessibility of surgery, a false-negative EBUS might eventually lead a proportion of SABR-treated patients to continue having residual disease, and therefore, experiencing more frequent regional recurrences. A Positron Emission Tomography / Computed Tomography (PET/CT) scan performed prior to SABR, however, has been shown to mitigate that risk.²⁵ Another point of concern is postoperative toxicities. A systematic review, published in 2008, found that 16% of patients receiving VATS reported complications.²⁶ In 2012, Paul et al reported an even higher proportion (40%) being affected by postoperative complications.²⁷ In comparison, SABR was found to have a 0.7% procedure-related cumulative mortality.²⁸ These two factors, toxicities and adjuvant therapy, were therefore considered mediators (intermediators between the exposure and outcome) given their effect on cancer survival.

After identifying potential confounders and mediators, we used the online software, DAGitty (V.3.0), to identify the minimal sufficient adjustment set to estimate the total effect of SABR on OS.²⁹ The adjustment set included: age, comorbidities (measured by the Charlson Comorbidity Index), lung function, performance status, hospital volume and the socioeconomic deprivation level.

Data source

In compliance with the Cancer Screening and Registry Act (Krebsfrüherkennungs- und -registergesetz), retrospective data on patients with NSCLC will be assembled from routinely collected hospital data from participating centres, including the University Hospital of Magdeburg, University Hospital of Halle (Saale), and University Hospital of Freiburg, supplemented by population-based clinical cancer registry data from the German federal states of North Rhine-Westphalia (approximately 18 million inhabitants), Saxony-Anhalt (approximately 2.2 million inhabitants) and Baden-Württemberg (approximately 11 million inhabitants). Additional hospitals in these states are currently being invited to participate.

Participating university hospitals will submit patient-level data requests to their respective cancer registries in order to receive additional information on patients treated at their centres. This includes data on disease progression reported outside the treating hospital and long-term outcomes such as vital status and cause of death, which are typically not captured by the hospitals. Access to this registry data is permitted under regional cancer registry law, provided that at least one case report for the patient has been previously submitted by the hospital. If no case report exists, registry data for that patient cannot be retrieved.

After linkage, these data will be merged to form a single anonymised patient-level dataset per centre, making it unlikely to extract any identifying information from the dataset.

The anonymised dataset provided by each hospital will include detailed demographic and clinical patient characteristics such as age, sex, comorbidities, continuous measures of heart and lung function measures, Eastern Cooperative Oncology Group (ECOG) or Karnofsky performance status, diagnostic tests used for staging (eg, PET/CT, EBUS), and information on the treatment decision-making process. Information on comorbidities will be collected with sufficient clinical detail to identify individual conditions for eligibility assessment and to calculate the Charlson Comorbidity Index for use in the statistical adjustment. Performance status will be harmonised across centres by applying standardised conversion between Karnofsky and ECOG scales where needed, and the distribution of performance categories will be reviewed by centre to identify and address potential inconsistencies. Information on tumour characteristics at diagnosis (date of diagnosis, Tumor-Node-Metastasis (TNM) stage, grading, topography, morphology) and the treatment received (date of treatment, duration of treatment in case of radiotherapy, procedure performed, reported complications, etc) will also be included, in addition to data from the cancer registries on vital status and cause of death.

The complete list of variables to be collected is available in online supplemental file 1. The study is planned to be conducted between 2026 and 2029, including data acquisition, harmonisation, analysis and dissemination.

Study population and eligibility criteria

The planned inclusion and exclusion criteria are as follows:

Inclusion criteria:

• Age 18 years or more (male and female patients).

• NSCLC determined histologically.

• Stage IA diagnosis defined by any combination of T1a, N0, M0, T1b, N0, M0 and T1c, N0, M0.

• PET/CT scan is required to confirm staging and nodal involvement.

• Performance score of Karnofsky ≥60% or ECOG score ≤2 before any treatment.

Exclusion criteria:

• Direct evidence of regional or distant metastases.

• Synchronous primary or prior malignancy in the past 3 years other than non-melanomatous skin cancer or in situ cancer.

• Previous lung or mediastinal radiotherapy.

• Reduced lung function determined by forced expiratory volume in 1 s (FEV1)<40% predicted, and/or diffusing capacity of the lung for carbon monoxide (DLCO)<40% predicted.

• Absence of severe chronic heart disease, severe cardiac or peripheral vascular disease and pulmonary hypertension

• Major surgery within the past 1 year.

• Serious medical comorbidities or other contraindications to SABR or VATS.

Briefly, the study population will include non-small cell cancer patients aged 18 years or more, diagnosed at stage IA from the linked dataset, between 2016 and 2023, with no record of previous malignancy in the past 3 years. In emulating a target trial, the eligibility criteria should adhere to the positivity assumption, ensuring that every patient has a non-zero probability of receiving either treatment option. This means that all included patients have to be eligible for both treatment strategies, SABR and VATS, before treatment assignment.

Diagnosis of Lung cancer will be based on the tenth edition of the International Classification of Diseases (ICD-10: C34) that is used by the cancer registry. Similarly, the non-small cell variant of LC will be determined by the ICD-O-3 morphology codes corresponding to the following subtypes (squamous cell carcinoma, adenocarcinoma, large cell carcinoma, unspecified and other specified). Stage IA will be defined through the following combinations of clinical TNM staging: T1a, N0, M0, T1b, N0, M0 or T1c, N0, M0. Confirmation that a PET/CT scan was performed will be required to confirm stage and nodal involvement in all patients. This information will be provided by the participating clinics.

To ensure that patients included in the study are medically eligible for both treatment options and thus comparable, eligibility will be based on detailed hospital clinical records provided by the participating centres. Only patients with a performance score of Karnofsky ≥60% (or ECOG ≤2) will be included. In addition, patients with documented serious medical comorbidities, significantly reduced lung function (FEV1 <40% predicted and/or DLCO <40% predicted), or other contraindications to either SABR or surgery will be excluded. Patients with a previous history of major thoracic surgery or prior radiotherapy to the lung or mediastinum will also be excluded. These criteria aim to emulate a clinical setting in which both treatment strategies would have been considered viable options.

For all included patients, the Charlson Comorbidity Index will be calculated and incorporated into the statistical adjustment to account for overall comorbidity burden.

Treatment strategies and assignment procedures

In our two-arm open label study design, the intervention arm will consist of SABR treated patients whereas patients receiving VATS will be in the control group. All individuals will enter the study at the time of diagnosis (time 0). As the date of treatment initiation does not usually coincide with the start of follow-up (time 0), we will allow a 3-month ‘grace period’ for treatment to begin to include individuals in both arms. This grace period will help us account for potential immortal time bias arising from the possible unequal waiting periods experienced by patients receiving SABR vs surgery.

VATS-treated patients will be identified through the German Operation and Procedure (OPS) codes (OPS code: 5–324.6 ff). SABR, on the other hand, will be defined using the ‘single dose’, ‘total dose’ and ‘date of radiotherapy’ variables. Patients receiving radiation doses of 5 or more Gy (Grays) in 10 or fewer fractions will be included in our intervention arm. Patients who did not begin SABR or have not yet undergone VATS within the first 3 months after diagnosis will not be considered in our analysis.

Randomisation will be emulated by cloning two exact copies of each patient with one clone allocated to each study arm, hence doubling the size of our dataset. Cloning the patients ensures that the study arms are identical at baseline, with regard to recorded demographics and clinical characteristics at the time of diagnosis. Therefore, a patient who, according to the hospital records data, has received surgery will be included in both arms, the intervention and control. The same procedure applies to patients who received SABR; they will be included in both arms as well. This cloning procedure ensures that baseline characteristics are defined consistently across treatment strategies at the start of follow-up.

Outcomes and follow-up period

For a target trial emulation to be successful, it is crucial to accurately define the baseline, or time zero, when follow-up begins in the observational data. This is when the eligibility criteria are met, and from which point study outcomes are measured. In our target trial, follow-up naturally starts when a treatment strategy is decided on, typically aligning with or just before treatment begins. Properly aligning eligibility verification, treatment assignment and follow-up commencement is essential to avoid flawed conclusions. Our 3-month grace period provides the necessary leeway to accommodate delays in treatment start, ensuring uniform initiation times across all study arms and preserving the validity of our comparisons.

Our primary outcome is OS after 1 and 5 years. OS will be calculated until the date of death from any cause or until the end of follow-up, whichever comes first. Vital status is ascertained using death certificates and information from the registration offices. Patients lost to follow-up before death or still alive at the last vital status assessment will be right-censored at the date of the last vital status assessment or at the censor date, whichever comes first. Secondary outcome measures will include cancer-specific survival and recurrence-free survival. Cancer-specific survival will be calculated up to the date of lung cancer-associated death, which in this case will be defined with the ICD-10 codes (C33–C34) under the ‘cause of death’ variable recorded by the cancer registry data. Recurrence-free survival will be calculated up to the date of first recurrence (local, regional or distant) or death, whichever occurred first.

Statistical analysis plan

Demographic and clinical characteristics according to study arm will be described using common descriptive statistics. A flow chart illustrating the number of individuals assigned to each treatment arm, those who follow the protocol, and those who are censored or excluded and reasons for exclusion will be included in our final analysis.

The effect of SABR versus VATS within 3 months of diagnosis on survival will be measured through the differences between the study arms in: (1) 1-year and 5-year survival probabilities; and (2) restricted mean survival times (RMSTs over a 1-year and 5-year window). Differences in RMST will reflect the average gain or loss in survival time associated with one treatment compared with the other, allowing separate evaluation of shorter-term (1 year) and longer-term (5 years) treatment effects.

After cloning our sample at baseline, we will then proceed to censor a clone when the treatment actually received by the patient is no longer compatible with the treatment strategy of the arm they entered. For example, a patient who, according to the registry data, received surgery will be censored at the time of surgery in the SABR treatment arm (where they are included as a clone). This will also apply for patients receiving SABR.

This artificial censoring induces informative censoring (ie, selection bias over time), as described by Hernan, since treatment received typically depends on individual characteristics and is not random. To account for this introduced selection bias, we use the inverse-probability-of-censoring weights (IPCWs) method where uncensored observations are up-weighted to represent censored observations with similar characteristics and thus to allow the unbiased estimation of the causal effect of interest. A standard approach to estimate the weights is to predict the individual probabilities of remaining uncensored at each time of event using a Cox regression model.

For the per-protocol analysis, we will focus exclusively on patients who completed the full course of SABR treatment, providing a more precise measure of its efficacy. This analysis will exclude patients who initiated but did not complete SABR or switched to surgery, thereby isolating the effect of complete adherence to the SABR protocol.

In parallel, we will conduct an analysis that approximates an intention-to-treat approach, including all individuals who initiated SABR within the 3-month grace period, regardless of subsequent treatment changes. This will offer an observational counterpart to the intention-to-treat principle used in RCTs. While cloning combined with IPCWs helps reduce bias due to informative censoring and treatment selection, this approach does not eliminate all confounding, particularly if relevant factors are unmeasured or imperfectly recorded. Therefore, sensitivity analyses will be further conducted to evaluate the robustness of the findings.

Survival curves will be estimated in each arm using a weighted non-parametric Kaplan-Meier estimator. We will calculate the 95% CIs for the difference in 1-year and 5-year survival and difference in RMSTs using non-parametric bootstrap with 1000 replicates.

All analyses will be performed using R.³⁰

Missing data

To handle missing data in our proposed analysis, we will use a multiple imputation fully conditional approach for all confounders with missing values.³¹ This method involves specifying conditional regression models for each missing value in a variable, based on the values of the other variables in the imputation model. We will follow this approach to minimise any potential bias that may arise due to missing information.

Subgroup and sensitivity analysis

While VATs are considered the gold standard treatment for stage IA, we will also be interested in observing the proportion of patients receiving other surgical procedures for stage IA NSCLC treatment. Previous studies have shown that less invasive procedures such as segmentectomy and wedge resection are better tolerated than lobectomy and may improve survival.³² As a subgroup analysis, we could compare OS of patients receiving lobectomy, wedge resection and segmentectomy to our intervention of interest, SABR.

Sensitivity analyses will be conducted to assess the effects of decisions made during data cleaning, variable transformation and key design assumptions. In our emulated target trial, follow-up is defined to start at diagnosis to provide a common time zero across treatment strategies and to reflect the point at which treatment planning begins. However, unrecorded clinical events occurring between diagnosis and treatment initiation, such as deterioration in performance status, may influence the treatment ultimately received. To evaluate the robustness of the findings to this timing structure, we will repeat the analyses using alternative grace period definitions, including shorter (1 month) and longer (6 months) windows. In addition, we will also assess the influence of comorbidity burden and patient fitness on study findings. Specifically, analyses will be repeated after excluding patients with higher Charlson Comorbidity Index scores, and the impact of alternative definitions of functional operability will be evaluated.

Finally, a quantitative bias analysis will be conducted to assess how potential unmeasured or imperfectly measured confounders could influence the estimated treatment effect.³³

Ethics and dissemination

The Ethics Committee of the Medical Faculty of Martin Luther University Halle-Wittenberg with an approved addendum with Dnr 2023–112 has approved this study. The study uses anonymised routinely collected hospital and cancer registry data in accordance with applicable data protection regulations. Results will be disseminated through peer-reviewed publications and presentations at scientific conferences.

Data availability statement

No data are available.