New Protocol Aims to Predict Serious Treatment Side Effects in Advanced Lung Cancer

Researchers in the United Kingdom have published a detailed protocol for a project called Lung Cancer Improved Decisions, or LUCID, which sets out how they plan to build and internally validate a risk score estimating the likelihood that a patient with late-stage lung cancer will suffer a serious adverse event after receiving a particular systemic anti-cancer therapy. The protocol, published in Diagnostic and Prognostic Research, describes a planned analysis of linked national health data covering more than 20,000 patients in England. Crucially, this is a study protocol rather than a report of finished results: no risk score has yet been developed, and the document lays out methods that will be carried out in future work.

The clinical problem the project addresses is stark. According to the protocol’s background section, less than 20 percent of patients diagnosed with advanced lung cancer survive beyond five years, and around half of those treated with systemic anticancer therapy experience a serious adverse event caused by the treatment that results in a hospital attendance. Serious adverse events are defined in the protocol as those requiring a hospital attendance. With several classes of therapy now available, including chemotherapy, immunotherapy, targeted therapy and combinations of these, the authors argue that a score predicting the risk of a serious event under each treatment option could improve communication with patients and support shared decision making. They note that no risk scores are currently available for use in patients with advanced stage lung cancer.

The scale of the affected population underlines why the team considers the work important. Lung cancer is the second most common cancer in both men and women in the UK and has the highest cancer mortality, with 35,000 patients a year dying of the disease. Only about 16 percent of patients diagnosed with lung cancer survive more than five years, a figure the authors attribute partly to the age and smoking history of the patient population and partly to late-stage diagnosis, with 21 percent of patients diagnosed at stage 3 and 50 percent at stage 4.

What distinguishes LUCID from a conventional prognostic model is its planned use of counterfactual prediction. The protocol explains that most clinical prediction models are what the authors call pure prediction models: they estimate risk under current or observed conditions, assuming that the relationship between variables and outcomes does not change. When such models are used to inform treatment choices, they can fail, in part because of poor reporting standards but mainly because they are applied to populations under different conditions from those they were trained on. Treatment choice itself changes the distribution of characteristics in the population, and current treatment allocation reflects confounding by indication, meaning that ignoring it yields spurious associations in the model.

Counterfactual prediction models take a different approach grounded in causal inference theory. The treatment choice is explicitly incorporated into the model, and the goal is to estimate what a patient’s risk would have been under each possible treatment scenario, whether or not that treatment was actually received. Such models can in principle answer questions of the form: what would the outcome have been if this patient had received Treatment A instead of Treatment B? The protocol notes that these models rely on assumptions including conditional exchangeability, meaning no unobserved confounding, consistency, and positivity. Because standard performance assessment tools are not valid for counterfactual prediction, the team plans to apply counterfactual measures of calibration, discrimination and overall prediction error for validation of predictions under interventions, as described by Boyer and colleagues.

The planned study population consists of patients aged 18 and over diagnosed with advanced non-small-cell lung cancer without actionable genomic alterations between 2016 and 2019, a four-year period. The exposure is first exposure to systemic anti-cancer therapy, with four groups of regimens compared against each other: regimen A, combination chemotherapy alone, including platinum, gemcitabine, taxanes and other cytotoxic combinations; regimen B, combination chemotherapy plus a PD-(L)1 checkpoint inhibitor; regimen C, immunotherapy only; and regimen D, single agent carboplatin. The primary outcome is any unplanned or emergency hospital admission occurring within 30 days of the first systemic anti-cancer therapy administration, a binary outcome. A secondary research question will use descriptive Kaplan-Meier plots to compare death within two years of the first treatment dose across the four regimen groups, a timeframe chosen because of the poor prognosis of this population.

The data will come from linked national sources in England: the Systemic Anti Cancer Therapy registry, which records all patients receiving such therapy across NHS England trusts along with diagnosis and treatment regimens; Hospital Episode Statistics covering admitted patient care, emergency department attendances and outpatient attendances; Office for National Statistics mortality data; and the National Cancer Registration and Analysis Service. Time zero for the analysis is the date of the first treatment dose, and eligibility will be assessed prior to the decision-to-treat appointment, defined as the appointment immediately before the first dose. Follow-up for the primary outcome ends at an event, death or the end of the study period, and patients who experience the primary outcome will still contribute to the survival analysis for death.

The statistical plan is set out in detail. The team will initially fit a multivariable logistic regression model of the outcome on the exposure and covariates, using the whole dataset for development, with internal validation by K-fold cross validation, where the number of folds depends on the effective sample size, defined as five times the number of outcomes. They will assess whether a penalised or non-penalised logistic regression is preferable based on the magnitude of the penalty parameter lambda. Variables will be grouped into a set L of confounders, some of which are also predictors, and a set Z of predictors of the outcome that are not related to treatment. Missing data will be handled where necessary with multiple imputation combined using Rubin’s rule, with comparisons to complete records analysis and simpler practical methods such as missingness indicators and median imputation. Model performance will be summarised using calibration and discrimination, and heterogeneity will be assessed, with fine-tuning and interrogation used to evaluate differences in performance between hospitals. Clinical utility will be assessed by calculating the net benefit in preventing serious adverse events.

Sample size considerations reflect the novelty of the approach. The authors state there are no previous studies in which a counterfactual prediction model was developed for lung cancer for any outcome, making conventional sample size calculation difficult. They aim for at least 10 variables per outcome event and also applied the method of Riley and colleagues to calculate the mean absolute prediction error achievable with the available data. Using a development dataset of 24,000 patients, an outcome proportion of 0.34 and 14 candidate predictor parameters, the calculation yields a mean absolute prediction error of 0.007, well below the recommended cut-off of 0.05, suggesting the dataset is large enough for the planned model.

The authors frame the potential benefits carefully as aims rather than demonstrated effects. If the counterfactual risk score performs as hoped, clinicians could more accurately inform patients of the serious-event risk each treatment would carry from a causal perspective, multidisciplinary teams would gain a shared understanding of that risk to aid shared decision making, and hospitalisation rates due to serious adverse events in this group may be reduced, which is particularly relevant to minimising patient suffering in the last years of life. The protocol was developed following the HARPER reporting template and aligned with the American Joint Committee on Cancer’s checklist for outcome prediction models. The team describes the planned tool as the first decision aid for advanced non-small-cell lung cancer patients using a counterfactual approach and states that a future study is planned for external validation, since the current work covers development and internal validation only. Until that work is complete and the score is tested in practice, its value for patients remains a hypothesis to be tested rather than an established clinical benefit.

Subject of Research: A protocol for developing and validating a counterfactual risk score to predict serious adverse events of systemic anti-cancer therapy in advanced lung cancer.

Article Title: Protocol for the development and validation of a risk score to predict risk of adverse events associated with systemic anti-cancer treatment in late-stage lung cancer: Lung Cancer Improved Decisions (LUCID)

Article References: Almossawi, O., Steventon, L., Keogh, R., Diaz-Ordaz, K., Wang, Z., Challenger, A., Dodwell, D., Forster, M., Man, K. K. C., Wei, L., Masento, S., Januszewski, A., & Chambers, P. (2026). Protocol for the development and validation of a risk score to predict risk of adverse events associated with systemic anti-cancer treatment in late-stage lung cancer: Lung Cancer Improved Decisions (LUCID). Diagnostic and Prognostic Research, 10(1), Article 10. https://doi.org/10.1186/s41512-026-00225-y

Image Credits: AI Generated

DOI: 10.1186/s41512-026-00225-y

Keywords: lung cancer, risk score, prediction model, systemic anti-cancer therapy, counterfactual prediction, study protocol, serious adverse events, internal validation, NHS England, non-small-cell lung cancer, shared decision making, logistic regression