Materials and methods

Patient data

We retrospectively identified 451 consecutive patients with stage I–IV non-small cell lung cancer (NSCLC) treated with radical-intent radiotherapy at University College London Hospital between 2015 and 2023. The study received ethics approval in March 2022 and was registered in the UK Integrated Research Application System under project ID IRAS 303606. Of these, 231 patients (51%) had a pre-treatment CT and two post-treatment follow-up CTs available, yielding 2708 CT scans. Median follow-up was 22 months (range, 1–97). Follow-up CT was generally performed every 3 months during the first year and every 6 months thereafter.

Patients were treated with either stereotactic body radiotherapy (SBRT; 34 Gy in 1 fraction to 60 Gy in 8 fractions; n = 132) or conventionally fractionated definitive radiotherapy (dRT; 55 Gy in 20 fractions to 66 Gy in 33 fractions, with some patients receiving 30 Gy in 10 fractions; n = 99). Both contrast-enhanced and non-contrast scans were included. Images were reconstructed using 55 kernels, grouped as SOFT or SHARP. None were ECG- or respiratory-gated. Clinical variables included age, sex, overall tumour stage, gross tumour volume, treatment type, survival time, and survival status.

Patients could contribute to one or both analysis-specific sub-cohorts, depending on anatomical coverage and scan availability. We aimed to include at least 150 patients in each sub-cohort.

For LV myocardial analysis, we included patients with a pre-treatment CT and at least three follow-up CT scans covering the entire thorax. Multiple time points were required to reduce variability caused by cardiac and respiratory motion on ungated CT. This sub-cohort included 150 patients, each with 3–12 follow-up scans, giving 1665 scans.

For SKM and fat assessment, measurements were obtained at L1 because L3 was not consistently covered on thoracic CT. Although L3 is more commonly used, L1 is a practical alternative when abdominal coverage is limited and correlates strongly with L3-derived measures [14]. Included patients had a pre-treatment CT scan and at least two follow-up CT scans that fully captured the L1 vertebral body and transverse processes. This sub-cohort included 184 patients, each with 2–11 follow-up scans, giving 1647 scans.

Segmentation

Automated 3D segmentations were obtained using TotalSegmentator (Version 2) [7]. Structures included LV myocardium, cardiac chambers, aorta, oesophagus, pulmonary artery, L1 vertebra (body and bridge), SKM, subcutaneous fat, and torso fat. Two board-certified clinical oncologists reviewed the auto-segmented contours for a subset of 30 patients to verify their conformity with accepted clinical standards. The resulting volumes were used for longitudinal analysis. Examples are provided in Supplementary Material A.

Measurement of longitudinal change

To reduce the influence of scan-to-scan variability in ungated imaging on the estimated longitudinal velocity, patient-specific trends were derived from all available scans rather than a single scan pair, providing a more robust estimate of change over time. For each patient, measurements were normalised to the pre-treatment value and fitted against time in months using linear regression. The slope represented the monthly rate of change. The linear assumption is evaluated and supported in Supplementary Material B.

For LV myocardial mass, the rate of change was denoted vmyo. Because L1 coverage varied between scans, SKM, subcutaneous fat, and torso fat volumes were divided by the measured L1 length in centimetres before baseline normalisation. The corresponding linear slopes were denoted vskm, vsubF, and vtorsoF, respectively. Examples of the fitted longitudinal changes are provided in Supplementary Material C. The effects of reconstruction kernel and contrast-agent use on the estimated velocities are reported in Supplementary Material D.

Survival analysis

We examined the associations between longitudinal tissue-change velocities and overall survival using weighted multivariable Cox proportional hazards models implemented in lifelines [15]. The LV model included vmyo, while the body-composition model included vsubF, vtorsoF, and vmuscle. Prior evidence suggests that organ-volume-change effects on mortality are non-linear e.g. U-shaped for heart, and that departures from normal cardiac size, towards either hypertrophy or atrophy, increased mortality risk [8], [16]. Each velocity was therefore represented using B-spline basis functions [17]. Models were adjusted for age, sex, treatment type (SBRT vs dRT), and gross tumour volume (see Table S3 and Table S5). Overall stage was not included due to high correlation with treatment type(r = −0.68). Inverse-probability-of-treatment weights (IPTW) was used to reduce baseline imbalances between SBRT and dRT.

Model stability and predictive performance were assessed using 1000 bootstrap samples. In each sample, the weighted spline Cox model was refitted, and predictors with Wald (p < 0.05) were recorded. Predictors selected in more than 60% of samples were retained in the final model. The optimism-corrected concordance index was calculated by subtracting the mean bootstrap optimism from the apparent C-index.

For risk stratification, the fitted spline curves were used to identify values at which the estimated hazard ratio exceeded 1.0 in each bootstrap. The median bootstrap-derived values were used as the final cut-offs defining high-risk and low-risk/stable groups. Hazard ratios for the resulting cardiac and skeletal-muscle risk categories were reported using the stable group as the reference.

Dose deposited in organs related to longitudinal changes

Biological effective dose (BED) was calculated as.

BED=d1+dα/βα/β=3

where n is the number of fractions, d is the dose per fraction, and α/βwas set to 3 Gy. BED-based dose-volume metrics from V5Gy to V65Gy, in 5 Gy increments, were extracted for four heart chambers, left ventricular (LV) myocardium, body, aorta, oesophagus, and pulmonary artery. Vx represents the volume of the structure receiving at least x Gy BED.

For dose–response analyses, each biomarker was binarized using the final bootstrap-derived cut-off from the spline Cox analysis. The high-risk group was coded as 1, and the low risk/stable group was coded as 0. Collinearity among dose variables was first reduced by hierarchical clustering of the pairwise correlation matrix into four clusters. Age and total dose per fraction were retained as prespecified covariates. For each outcome model, feature selection was then performed using bootstrap stability selection. In each bootstrap sample, one representative predictor was selected from each feature cluster based on clinical relevance and its association with the outcome, evaluated using Firth logistic regression with Bonferroni correction. VIF-based filtering was then applied to reduce residual multicollinearity (VIF ≤ 3). The 5 predictors with the highest bootstrap selection frequencies were retained, underwent a final VIF check, and were further evaluated using L1-penalized logistic regression with balanced class weights and 5-fold cross-validation [18]. Predictors with non-zero coefficients were included in the final weighted multivariable logistic regression model. Model performance was reported as internally assessed apparent performance. The analysis code is available at https://github.com/YingClara92/BodyComposition.git.