Results

Of the 451 patients with stage I–IV NSCLC treated at UCL Hospital between 2015 and 2023, 220 were excluded because pre-treatment or at least two follow-up CT scans were unavailable, leaving 231 patients for analysis (Table 1; Fig. 1).

 | All | dRT (99) | SBRT (132) | p value between SBRT and dRT
Age [mean ± SD] | 75.0 ± 9.8 | 70.6 ± 9.2 | 78.1 ± 9.0 | <0.005
GTV volumes (cm3) [mean ± SD] | 161.6 ± 222.6 | 329.8 ± 262.0 | 48.4 ± 71.2 | <0.005

Gender |  |  |  | 
Female | 123(53.2%) | 56 (56.6%) | 67 (50.8%) | 0.46
Male | 108(46.8%) | 43 (43.4%) | 65 (49.2%)

Overall Stage |  |  |  | 
I | 122(52.8%) | 8 (8.1%) | 114 (86.3%) | <0.005
II | 33(14.3%) | 28 (28.3%) | 5 (3.8%)
III | 60(26.0%) | 55 (55.6%) | 5 (3.8%)
IV | 7(3.0%) | 4 (4.0%) | 3 (2.3%)
Unknown | 9(3.9%) | 4 (4.0%) | 5 (3.8%)

RT (Total Dose/Fractions) |  |  |  | 
60- 66Gy/30–33 | 58 (25.1%) | 58 (58.6%) |  | /
50–55/20 | 35(15.2%) | 35(35.4%) |  | /
60/15 | 2 (0.9%) | 2 (2.0%) |  | /
30/10 | 4 (1.7%) | 4 (4.0%) |  | /
60/8 | 31 (13.4%) |  | 31 (23.5%) | /
45–55/5 | 59 (25.5%) |  | 59 (44.7%) | /
30–54/3 | 38 (16.5%) |  | 38 (28.8%) | /
34/1 | 4 (1.7%) |  | 4 (3%) | /
Number of sequential imaging [median ± 95%CI] | 7 [3−13] | 8 [3−12] | 7 [3–13] | 0.38
Imaging time[median ± 95%CI] | 18[1–66] | 19[1–69] | 1[0–61] | <0.005

Fig. 1: Patient inclusion and exclusion flow chart. NSCLC = non-small cell lung cancer, LV = left ventricular.

After IPTW adjustment, tissue-change velocities differed between SBRT and dRT for subcutaneous fat, torso fat, and LV myocardium, but not for skeletal muscle; however, overall survival did not differ significantly by treatment type in either analysis-specific cohort (Supplementary material E).

We fitted a univariate IPTW Cox proportional hazards model using B-splines with vskm only because vsubF, vtorsoF, age, sex, treatment group (SBRT vs dRT), and gross tumour volume did not meet the predefined bootstrap stability criterion for inclusion in the final model. The optimism-corrected C-index was 0.70. Spline modelling improved the concordance index by approximately 7% compared with a linear model (Supplementary Material F). Additional sensitivity analyses using different cropped longitudinal follow-up windows showed that concordance was highest when more complete longitudinal data were used, supporting the added value of incorporating additional longitudinal measurements; details were provided in Supplementary material G.

The spline-based analysis identified −0.4±0.1%/month as the cut-off value for vskm across bootstrap resamples. The spline-based partial hazard plot from all data (in Fig. 2a) declined monotonically and then nearly plateaued, with only a modest upturn. We defined two groups using a median bootstrap-derived threshold: 1) SKM preserved/minor loss, including patients with increased, stable SKM, or minor SKM loss (vskm≥−0.4%/month), where the hazard was lower and stable; and 2) Severe SKM loss, including patients with significant SKM loss (vskm<−0.4%/month), which was associated with a sharp increase in hazard. Patients with SKM loss had a substantially higher risk of death than those with preserved SKM (HR 4.41, 95% CI 2.46–7.91, p < 0.005; Table 2). The corresponding KM curves differed significantly between the two groups (Fig. 2b).

Fig. 2: Partial hazard plot-based Kaplan-Meier (KM) survival analysis. (a) is the partial hazard plot vmuscle. SKM loss (vmuscle<−0.4) represents patients with significant skeletal muscle loss and SKM preserved (vmuscle≥−0.4) are those without. (b) is the KM survival curve of groups for vmuscle. (c) is the partial hazard plot vmyo. Cardiac stable is the LV myocardial mass stable group (−0.3≤vmyo≤0.3), Cardiac atrophy (vmyo≤−0.3) is the LV myocardial mass loss group, and Cardiac hypertrophy (vmyo>0.3) is the LV myocardial mass gain group. (d) is the KM survival curve of groups for vmyo.

Category | HR (vs stable) | 95% CI | p value
Cardiac Hypertrophy | 7.90 | 2.82–22.15 | <0.005
Cardiac Atrophy | 4.12 | 1.65–10.27 | <0.005
SKM Loss | 4.41 | 2.46–7.91 | <0.005

We also fitted a univariate spline Cox model for vmyo in the LV myocardium cohort because age, sex, treatment group (SBRT vs dRT), or gross tumour volume did not meet the predefined bootstrap stability criterion for inclusion in the final model. The spline model captured a U-shape hazard curve and improved predictive accuracy with optimism-corrected C-index of 0.75 (+22% vs. linear). A standard linear Cox model failed to detect any association between vmyo and OS (Supplementary material F), likely because the U-shaped hazard masked the effect, with opposing slopes that average to ∼0. The bootstrapping identified the thresholds at which HR > 1, which were − 0.3%/month (95% CI: −0.4 to −0.2) and 0.3%/month (95% CI: 0.2 to 0.5), respectively. The partial hazard curve in Fig. 2c was asymmetric and U-shaped: the hazard was lowest near stable mass, rose steeply for hypertrophy (positive vmyo) and rose more gradually for atrophy (negative vmyo).

We defined three categories: 1) cardiac stable: −0.3%/month≤vmyo≤0.3%/month, 2) cardiac atrophy: vmyo< -0.3%/month, and 3) cardiac hypertrophy: vmyo>0.3%/month. Compared with patients in the stable group, both cardiac hypertrophy (HR 7.9, 95% CI 2.82–22.15, p < 0.005) and cardiac atrophy (HR 4.12, 95% CI 1.65–10.27, p < 0.005) were associated with increased mortality (Table 2). KM analysis showed significantly poorer survival in both the atrophy and hypertrophy groups compared with the stable group (Fig. 2d). Although the log-rank comparison between the cardiac atrophy and hypertrophy groups did not reach statistical significance (p = 0.12), the KM estimates at 30 months post-treatment suggested a clinically meaningful divergence in survival probabilities (92% [95% CI: 72–98%] vs. 70% [95% CI, 38–87%]).

Higher low-dose exposure to selected thoracic organs was associated with adverse longitudinal changes in SKM and LV myocardial mass (Table 3). Oesophageal V10Gy was independently predictive of SKM loss (OR 1.40, 95% CI 1.04–1.89; p = 0.03), albeit with modest discrimination (AUC 0.63, 95% CI: 0.53–0.73). For LV myocardial atrophy, aorta V5Gy was a significant predictor (OR 1.71, 95% CI 1.08–2.70; p = 0.02), while heart myocardium V10Gy showed a weaker association that did not reach statistical significance (p = 0.09). The LV myocardial atrophy model showed discrimination of 0.71 (95% CI: 0.61–0.81). Heart right atrium V10Gy (OR 1.63, 95% CI 1.09–2.43; p = 0.02) and total dose fraction (OR 1.51, 95% CI 1.02–2.22, p = 0.04) were significantly associated with LV myocardial hypertrophy and demonstrated AUC of 0.65 (95% CI: 0.51–0.78).

Outcome | Predictor range | Odds ratio (95% CI) | p | AUC (95% CI) | Accuracy (95% CI)
Skeletal muscle loss | Oesophagus V10Gy (cm^3, mean ± SD)12.7 ± 12.6 | 1.40 (1.04–1.89) | 0.03 | 0.63 (0.53–0.73) | 0.69 (0.54–0.71
 | Heart atrium right V20Gy (cm^3, mean ± SD)6.8 ± 17.5 | 1.32 (0.98–1.79 | 0.07 |  | 
 | Age (years, mean ± SD)75.8 ± 8.8 | 1.11 (0.82–1.49) | 0.50 |  | 
Cardiac Atrophy | LV myocardium V10Gy (cm^3, mean ± SD)12.5 ± 25.5 | 1.50 (0.94–2.39) | 0.09 | 0.71 (0.61–0.81) | 0.71 (0.61–0.77)
 | Aorta V5Gy(cm^3, mean ± SD) 111.2 ± 99.2 | 1.71 (1.08–2.70) | 0.02 |  | 
 | Age (years, mean ± SD)75.8 ± 8.8 | 1.37 (0.93–2.04) | 0.11 |  | 
 | TotalDosePerFraction | 1.17 (0.79–1.74) | 0.42 |  | 
Cardiac Hypertrophy | Heart atrium right V10Gy (cm^3, mean ± SD)9.4 ± 18.7 | 1.63 (1.09–2.43) | 0.02 | 0.65 (0.51–0.78) | 0.65 (0.45–0.71
 | TotalDosePerFraction(Gy, middle [95% CI])7.5 [2–34] | 1.51 (1.02–2.22) | 0.04 |  |