Session Item

There is a need for quantitative biomarkers to guide treatment decisions in patients with lung cancer treated with radiotherapy. Decisions on treatment intent consider variables, such as age, stage and performance status, but predicting early mortality is challenging. 90-day mortality is used in surgical decision making, but no models have been developed for radiotherapy.  In this work, we investigated muscle characteristics as an image-based biomarker to predict early mortality.

586 patients with NSCLC lung cancer treated with 55Gy in 20 fractions were identified from a retrospective database. The muscle compartment at T12 was segmented on planning CT using an in-house AI model. Muscle area and density were extracted for all patients. Early mortality was defined as death within 90-days of start of treatment. A logistic regression model predicting early mortality was built including known prognostic variables and variables that impact muscle quality; age, sex, performance status, T-stage, and N-stage. Models with muscle area, muscle density and performance status were individually built. 5-fold cross-validation was used to assess model accuracy and Akaike Information Criteria (AIC) ranked model performance.

39 segmentations failed leaving 547 patients for analysis. Patient characteristics are included in table 1. Median muscle area of 40.4cm³ (11-110cm³) and median density 12.7HU (-7.4-34.2HU) were found, indicating poor muscle health across the population. 64 (11.7%) patients died within 90-days of treatment, higher than reported in surgical series. These patients had significantly lower muscle density (Mann-Whitney, p=0.03) but not muscle area (Mann-Whitney, p=0.28). Individual logistic regression models showed significance of muscle density (p=0.02) and muscle area (p=0.003), reduction in muscle was associated with early mortality. Male sex was significant in both models, accounting for age, T-stage, N-stage (all non-significant), table 2.  Male sex showed a significant difference in muscle characteristics (Mann-Whitney, density p=0.04 and area p=0.004) indicating this effect is driven by gender (figure 1a).

5-fold cross-validation showed model accuracy with muscle density of 0.87 and for the model with muscle area an accuracy of 0.88. The AIC for the logistic regression including muscle area was significantly improved compared to the model including density. Finally, the model with performance status showed no significant association with ninety-day mortality.

We have shown that muscle characteristics, collected from routine imaging, are a potential tool for predicting 90-day mortality for patients with lung cancer treated with radiotherapy. Model performance showed good accuracy with muscle area or density, with males showing greater impact of poor muscle condition on 90-day mortality. Muscle characteristics showed superior predictive performance over performance status. Further internal and external validation will be performed.