Development and internal validation of a multinomial NTCP model for the severity of acute dyspnea after radiotherapy for lung cancer

Background and purpose: Dyspnea evolution after radiotherapy for lung cancer is complex with potential symptom deterioration and improvement from baseline. We developed and internally validated a multinomial normal tissue complication probability (NTCP) model predicting dyspnea grade. Materials and methods: Patient-reported dyspnea was collected pre-treatment and during 6 months follow-up for 182 stage I–IV lung cancer patients treated with radical (chemo)radiotherapy. Dyspnea changes (ΔDys) from the baseline grade (Dys ₀ ) to the follow-up grade (Dys) were evaluated. A multinomial logistic regression model simultaneously predicting 3 grades of Dys (Dys ≥ 3, Dys = 2 and Dys ≤ 1 (reference level)) was optimized. Reference NTCP models predicting Dys ≥ 2 and Dys ≥ 3 risks irrespective of Dys ₀ were generated for comparison. Models were shrunken and performance was assessed using optimism-corrected AUC (bootstrapping). Results: Rates of ΔDys ≥ 1 (deterioration) and ΔDys ≤ -1 (improvement) at 6 months were 31.9% and 12.6%. Dys ≥ 3, Dys = 2 and Dys ≤ 1 rates were 13.7%, 20.9% and 65.4%, respectively. The multinomial model (combining the risk factors Dys ₀ and MLD and the protective factor chemotherapy treatment) predicted Dys ≥ 3, Dys = 2 and Dys ≤ 1 with AUC (95% CI) of 0.72 (0.65–0.75) 0.76 (0.72–0.79) and 0.78 (0.74–0.80), respectively. Reference Dys ≥ 2 and Dys ≥ 3 models showed worse AUC: 0.64 (0.59–0.67) and 0.66 (0.50–0.70), respectively. Conclusions: Dyspnea grade could be predicted with high accuracy using a multinomial NTCP model, yielding personalized dyspnea symptom improvement and deterioration risks.