Abstract 196P
Background
Radiation-induced lung diseases (RILD) such as dyspnea, cough, and fever can have a significant impact on the quality of life of a patient and, in rare situations, be fatal. Severe RILD is observed in about 15% of lung cancer patients treated with (chemo-) radiotherapy within the first 6 months after radiation. This study aims to develop a causal Bayesian network (BN) structure based on expert knowledge to predict dyspnea in lung cancer patients receiving (chemo-) radiation treatment.
Methods
Social, clinical, and dosimetric parameters of 730 NSCLC patients treated with (chemo) radiation at Maastro Clinic, Netherlands, from 2009–2014 were used to develop the structure. To develop the structure, eight lung cancer experts determined the causal relationship between the available variables independently. Only causal relationships that more than 50% of the expert agreed on were used to build the final structure. Structure performance was evaluated based on the Area Under the Curve (AUC) and calibration.
Results
The resulting structure included ‘chemotherapy’, ‘Post-RT Dyspnea’, ‘Baseline Dyspnea’, ‘Mean Lung Dose, ‘FEV’, ‘Smoking Status’, and gtv1as nodes. However, gtv1 was removed from the final structure because it had too much missing information. The structure had an AUC value of 0.86 (95% CI 0.83–0.88) with a sensitivity and specificity of 0.91 and 0.73 respectively.
Conclusions
We developed a BN structure based on experts’ opinions to predict if a NSCLC patient will have difficulty breathing after treatment. Our model accurately predicted post-radiation therapy dyspnea. However, the performance of the structure is still not optimal even after using a 50% concordance threshold between experts which suggests further research to choose the optimal threshold to improve structural performance. We also observed that the association between smoking status and baseline dyspnea was much stronger in the data than the number of experts agreeing that this relationship is important. This shows that clinical structures built by experts for clinical plausibility should be complemented by data evidence for optimal performance.
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.