Effective use of PROs for survival prediction: Transformer-based modelling in NSCLC patients

Accurate survival prediction is a key component to patient quality of life (QoL)-centered treatment. It allows, for example, setting reasonable treatment goals or timely palliative care referral. However, clinical estimates are often too optimistic leading to lower patient QoL before death. Patient-reported outcomes (PROs) have proven to be an important survival predictor improving the overall prognostic accuracy. In this study, transformer architecture was explored to leverage PRO trajectories to improve survival accuracy and identify the most prognostic PRO symptoms for survival prediction.

We analyzed 475 (cross-validation discovery set: 380; held-out testing set: 95) early-stage non-small cell lung cancer (NSCLC) patients who underwent SBRT treatment and routinely completed the Edmonton Symptom Assessment Scale (ESAS). A transformer-based model was developed to perform longitudinal modeling of overall survival (OS), incorporating PROs collected at multiple post-treatment follow-ups, as well as clinical and demographic variables. The performance of the proposed model was compared to traditional outcome modeling approaches, including univariate and multivariate (time-varying) Cox proportional hazards regression (CoxPH) and joint probability survival modelling. The best-performing transformer model was interpreted using the SHapley Additive exPlanation (SHAP) values, and the most prognostically relevant ESAS symptoms were identified through a backward elimination procedure guided by concordance index (c-index) and area under the ROC curve (AUC).

The best-performing transformer model achieved a cross-validated c-index of 0.753 [95 % CI: 0.742-0.764] and an AUC of 0.862 [95 % CI: 0.846-0.878] on the discovery set. On the heldout test set, the model reached a c-index of 0.694 and an AUC of 0.785, evaluated at the last time point. It significantly outperformed both Cox models and the joint probability model. Model interpretation using SHAP values and backward elimination identified appetite loss, pain, overall well-being, and shortness of breath as the most prognostically relevant symptoms for survival prediction.

Transformer-based survival models that integrate longitudinal PROs significantly enhance prognostic accuracy in SBRT-treated NSCLC patients. Loss of appetite and pain emerged as the most predictive symptoms, followed by overall wellbeing and shortness of breath. These findings suggest that targeted, symptom-focused PROs tracking could streamline clinical implementation and improve survival estimation in routine oncology care.