Abstract 417P
Background
Improving prognostic models and directing clinical decisions, especially in limited-resource settings where chemotherapy is still the standard of care, requires an understanding of the factors impacting patient survival outcomes. We sought to assess the impact of non-omics data on patient survival using a Cox neural network that combines deep learning features with conventional Cox proportional hazards modelling.
Methods
A large-scale database comprising 951 survival cases from 1370 lung cancer patients, diagnosed between January 2013 and December 2022, was split into training and validation sets. A multilabel neural network model was trained to extract deep learning features from demographic, clinical, pathological, and therapeutic inputs. These features were combined with traditional variables to construct a comprehensive model for predicting overall survival (OS) at 6, 12, and 24 months. Feature importance was assessed, and model performance was evaluated using ROC curves.
Results
The concordance index improved from 0.68 (previously reported with traditional Cox regression) to 0.812 with the proposed model, with a mean standard error of 0.012, demonstrating strong predictive ability. Significant leads identified: tobacco status, cardiac comorbidities, stage at diagnosis, pleural effusion, initial bone and brain metastases, ECOG PS, surgery, radiotherapy, and chemotherapeutic regimens. ROC curve analysis was stable across the stated time frames, yielded an AUC of 0.76 and 0.75 for the training and validation sets, respectively, indicating strong discriminative power. The training and validation loss curves showed convergence, with decreasing and stabilising loss values, suggesting effective model learning and generalisation. An online calculator has been developed based on various leads to estimate the OS.
Conclusions
Integrating deep neural network-derived features with traditional variables significantly enhances the predictive power of survival models. This model is well suited for use by clinicians still operating under older treatment regimens, offering strong predictive capabilities. We call for future collaborations focused on validating and optimising the model in the context of immunotherapy and targeted therapy.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.