Bridging the Gap in Cancer Survival Analysis: Deep Neural Networks for Left-Truncated Data

Deep learning has demonstrated better performance than traditional regression methods in handling right-censored cancer survival data; however, its application in survival analysis remains limited due to censoring-related data loss and the lack of approaches addressing left-censored or left-truncated cases. The aim of this study is to introduce a novel deep neural network algorithm designed to incorporate left truncation in survival analysis. Building upon prior research, we propose an innovative approach that integrates deep learning with survival analysis by applying complementary and consensus statistical principles to simultaneously account for left truncation and right censoring. The cumulative hazard function was estimated using the Breslow estimator and incorporated as an output variable in the proposed model. Model performance was assessed using the integrated area under the curve (iAUC), which quantifies predictive accuracy across all time points. The proposed method demonstrated robust predictive performance with an iAUC of 0.745 (95% CI: 0.705–0.785) when applied to left-truncated and right-censored survival data. Its performance was comparable to widely recognized statistical and machine learning techniques for survival analysis, including the Cox proportional hazards model (iAUC = 0.779, 95% CI: 0.739–0.812) and random survival forests (iAUC = 0.829, 95% CI: 0.795–0.858). This study introduces a methodological extension of deep neural network models, previously restricted to right-censored survival data, to also accommodate left-truncated survival data. The validation results provide evidence that the proposed model achieves predictive performance comparable to well-established methods, thereby broadening the applicability of deep learning in survival analysis.