Chronic diseases, particularly those with progressive neurological impairment, present a significant challenge in healthcare due to their impact on millions globally and the limited availability of effective therapies. Addressing this challenge requires innovative approaches, such as leveraging individuals’ genetic features for early intervention and treatment strategies. Due to the irregular intervals of patient visits, clinical data typically appear as censored, necessitating advanced analytical methods. Thus, this study introduces GTsurvival, a novel network architecture that combines graph convolutional networks (GCN) with a neural decision tree, providing promising advancements in disease prediction. GTsurvival utilizes restricted mean survival time (RMST) as pseudo-observations and directly connects them with baseline variables. Through the joint simulation of RMST, GTsurvival can effectively utilize shared information and enhance its predictive ability for patients’ future survival status. Firstly, GTsurvival is introduced to handle complex censored data, emphasizing the crucial role of graphs utilized in GCNs for processing related information among samples. Secondly, the neural decision tree within GTsurvival enhances decision-making by mitigating uncertainty at split nodes, effectively minimizing the global loss function and optimizing survival analysis in high-dimensional datasets. Thirdly, evaluations on simulated datasets and a real-world neurodegenerative disease cohort verify that the proposed GTsurvival method surpasses existing approaches. This superiority is partly attributed to the inclusion of a generalized score test during feature selection, which helps capture variants associated with disease progression.
Accurately estimating the incubation period of emerging infectious diseases, defined as the time interval from infection to symptom onset, is essential for informing quarantine strategies and allocating public health resources. To address the challenges of interval-censored data commonly encountered in incubation period studies, as well as the possibility that some infected individuals may never develop symptoms, we propose a semiparametric probit mixture cure model. By introducing two latent variables to simplify the likelihood function, we develop a corresponding EM algorithm for parameter estimation. Under mild regularity conditions, we establish the consistency and asymptotic normality of the proposed estimators. Numerical simulation studies demonstrate that the proposed method performs well even with limited sample sizes. Furthermore, an empirical analysis of real-world COVID-19 data validates the practical applicability of the model. The proposed semiparametric probit mixed cure model offers methodological support for early prevention and control decision-making in response to sudden infectious disease outbreaks.