Abstract
Abstract
In patients with acute ischemic stroke and unknown symptom onset, reliable estimation of time since stroke onset is important for guiding reperfusion treatment decisions, particularly in settings where advanced imaging is unavailable. In this study, we propose a fully automated onset time estimation pipeline based on non-contrast CT (NCCT) using radiomics features extracted from automatically segmented infarct regions. A segmentation network is used to localize the infarct core, after which radiomics features are derived from both the lesion and the corresponding contralateral region. These features are used to train machine learning regressors, including kernel-based models and a multilayer perceptron, and are compared against pure attenuation-based baselines reflecting net water uptake (NWU). As approximately 25% of patients present without exact onset time and only a last-known-well time (LKWT) is available, these cases are commonly omitted from model development, further reducing already limited training cohorts. To address this problem, we investigate multiple strategies for incorporating LKWT, including surrogate target assignment, stochastic target sampling, and censoring-aware learning formulations. In addition, we assess the robustness of the deployed models to variations in infarct delineation by simulating multiple plausible segmentation boundaries at inference. The literature-parameterized NWU model remained a competitive baseline, while radiomics-based censoring-aware models achieved the lowest errors. Incorporating LKWT through censoring-aware formulations reduced the error compared with known-onset-only training, although this difference was not statistically significant. On the external test set of 32 patients with documented onset time, the censoring-aware support vector regression formulation achieved the lowest median absolute error of 1.12 h (95% BCa CI: 0.91–1.42 h). Censoring-aware formulations also showed low variability under the investigated infarct-boundary perturbations. These findings highlight the potential of NCCT-based regression models for continuous and interpretable onset time estimation. By incorporating LKWT cases, the proposed framework can expand otherwise limited training cohorts, while providing flexible predictions independent of fixed decision thresholds and demonstrating robustness to segmentation uncertainty. Together, these properties suggest that the proposed framework may have future value for clinical decision support, although validation in larger cohorts and prospective studies remains necessary before clinical translation.