This work aims to estimate the time since stroke onset – a key factor in determining clinical decisions, such as eligibility for thrombolysis – using non-contrast CT (NCCT) images. In many cases, the exact onset time is unknown, limiting treatment options under current guidelines. To address this challenge, we propose a method to derive the onset time directly from imaging data. Our approach utilizes Radiomics features extracted from the infarct region and its contralateral counterpart to train machine learning (ML)-based regressors including a multi-layer perceptron (MLP) for onset time estimation, comparing performance against net water uptake (NWU)-based methods. To adequately incorporate Last-Known-Well Time (LKWT) into the MLP training process, we introduce a loss function inspired by survival analysis. Experimental results demonstrate that a combination of L1 and the proposed ranking-aware loss achieves the best performance, with a median absolute error of 1.90 h, outperforming the MLP trained with L1 loss alone, other ML-based regressors, and NWU-based methods. These results highlight the potential of the dedicated incorporation of LKWT, though further research is needed to refine this approach for clinical applicability.