Continuous monitoring of blood pressure (BP) is propitious to treatment adherence, yet invasive methods are pricey and perilous, while non-invasive techniques are constrained by human discomfort and observer bias. This study proposes a photoplethysmography (PPG)-based neural network model for BP estimation using short-duration PPG data from 168 patients across various BP stages. Results show that deep learning models, particularly a multi-task CNN, excel in accurately estimating BP from PPG signals by effectively capturing cardiac-related features. This CNN model performs comparably to feature-engineered datasets and demonstrates strong potential for computing BP trends. The study presents that short-duration PPG signals can be directly utilized for BP trend analysis, offering promising applications in early detection and prediction of cardiac-related diseases.