Unattended radiological sensor networks must take advantage of contextual data e.g., open-source data in addition to historical sensor signals to anticipate nuclear isotope signatures and mitigate nuisance alarms in urban environments. To address these challenges, we have developed novel AI-driven predictive analytics – using machine learning and deep learning models – to predict radiological isotope signatures by learning from historical sensor data for 9 months in DC and 7 months in Fairfax in 2019 and 2020. Our sensor data includes alerts from three medical Tc-99m, I-131 and 511 from Positron Emission Tomography (PET) and one industrial Cs-137 isotopes. Our AI-driven analytics leverage historical data [1] to anticipate the number of alarms per isotope per sensor in the next hour across nine sensors in two locations: Fairfax, VA and Washington, DC. We design experiments to contrast performance of the state-of-the-art ML models (Logistic Regression, Random Forest, ARIMA, SVM, and K-Nearest Neighbors) with deep learning models that rely on Long-Short Term Memory (LSTM) [2] and Transformer [3] architectures. n Table 1, we present experimental results of the LSTM model compared to two top performing ML models – Logistic Regression (LR) and Random Forest (RF) – over two locations: Fairfax, VA (4 sensors over 7 months) and Washington, DC (5 sensors over 9 months). The LSTM model outperformed the top ML models for the Fairfax, VA location and met or exceeded the performance of the top ML models for the Washington, DC location across several metrics.