The eastern Adriatic coast is a known hotspot of strong meteorologically induced high-frequency sea-level oscillations, occurring at periods shorter than 1 hour and reaching wave heights of several metres. When highest, these oscillations are termed meteotsunamis. In this study, we test deep-learning methods for predicting maximum daily amplitudes of high-frequency (T < 1 hour) sea-level oscillations at two Adriatic locations, Bakar and Ploče, using convolutional neural networks driven by past sea-level observations and atmospheric predictors from the ERA5 and CERRA reanalyses. We evaluate two deep-learning architectures designed to test different approaches to representing sea-level and atmospheric forcing. The first architecture, HFNet, is based on the HIDRA family of models, a general low-frequency sea-level forecasting framework that has been extensively evaluated in the Adriatic and shown to provide a credible baseline for sea-level prediction. The second architecture, HFNet<sub>JE</sub>, extends this approach through joint encoding of atmospheric predictors and a more extensive processing of past sea-level information, with the aim of improving the representation of processes associated with high-frequency sea-level oscillations. Analysis of more than 20 years of data shows that high-frequency sea-level extremes are larger in Bakar (> 60 cm) than in Ploče (< 35 cm), occur ~6 times per year, and are most common during the warm season. Both architectures reproduce the observed variability, with higher skill for typical than for extreme events. HFNet<sub>JE</sub> performs best overall and under typical amplitude conditions, whereas HFNet more effectively captures extreme events, although these remain systematically underestimated in both architectures. Model performance is higher at Ploče, likely because of its smaller sea-level range and simpler response to atmospheric forcing. Models forced with ERA5 consistently outperform those using the higher-resolution CERRA in predicting extremes, suggesting limited added value from increased spatial resolution. Ablation experiments indicate that several predictors are redundant for average forecasting performance, whereas extreme-event prediction generally benefits from the full predictor set. Overall, the results demonstrate the potential of deep learning for prediction of high-frequency sea-level oscillations in the Adriatic, but also highlight persistent limitations in forecasting rare high-amplitude events.