Accurate dissolved oxygen (DO) modelling is crucial to better understand the drivers and timing of oxygen declines, supporting early warning and mitigation strategies. This study aims to predict DO dynamics by integrating hydrological and aquatic ecosystem modelling with advanced deep learning techniques based on Long Short-Term Memory (LSTM) networks. We focus on the Mar Menor lagoon in southern Spain, an ecosystem of high ecological and socioeconomic importance, where oxygen depletion events have caused massive fish kills with significant impacts on fisheries and tourism. We simulated 20 years (2003 - 2023) daily hydrological dynamics using SWAT+ model for Mar Menor catchment, generating discharge, nutrients and sediments outputs. These were used as inputs for GOTM-WET lagoon model, which produced daily and hourly outputs of key biogeochemical variables, with DO predictions focused on the deepest layer (6.5 m). We pre-trained an LSTM model to predict DO using combinations of meteorological data, SWAT+, and GOTM-WET outputs as inputs, using the GOTM-WET DO output as targets for model pre-training to overcome the limited availability of observed data. The model was then fine-tuned with weekly DO observed data. SHAP analysis identified water temperature, density, and macrophyte dry weight as key DO predictors. The model performance improved substantially after fine-tuning, with R2 increasing from 0.59 to 0.77 and RMSE dropping from 0.89 to 0.49 mg/L, outperforming the GOTM-WET baseline (R2 = 0.45, RMSE = 1.03 mg/L). These results highlight the potential of coupled hydrological and aquatic ecosystem modelling with advanced deep learning techniques to enhance DO modelling.