<p>This article examines the influence of radio and traffic features on perceived downlink throughput in an Orange Senegal 4G LTE small-cell network, using a set of real data collected from seven small cells over a period of more than three months. The main objective is to identify the feature that best explains variations in perceived downlink throughput and to assess the performance of machine learning models (LR, DT, RF, and MLP) and a deep learning model (DNN) for its prediction. After rigorous data preprocessing and the selection of 13 relevant features, correlation and prediction analyses are conducted. The results show that, among all the features studied, the average Channel Quality Indicator (CQI_Avg) is the most decisive factor affecting perceived downlink throughput. This dominance is explained by the CQI’s direct causal role in link adaptation mechanisms (modulation and coding), in contrast to load or traffic features, and whose impact remains indirect. From a predictive standpoint, the deep neural network (DNN) outperforms all other models, achieving an accuracy of 96.1%, an RMSE of 0.73, and an MAE of 0.49, demonstrating its ability to capture complex non-linear relationships between radio quality, resource allocation, and traffic. The study thus highlights how AI-based approaches can overcome the limitations of classical statistical analyses. These findings have major operational implications for throughput optimization and QoS management in 4G LTE (Long Term Evolution) networks. By extension, 5G networks should primarily rely on intelligent monitoring and prediction of CQI, integrated into machine learning and deep learning based solutions.</p>