<p>Accurate prediction and adaptive selection of the modulation and coding scheme (MCS) in wireless communication systems are crucial for enhancing spectral efficiency, optimizing data transmission, and ensuring quality of service. Traditional rule-based MCS methods lack the flexibility to adapt to dynamic wireless environments. This work presents an open radio access network (O-RAN) framework that enables adaptive MCS selection for multi-user multiple-input multiple-output systems by using instantaneous uplink channel estimates to predict MCS without requiring user feedback. Five machine learning (ML) techniques–convolutional neural network–long short-term memory (CNN-LSTM), deep neural network (DNN), support vector machine, random forest, and bagging k-nearest neighbours–are trained and evaluated using both simulated and over-the-air (OTA) measurements collected from an indoor O-RAN testbed. The models are deployed as external applications (xApps) using parallel microservices to enable near-real-time inference and adaptive MCS control. In controlled simulation conditions the proposed approach achieves very high classification accuracy (CNN-LSTM up to 98.89%; DNN 98.55%), and OTA experiments on the indoor testbed confirm strong predictive performance. Uplink throughput, used as a key performance indicator, improved by approximately 10–20% in typical test scenarios and reached gains up to 30% in favorable multi-user simulation conditions. These results demonstrate the potential of ML-integrated O-RAN systems for adaptive rate control, while underscoring the need for further validation in larger-scale, mobile, and more heterogeneous deployment environments.</p>