<p>Accurate prediction of groundwater conditions during tunnel excavation by Tunnel Boring Machine (TBM) is essential for preventing water inrush hazards and ensuring safe TBM operation. In previous studies, extensive geological information was collected in drill-and-blast tunnels and utilized in different machine learning models to predict groundwater conditions. However, most of these geological parameters cannot be directly obtained during TBM excavation due to the invisibility of the tunnel face, resulting in degraded model performance. In contrast, a large volume of TBM operational parameters can be continuously recorded during excavation, and several of them may provide useful information on groundwater conditions. Despite this potential, these parameters have not been effectively incorporated into machine learning models for predicting groundwater conditions. This study proposes a machine learning method that integrates available geological information with abundant TBM operating data to address this gap. In the proposed method, groundwater conditions are categorized into four classes according to the observed water inflow rate. A LightGBM-based classifier is used to develop machine learning models, and Bayesian optimization is applied to tune hyperparameters. Class-weighting and SMOTE techniques are used to mitigate data imbalance. Results on the test dataset demonstrate that the proposed LightGBM model achieves an accuracy of 0.963 and an F1-score of 0.952. Compared with the LightGBM model relying solely on geological information, the accuracy and F1-score are improved by 0.083 and 0.094, respectively. Compared with the LightGBM model using only TBM operational data, the corresponding improvements are 0.055 and 0.065, respectively. Notably, the model exhibits strong capability in identifying water-inflow conditions, providing valuable support for early warning and risk mitigation in TBM tunneling projects.</p>