Increasing concerns over groundwater drought risks, which threaten water availability and adversely impact ecosystems, agriculture, and human activities, underscore the necessity of comprehensive evaluation methods. This research introduces a meticulous approach to evaluating groundwater drought risk (GWDR) in the semi-arid expanse of the Kansai River Basin, West Bengal, India. It intricately amalgamates the Soil and Water Assessment Tool (SWAT) model with three distinct machine learning algorithms namely, Support Vector Machine (SVM), Random Forest (RF), and Neural Networks (NN). The assessment relies on a diverse array of 26 thematic datasets encompassing hydrological, meteorological drought risk, and socioeconomic conditioning variables. The SWAT model has been used to derive hydrological parameters including groundwater recharge, lateral flow, base flow, surface runoff, evapotranspiration, return flow, and soil water content. Simultaneously, a pre-monsoonal water level dataset from 503 well locations is adhered to an impartial sampling strategy, maintaining a 70:30 ratio for training and testing datasets. The ensuing GWDR maps, derived through SVM, RF, and NN models, reveal four discerning risk classes across the study area. High-risk zones conspicuously predominate in upper catchment areas, while low-risk zones find their strategic position in the lower catchment regions. The area under the receiver operating characteristic curve (AUC-ROC) for the RF model, showcases an impressive 91% success rate, surpassing its counterparts SVM and NN models, which attained success rates of 88.4% and 80.7%, respectively. The Mann–Kendall test with Sen's slope analysis confirms a noticeable decline in groundwater levels within high to moderately high GWDR zones, supporting the study's findings. These findings significantly impact water resource management in semi-arid regions, emphasising the need for proactive measures to address evolving groundwater drought risks.