Process-aware information systems support the execution and monitoring of business processes. The recorded event data can be leveraged to predict the remaining time of running cases, enabling organizations to meet service-level targets, and enhance operational performance. However, most existing approaches treat cases as being independent from each other, even though there often is contention between cases for shared resources. While some approaches account for inter- case dependencies, they overlook the queuing dynamics arising from work handovers between resources. Moreover, inter-case-aware approaches often use traditional tree-based models, which generally achieve lower accuracy than modern deep learning models. To overcome these limitations, we introduce QTab, a novel remaining time prediction approach powered by two components: (1) modeling of resource interactions and work handovers through a queuing network to derive informative inter-case features; (2) Modern tabular neural networks trained in sophisticated ensembling pipelines. Experimental results on eight real-world event logs show that QTab achieves competitive or superior predictive accuracy compared to existing tree-based and deep learning approaches in remaining time prediction. At the same time, it offers better scalability in feature extraction and improved efficiency in model training.

Top