IoT-Enhanced Business Processes generate heterogeneous data originating from both IoT devices and process execution engines. While this data provides opportunities for predictive analytics and machine learning, constructing datasets is challenging due to heterogeneity in data sources, formats, and temporal semantics. In particular, integrating continuous IoT data with discrete process events requires explicit decisions on data correlation and temporal alignment. This paper addresses these challenges by proposing a Domain-Specific Language that enables the declarative specification of datasets composed of features derived from heterogeneous data sources. The language makes integration and alignment decisions explicit and is supported by a runtime interpreter that defines precise execution semantics. The approach is illustrated through a logistics case study.