Model training describes scenarios in which data are primarily used to develop, train, validate, or refine predictive, classificatory, or generative models. In contrast to data products, where the dataset itself represents the central artifact, and data-driven services, where data are embedded into an ongoing service process, the primary output of model training is a trained analytical model.

Machine learning research defines model training as the process of fitting an analytical function to historical data in order to identify patterns and enable prediction or classification (Jordan & Mitchell, 2015; Hastie et al., 2009). In this context, data do not merely inform a decision process but actively shape the internal structure and parameters of the resulting model (Goodfellow et al., 2016). The trained model thereby becomes the primary artifact of value creation.

Within applied domains such as healthcare, the development of predictive models requires careful attention to data representativeness, training and validation procedures, and performance evaluation criteria (Beam & Kohane, 2018). Predictive modeling literature further emphasizes the conceptual distinction between explanatory data analysis and predictive model construction, highlighting the importance of explicitly defining modeling objectives and evaluation metrics (Shmueli, 2010).

Within the specification phase, model training therefore requires explicit clarification of:

• the modeling objective (e.g., prediction, classification, risk scoring),
• the required training and validation data,
• performance criteria and evaluation metrics,
• and the intended organizational embedding of the trained model.

In regulated environments such as the EHDS context, additional governance considerations arise, particularly regarding lawful access to training data and secure processing requirements.