Heart rate response to physical activity is widely investigated in clinical and training practice, as it provides information on a person's physical state. For emerging digital phenotyping approaches, there is a need for individualized model estimation. In this study, we propose a zero-poles model and a data-driven evolutionary learning method for identification. We also perform a comparison with existing first and second order models and gradient descent identification methods. The proposed model is based on a five-phase description of heart rate dynamics. Data was collected from 30 healthy participants using a treadmill and a thoracic sensor in two protocols (static and dynamic), for increasing and decreasing activity. Results show that the zero-poles model is a good fit for heart rate response to exercise (Pearson's coefficient ρ>
.95), while first and second order models are also suitable (ρ>
.92). The evolutionary learning method shows excellent results for fast model identification, in comparison with least-squares methods (p<
.03). We surmise that the parameters of investigated linear dynamic models make good candidates for digital biomarkers and continuous monitoring.