*4.3. Discussion*

On one first hand, from the results presented in Section 4.1, where the impact of modifiers are evaluated, we observe the preferences for the neutral model, where neutral modifier *α<sup>m</sup>* = 1.0 is applied to the linguistic terms. It indicates the non-predominance of a linguistic term over another when defining transition zones between OHRTZs and the aerobic thresholds *VT*1, *VT*2. We note that in the case of high values of HR, which are more sensitive for patients, the yielding model is strongly not recommended by the experts.

On the other hand, based on expert evaluation presented in Section 4.2 where the impact of temporal windows are evaluated, we note that the short-term temporal window suits in *high* zones detecting immediately critical heart rates. Model B (the middle-term temporal window) suits in the *adequate* zone requiring a minimal permanence within, and also suits in the *low* zone without critical differences with regard to other models.

In this way, we can note the adequacy of the clinical protocol for real-time monitoring the CRSs in wrist-worn devices. In addition, the use of a fuzzy model including modifiers and temporal windows has provided a methodology to obtain more accurate terms. This methodology can be extended to model other health contexts based on data stream processing.

Although previous works have been mainly focused on ECG sensors [13,26], the use of a wrist-worn device with the new generation of heart rate sensors provides high accuracy with respect to ECGs [43] for low-risk patients performing low- and medium-intensity exercise. The proposed approach has been implemented within Polar M6000 with Android Wear. Moreover, this wrist-worn device is noninvasive, light and comfortable, but with powerful computing capacity.

On translating the approach to other devices and health contexts, we advise that the quality and precision of heart rate is critical to ensure patient safety. High-risk patients and those with other pathologies could require more accurate devices such as ECGs. In this way, the proposed wrist-worn device just provides a measurement of HR. ECG devices could provide further signal processing of HR where heart rate variability or QRS could be described by means of the here-proposed linguistic terms and fuzzy temporal windows, due to the expanding importance of short-term beat windows in patient analysis [44].

Finally, we note the light processing to compute our methodology, which is based on fuzzy logic, enabling low-cost wrist-worn devices to incorporate it without a computational burden. Other approaches based on similar devices, such as the Fitbit [28] or Garmin [45], could be extended to develop embedded applications providing real-time monitoring during rehabilitation sessions.
