Functional Near-Infrared Spectroscopy (fNIRS) is gaining popularity in detection and classification of cognitive and emotional states. In addition to hemodynamic responses arising from functional activity changes in the brain areas of interest, fNIRS signals contain components related to other physiological processes, such as respiration (frequency oscillations around 0.3 Hz) and cardiac pulsation (around 1 Hz). While heart rate and respiration measures have been successfully used as separate modalities to assess mental workload, these components are often discarded in fNIRS studies during the pre-processing. In this study, we examined whether including features related to heart and breathing rate improves the accuracy of mental workload level classification. Data collected with wearable fNIRS devices from 14 healthy participants performing mental workload task (n-back) were used to extract features for the classification. Machine learning classifiers were trained and tested using conventional features separately and in combination with the features derived from the oscillatory activity of respiration and heart pulsation. By comparing the performance, we demonstrated the effect of including proposed features on the classification accuracy of mental workload. In future studies, the examined features might be beneficial for other classification problems where modulations in heart and breathing rates are expected.