Human activity recognition (HAR) is an important research area due to its potential for building context-aware interactive systems. Though movement-based activity recognition is an established area of research, recognising sedentary activities remains an open research question. Previous works have explored eye-based activity recognition as a potential approach for this challenge, focusing on statistical measures derived from eye movement properties---low-level gaze features---or some knowledge of the Areas-of-Interest (AOI) of the stimulus---high-level gaze features. In this paper, we extend this body of work by employing the addition of mid-level gaze features; features that add a level of abstraction over low-level features with some knowledge of the activity, but not of the stimulus. We evaluated our approach on a dataset collected from 24 participants performing eight desktop computing activities. We trained a classifier extending 26 low-level features derived from existing literature with the addition of 24 novel candidate mid-level gaze features. Our results show an overall classification performance of 0.72 (F1-Score), with up to 4% increase in accuracy when adding our mid-level gaze features. Finally, we discuss the implications of combining low- and mid-level gaze features, as well as the future directions for eye-based activity recognition.