Development of an assistive device based on multisymbolic pupillary communication for patients in a state of complete paralysis.

In patients affected by Complete Locked-In Syndrome (CLIS), all forms of voluntary communication are impaired due to total motor paralysis, despite preserved consciousness and cognitive abilities. In this context, conventional communication methods prove ineffective, highlighting the need for assistive devices that rely on residual physiological signals that can still be voluntarily modulated. The pupil accommodative response (PAR), which regulates pupil diameter based on the distance of the visual focal point, is typically preserved even in individuals with CLIS. This makes it a promising communication channel, as it can be voluntarily triggered by focusing gaze on visual targets positioned at different depths. The aim of this thesis is the design and development of a non-invasive, compact, and low-cost interaction system based on real-time pupil area monitoring and the detection of constriction events associated with PAR. The device is built around a Raspberry Pi Zero equipped with an infrared camera and lighting system, and integrates a segmentation algorithm capable of extracting pupil area in real time from the video stream. Unlike previous approaches that exploit PAR as a binary signal, the proposed system leverages its continuous nature by implementing a four-level coding paradigm based on the combination of two focal distances and two possible durations of the pupillary constriction. This strategy enables the transmission of a larger amount of information, significantly enhancing the communicative potential of the system. The device is supported by a computer responsible for processing and hosting a modular graphical interface, through which the user can access various applications. Among these, a text-based communicator has been developed, enabling word composition through modulation of the PAR. The system was validated through a series of experiments involving healthy subjects, virtually in the absence of any external movement, including convergence/divergence movements of the relevant eye. The results demonstrated effective pupil segmentation, with Dice coefficients exceeding 92%, and an excellent detection of pupillary constriction events, achieving both precision and recall values above 95%. Event classification accuracy exceeded 80%, indicating a good capacity to discriminate the 4 different patterns of response, despite the high intra and inter individual variability of the signal. Nevertheless, to ensure a more robust interaction, higher accuracy would be desirable and could possibly be achieved by future developments involving artificial intelligence-based techniques. Overall, the device demonstrated promising capabilities in supporting communication without requiring user training, suggesting potential applicability in both clinical and home settings.