Finding the Flow in the Heartbeats: Creation and Analysis of an Affective Computing Framework for detecting the Flow State through HRV

We explore the potential of detecting the Flow state using affective computing techniques, with a focus on Heart Rate Variability (HRV) analysis. Flow, as defined by Csikszentmihalyi, is a mental state characterized by a balance in the perceived challenges and individual skills. While studies abound in literature, real-time detection of Flow remains challenging due to its self-absorbing nature and reduction of self-awareness. The research develops along two primary directions: psychological and technological. On the psychological side, the thesis aims to better understand Flow and its relationship with emotions and engagement, while investigating individual traits that may facilitate people to experience Flow. On the technological side, the work seeks to build a framework for detecting Flow through biosignals, optimizing the acquisition and synchronization of physiological data, and implementing machine learning techniques to identify different patterns of cardiac activity. First, we demonstrate that HRV features alone are effective in recognizing emotional states with Machine Learning (ML) using the MAHNOB-HCI database, achieving improved performance compared to previous studies that utilized multiple physiological signals. Second, we address inter-individual variability through baseline clustering, which showed improved classification performance for certain emotional categories. After that, we developed an open-source experimental framework, made available on GitHub, integrating Tetris gameplay, PolarH10 Inter-Beat Intervals (IBIs) recording, and video data, with the implementation of multiple timestamps ensuring temporal precision in multimodal data collections. With the obtained framework, we performed a data collection phase with 53 participants, showing significant correlations between HRV features and self-reported states of Flow, emotions, and engagement. Finally, machine learning models were applied to classify mental states, with results in line with some studies in the existing literature: HRV features were calculated over various time windows (3 minutes, 1 minute, 30 seconds, 10 seconds), and we found that the 3-minute time windows gave the most significative results. We also applied the baseline clustering technique, however, without obtaining conclusive results. The study also examined how the experience of Flow is related to other emotions resulting shaped by higher dominance, moderate to low arousal, and somewhat positive valence. The correlations between Flow disposition and personality traits, instead, revealed significant associations with traits like agreeableness, openness, intellectual curiosity, and trust. The findings indicate that HRV features are promising for detecting Flow, and the methods developed in this thesis contribute to both the theoretical understanding and practical detection of Flow in real-time applications.