[Paper Review] Detecting Affective Flow States of Knowledge Workers Using Physiological Sensors

Matthew Lee|arXiv (Cornell University)|Jun 18, 2020

Human-Automation Interaction and Safety4 citations

TL;DR

This study proposes a machine learning approach to detect affective flow states in knowledge workers using physiological sensors, leveraging heart rate variability, skin conductance, and other signals to classify flow versus non-flow states. It achieves 0.889 AUC for distinguishing flow from non-flow and 0.98 AUC for rest versus working states, demonstrating strong potential for real-time productivity and well-being support in the workplace.

ABSTRACT

Flow-like experiences at work are important for productivity and worker well-being. However, it is difficult to objectively detect when workers are experiencing flow in their work. In this paper, we investigate how to predict a worker's focus state based on physiological signals. We conducted a lab study to collect physiological data from knowledge workers experienced different levels of flow while performing work tasks. We used the nine characteristics of flow to design tasks that would induce different focus states. A manipulation check using the Flow Short Scale verified that participants experienced three distinct flow states, one overly challenging non-flow state, and two types of flow states, balanced flow, and automatic flow. We built machine learning classifiers that can distinguish between non-flow and flow states with 0.889 average AUC and rest states from working states with 0.98 average AUC. The results show that physiological sensing can detect focused flow states of knowledge workers and can enable ways to for individuals and organizations to improve both productivity and worker satisfaction.

Motivation & Objective

• To objectively detect flow states in knowledge workers using physiological signals, overcoming limitations of self-report and diary-based methods.
• To design work tasks that induce distinct focus states—balanced flow, automatic flow, and non-flow—based on the nine characteristics of flow.
• To develop machine learning classifiers that can distinguish between rest, working, non-flow, and flow states using physiological data.
• To enable real-time interventions that help workers enter and sustain flow states, improving productivity and job satisfaction.
• To support organizational use of flow data for task assignment, training, and performance monitoring while preserving worker well-being.

Proposed method

• Designed lab tasks based on the nine phenomenological characteristics of flow to induce three distinct focus states: balanced flow, automatic flow, and non-flow (overly challenging).
• Collected physiological data including heart rate variability (rmssd, sdNN, total_power), skin conductance levels, and skin conductance response peaks from participants during task performance.
• Used the Flow Short Scale as a manipulation check to validate that participants experienced the intended focus states.
• Trained non-personalized machine learning classifiers (e.g., SVM, random forest) to classify focus states using physiological features.
• Evaluated models using area under the ROC curve (AUC) and accuracy across participants, with cross-validation to ensure generalizability.
• Applied statistical analysis to identify physiological signatures distinguishing each state, such as higher skin conductance peaks in balanced flow and lower variability in automatic flow.

Experimental results

Research questions

• RQ1 Can physiological signals reliably distinguish between non-flow and flow states in knowledge workers during work tasks?
• RQ2 Do distinct physiological patterns correspond to different types of flow states, such as balanced flow and automatic flow?
• RQ3 Can machine learning models trained on physiological data accurately classify rest, working, non-flow, and flow states across diverse participants?
• RQ4 How do physiological markers like heart rate variability and skin conductance vary across different focus states?
• RQ5 What are the implications of detecting flow states for real-time workplace interventions to enhance productivity and well-being?

Key findings

• The machine learning classifier achieved an average AUC of 0.889 in distinguishing non-flow from flow states, indicating strong discriminative performance.
• The classifier for rest versus working states achieved an average AUC of 0.98, demonstrating high accuracy in detecting task engagement.
• Balanced flow was most strongly associated with the highest number of skin conductance peaks, suggesting heightened arousal during optimal challenge-skill balance.
• Automatic flow was characterized by the lowest variation in skin conductance measures, indicating stable, low-arousal engagement.
• The non-flow state (overly challenging task) was associated with the highest heart rate variability (rmssd, sdNN, total_power), indicating parasympathetic activation to manage stress.
• Overall classification accuracy reached 67% across all focus states, 75% for non-flow vs. flow, and 74% for balanced vs. automatic flow, showing consistent performance across participant groups.