TL;DR
Researchers have developed an "ambient-physical" system that uses passive sensors to predict worker concentration levels in open-plan offices with high accuracy. By fusing environmental data (like CO2 and noise) with physical activity data from smartphones, the system can identify the key drivers of focus, potentially allowing managers to optimize office layouts for better productivity.
What Problem Does This Paper Address?
Open-plan offices are a double-edged sword: they promote collaboration and save costs, but they are notorious for distractions, noise, and a lack of privacy. While we know these factors hurt productivity, it is difficult to measure how and when they impact an individual's concentration in real-time without being intrusive (e.g., using cameras or constant surveys).
The authors tackle the challenge of "concentration inference"—automatically determining how well a worker is focusing based on their surroundings and their physical behavior. The goal is to move away from subjective "gut feelings" about office design toward data-driven environments that can provide personalized recommendations or adjust ambient conditions to help workers stay in the "zone."
Methods
The authors propose a multi-layered system architecture consisting of:
- Passive Sensing Pool: They deployed Netatmo Weather Stations to collect ambient data (temperature, humidity, CO2, noise, and pressure) and used a custom smartphone app, OpenSense, to collect physical data (accelerometer, gyroscope, and pedometer) at 50 Hz.
- Window Stream Processing: To handle the different "tempos" of data, they used a sliding window approach. Ambient features were extracted over 30-minute windows, while physical movement features used 5-minute windows.
- Feature Fusion: The system aligns these disparate data streams by timestamp, creating a combined "ambient-physical" feature vector.
- Machine Learning Classifiers: They tested eight different classifiers, including Random Forest, XGBoost, and SVMs. The standout performer was CatBoost, a gradient-boosting algorithm that handles categorical data and imbalanced datasets effectively.
Datasets
The study was conducted over four weeks across two distinct pilot sites of an engineering consultancy firm in Australia:
- Site-1: A traditional single-floor open-plan office.
- Site-2: A modern, multi-layered space with a central void and different "zones."
- Participants: 31 volunteer corporate workers.
- Ground Truth: Workers self-reported their concentration levels twice a day (10:00 AM and 3:30 PM) using a 5-point Likert scale via the OpenSense app.
Key Insights and Ideas
- The Power of Fusion: The most significant finding is that neither ambient nor physical data is sufficient on its own. Combining them (A+P) led to a substantial jump in accuracy (often 10-30% higher than using physical data alone).
- Environmental Triggers: High CO2 density and lower temperatures (specifically between 21.5–22.5°C) were correlated with lower concentration levels for many occupants.
- The "Meeting" Effect: In Site-1, concentration peaked at a neutral level with 0-3 formal meetings but plummeted when workers had 5 or more. Interestingly, informal meetings in the modern Site-2 layout seemed to correlate with slightly better median concentration compared to Site-1.
- Context Matters: The system performed differently at the two sites, suggesting that office "topology" (layout, voids, and zoning) significantly changes how environmental factors influence focus.
Key Results
- Top Performance: The CatBoost classifier consistently outperformed all other models.
- Accuracy Milestones:
- In Site-1, the system achieved a staggering 97.16% accuracy for morning sessions and 96.56% for afternoon sessions when using the combined feature set.
- In Site-2, which had a more complex multi-layered layout, accuracy was lower but still impressive, peaking at 76.07%.
- Feature Importance: Magnetometer, accelerometer, and gyroscope readings were among the top indicators, alongside humidity and CO2 levels.
Limitations and Future Work
The authors acknowledge several hurdles for scaling this technology:
- Sample Size and Diversity: The study involved 31 participants from a single firm. Expanding this to different industries and larger cohorts is necessary to build a truly "generalized" model.
- Interoperability: Moving between sites revealed challenges in sensor infrastructure (e.g., Site-2 only supported 5GHz Wi-Fi, which affected some older ambient sensors).
- Ground Truth Collection: While the current system uses "active" surveys (asking the user), the authors want to move toward even less obtrusive ways to collect ground truth to avoid interrupting the very concentration they are trying to measure.
- Sequential Modeling: Future work will look at concentration as a sequential time-series task rather than a series of isolated classification snapshots.
Summary updated