SLICE (Simulated Location-based Identification of Compulsive Events)

The SLICE (Simulated Location-based Identification of Compulsive Events) study was a semi-controlled feasibility study investigating the potential of wearable devices and indoor localization to detect routine and repetitive activity patterns. The study consisted of a one-hour session in which participants engaged in both naturalistic and protocol-driven activities within a semi-controlled multiroom residential lab environment. It was approved by the ethics commission of the University of Potsdam (Approval No. 38/2022). 

Study Design and Procedure

The feasibility study was conducted at the Hasso Plattner Institute in Potsdam, Germany, and was designed to replicate everyday living conditions (see Floor Plan picture in the datasest). The environment consisted of a kitchen, bathroom, main room, and hallways, and was designed to support the collection of high-resolution, multimodal sensor data.

The main session lasted approximately one hour, with the total time commitment per participant (including setup and debriefing) being around two hours. During the one-hour recording, participants were left alone in the residential lab to minimize the Hawthorne effect and to create a realistic everyday setting. They were instructed to engage in natural activities such as reading, preparing food, or working on a laptop, thereby generating baseline or NULL data. At the same time, they were asked to perform a specified number of simulated compulsive-like activities, namely handwashing, table cleaning, and door checking, at self-chosen times during the session. Each behavior was performed in two distinct ways to represent routine and compulsive-like patterns, as described below:

Handwashing: Once for routine execution; three consecutive repetitions for compulsive-like execution.
Table Cleaning: Once for routine; three repetitions in a row for compulsive-like.
Door Checking: Once for routine; for compulsive-like execution, participants first checked the door once, then tested the lock five additional times, each with a minimum pause of one second between checks.

Sensor Setup

Participants wore two wrist-mounted IMU devices that recorded tri-axial free acceleration, angular velocity, and Euler angles at a rate of 60 Hz. A head-mounted egocentric camera provided first-person video for annotation, which was deleted after processing. Indoor localization was supported by UWB beacons placed in key areas (kitchen, bathroom, hallway) to provide precise distance estimation.

Data Annotations

All recordings were manually annotated using the ELAN open-source annotation platform based on the first-person video footage to establish ground truth for supervised learning. For having a consistent annotation procedure, we defined the following rules:

Handwashing:

• Starts when the participant stands at the sink to begin washing their hands.
• Handwashing is more complex and individual than the other two activities; for example, some participants first open the tap, while others take soap first.
• Ends with the shaking-off gesture (drying hands is not included at this stage).

Table Cleaning:

• Starts when a cloth is placed on the table to be cleaned.
• Ends when the cloth is removed from the table.

Door Checking:

• Starts when the hand touches the doorknob.
• Ends when the hand releases the doorknob.

A total of 893 activity instances were labeled across the 12 complete datasets. The single activity counts and the mean execution duration in seconds are listed in the Table below. With these annotations, we created a reliable dataset for training and evaluating multimodal activity recognition models, with a particular focus on identifying behaviors that resemble compulsive-like routines in a setting that closely reflects everyday life. 

Issues

Issues included a sensor failure on participant 14's right wrist, leading to incomplete data for that hand. Furthermore, the provided activity descriptions were partly insufficient, causing two participants to perform most activities consecutively. As a result, distinguishing between routine and compulsive executions became difficult, since it was unclear which repetitions represented compulsive-like behavior and which routine ones.

Resulting Dataset

For each participant, there are two CSV files: one for the left wrist and one for the right wrist. Each file contains the IMU data, the beacon data (in meters), and the annotations. The annotations appear in the label column (Handwashing, Table, or Door) and in the kind column (compulsive or routine).