IMU50: a high-frequency multi-modal wrist-worn IMU dataset of 50 subjects over 5 continuous days for supervised and...

We present a multiday, high-resolution inertial measurement unit (IMU) dataset collected from 50 healthy volunteers using the Actigraph Leap wrist-worn device. Each participant was monitored continuously for five consecutive days and nights, yielding an average of approximately 136 h of raw sensor recordings per subject. The dataset includes three-axis accelerometer signals, three-axis gyroscope signals sampled at 128 Hz, double-channel photoplethysmogram (PPG) sampled at 25 Hz, and wrist skin temperature acquired every minute. Hourly estimates of metabolic equivalents of task (METs) and caloric expenditure are also provided. These estimates are obtained by the Actigraph using Freedson algorithms computed over an additional 32 Hz embedded accelerometer. In addition to sensor data, the repository contains participant-level metadata encompassing age, sex, body weight, height, body mass index (BMI), hip and waist circumference, educational level, and lifestyle. All recordings were conducted under free-living conditions without restrictions on daily activities, capturing a realistic spectrum of behaviours including sleep, sedentary periods, and light daily activities. Data are stored in open, well-documented formats with accompanying code examples to facilitate loading and preprocessing. The dataset is intended to support research in two complementary directions: (1) self-supervised representation learning from raw IMU signals, exploiting the large volume of unlabelled continuous recordings; and (2) supervised regression and classification tasks targeting metabolic outcomes, morphological parameters, skin temperature, and heart rate from PPG data. No activity labels were collected during recording, making this resource particularly suitable for unsupervised and weakly supervised learning paradigms. The dataset fills a gap in the public availability of long-duration, multi-day wrist IMU recordings linked to metabolic and clinical metadata and is expected to serve as a benchmark resource for the wearable computing and digital health communities.