**Background:** The Ten Meter Walk Test (10MWT) is a standard physical test used to assess gait speed, mobility, and vestibular function, typically measured by a healthcare professional. Technological devices, such as inertial sensors, can provide objective and automated measurements, but existing datasets are often private or rely on expensive equipment. This study aimed to create a public, low-cost dataset using a smartphone's built-in accelerometer, magnetometer, and gyroscope to facilitate the development of automated diagnostic frameworks for physical status, disease progression, and telemedicine applications.
**Methods:** Data were collected from 109 volunteers (30 males, 79 females) aged 55 to 97 years (mean 81.0642 ± 9.2849 years) living in senior residences in the Fundão municipality, Portugal. Participants placed a Xiaomi Poco X3 Pro smartphone in a waistband. A custom mobile application recorded raw sensor data at 100 Hz for accelerometer and gyroscope (LSM6DSO model) and 50 Hz for magnetometer (ak0991x model) during a 10-meter walk. Each participant performed two captures, yielding 218 expected captures; after excluding invalid ones (due to hardware/software issues or participant incapacity), 208 captures were retained, of which 204 were valid (98.07% of real captures). The dataset includes accelerometer (m/s²), gyroscope (rad/s), and magnetometer (µT) data with Unix timestamps, along with user metadata (age, gender, weight, height, walking aids, diseases, physical activity, nutrition, etc.). Diseases were categorized into 16 groups (e.g., cardiovascular, neurological, musculoskeletal). The authors also computed features such as total time (mean 16.09 s), steps (mean 66.43), mean speed (108.12 m/s), mean force (1.63 N), distance (25.18 m), and mean power (14,615.08 W). Machine learning classifiers (k-Nearest Neighbors, Linear SVM, RBF SVM, Decision Tree, Random Forest, Neural Networks, AdaBoost, Naive Bayes) were applied to classify gender and disease categories.
**Key Results:** For gender classification, the best performance was achieved by RBF SVM (precision 96%, recall 96%, F1-score 96%, accuracy 96%), followed by Neural Networks and AdaBoost (both 95% accuracy). For disease classification, accuracy was low overall: Linear SVM achieved 23% accuracy (precision 22%, recall 64%, F1-score 25%), and AdaBoost achieved 18% accuracy (precision 18%, recall 40%, F1-score 22%). The authors note that the dataset is imbalanced and heterogeneous, limiting disease classification performance. The dataset contains 208 folders (each with accelerometer.txt, gyroscope.txt, magnetometer.txt, and user_data.json).
**Clinical Implications:** This public dataset provides a valuable resource for developing automated, low-cost methods to assess gait and mobility using smartphone sensors, which could support remote monitoring, telemedicine, and personalized medicine for elderly populations. The high accuracy in gender classification suggests that sensor data capture meaningful individual differences. However, the poor disease classification accuracy highlights the need for larger, more balanced datasets and advanced feature engineering. The study also emphasizes the importance of standardized device placement, data quality control, and addressing sensor limitations. Future work should focus on expanding the dataset to include diverse populations and improving algorithms for clinical applications such as fall risk assessment, rehabilitation monitoring, and early detection of mobility impairments.