# InclusiveHAR: A smartphone-based dataset for human activity recognition across diverse physical abilities

## TL;DR

InclusiveHAR is a smartphone-based dataset for human activity recognition that includes participants with and without disabilities, aiming to improve healthcare AI systems.

## Contribution

The novel contribution is a diverse HAR dataset including individuals with disabilities, enabling inclusive AI development.

## Key findings

- The dataset includes 20 participants with varied physical abilities performing six daily activities.
- Performance was evaluated using K-NN, SVM, and XGBoost models.
- The dataset is accompanied by detailed documentation for transparent and reproducible analysis.

## Abstract

Human Activity Recognition (HAR) has numerous applications in healthcare, rehabilitation, athletics, and smart environments. Effective AI models rely on diverse and representative datasets to achieve robust generalization. However, the majority of existing HAR datasets are collected exclusively from non-disabled individuals, limiting their applicability in real-world healthcare scenarios involving the elderly or individuals with disabilities. To address this limitation, we introduce InclusiveHAR, a novel smartphone-based HAR dataset collected from 20 participants, including 10 non-disabled individuals and 10 individuals with disabilities, of whom five had a single disability and five had multiple distinct conditions. Participants performed six daily activities: walking, standing, sitting, jogging, ramp ascent, and ramp descent. The dataset captures a wide range of movement patterns and behavioural variability, with particular emphasis on differences in activity execution observed in individuals with disabilities. Data were collected using an iPhone 14 Pro at a sampling rate of 50 Hz (one sample every 20 ms). The SensorLog app was used to lock the rate at 50 Hz. To illustrate the potential use of the dataset, a baseline evaluation is provided under multiple training scenarios using the MLP machine learning model. In this paper, we report and evaluate the performance of dataset against K-NN, SVM, and XGBoost models. In addition, the dataset is accompanied by detailed feature descriptions and comprehensive documentation of the data collection protocol, enabling transparent analysis, reproducibility, and future comparative studies. The InclusiveHAR dataset offers a valuable resource for investigating activity recognition performance across diverse participant groups and for supporting the development of inclusive HAR systems in healthcare and assistive technology applications.