From Lab to Life: Enhancing Wearable Airbag Fall Detection Performance with Minimal Real-World Data

Falls are the leading cause of accidental injury or death among older adults, particularly those with neurological conditions like stroke or Parkinson’s disease (PD) which impair mobility and balance. In these populations, falls are nearly unavoidable, but wearable airbags equipped with pre-impact fall detection algorithms may offer life-saving protection. However, collecting real-world fall data to train these pre-impact algorithms is time-consuming and costly, often leading to the use of simulated falls for model training. This study aimed to 1) identify the best-performing machine learning algorithms for real-world pre-impact fall detection using only simulated falls for training (independent environment approach) and 2) evaluate whether integrating a small amount of real-world data improves detection performance (combined environment approach). Real-world fall data were collected from 22 individuals (N = 12 stroke; N = 10 PD) wearing a waist-mounted wearable airbag device with inertial measurement units (IMUs). A simulated dataset (645 falls, 979 non-falls) was used to train models, while real-world data (32 falls, 32 non-falls) were used for testing and refining models. In the independent environment approach, random forest classifiers achieved the highest performance (F1 = 0.86). Incorporating real-world data and model fine-tuning improved performance, with the best combined environment model reaching an F1 score of 0.93. Feature analysis identified gyroscopic data as the most critical for classification. While real-world data collection remains challenging, integrating even a small amount of real-world falls significantly improves model generalizability. These findings highlight the potential of pre-impact fall detection algorithms for real-world applications, particularly in high-risk populations.