Journal article
A lead-time-aware approach to fall prediction using deep and transfer learning
Applied artificial intelligence, Vol.40(1), pp.1-30
09/2026
Metrics
1 Record Views
Abstract
Falls pose imminent threats to the elderly population that cause injuries, deaths, dis-abilities, and restricted mobility. Predicting falls before they occur is crucial for prevent-ing these outcomes. Existing fall prediction studies have limitations, including limited falldata, inconsistent prediction lead-times, reliance on linear features and traditionaltechniques, and a focus on future fall risk estimation. This study establishes a lead-time-aware temporal classification framework that systematically evaluates multiple pre-fall lead times using deep and transfer learning techniques. The proposed method isevaluated on two public datasets to determine whether pre-fall-onset motion signalscontain distinct patterns that can predict near-future falls at different lead times. Thedeep learning approach achieves a mean accuracy of 80%-94% across the lead-time-based labels, datasets, and cross-validations. Transfer learning marginally improves theaccuracy to 82%-94%, while feature extraction further increases it to 88%-97%. For thebest-performing models, combining class weighting with PR-curve-based optimalthreshold selection achieved F1-scores of 46%–84%, precision of 60%–82%, and recallof 74%–86%, despite significant class imbalance. The results also indicate that shorterpre-fall-onset lead-times are more closely associated with fall events; however, longerlead-times exhibit a weaker association with fall data.
Details
- Title
- A lead-time-aware approach to fall prediction using deep and transfer learning
- Creators
- Prithwi Chakraborty - Southern Cross UniversityGolam Sorwar - Southern Cross University
- Publication Details
- Applied artificial intelligence, Vol.40(1), pp.1-30
- Publisher
- Taylor & Francis; PHILADELPHIA
- Identifiers
- 991013402370202368
- Copyright
- © 2026 The Author(s).
- Academic Unit
- Information Technology; Faculty of Science and Engineering
- Language
- English
- Resource Type
- Journal article