Multi-modal detection of acute fear of falling in older adults: A proof-of-concept study

<ns3:p>Fear of falling in older adults is generally studied as a chronic condition, yet spikes in acute fear of falling remain underexplored, despite previous research showing that they often precipitate falls. This work introduces a multi-modal sensor-based framework for detecting theoretically defined 'potential fear of falling' by combining gaze elevation and heart rate signals captured from wearable eye tracking and wrist devices in older adults living in the community. We trained five conventional classifiers (logistic regression, KNN, random forest, XGBoost, CatBoost), optimized for minority class F1, and combined two ensembles: (1) random forest + CatBoost + KNN and (2) random forest + logistic regression + KNN. We also applied spectrogram-based transfer learning by fine-tuning the pre-trained VGG16 and ResNet50 models on accelerometer data. In the individual-classifier analysis, XGBoost, KNN, and random forest achieved ROC AUC = 0.99 and minority-class F1 of 0.93, 0.90, and 0.85, respectively. The ensemble models performed better than individual classifiers on multi-modal and accelerometer-only inputs, though overall performance remained modest without multi-modal signals in the latter case (minority-class F1 = 0.39). Transfer models outperformed ensembles. These results demonstrate that ensemble and spectrogram-based transfer learning models provide robust, high-sensitivity detection of potential acute fear of falling in multi-modal signals. This work lays the foundation for future studies to explore acute fear of falling biomarkers in larger cohorts and paves the way for personalized fall prevention interventions in everyday settings.</ns3:p>