Note , This project was funded by Liverpool John Moores University and part of a wider project wherein I was employed as a research assistant (support role).
Objective
Stair falls have been a major healthcare concern, especially among older adults who are more likely to suffer serious injury, or in extreme cases, death, when they fall on stairs. Generic fall-risk assessment tools exist but are not stair-specific, which motivates the need for biomechanical stair-specific testing to profile stepping strategies and understand the risky ones.
A study within LJMU highlighted the potential of stepping profiling to predict profiles risky to stair falls in older adults using statistical and unsupervised machine-learning approaches, using non-stairs data (physical/psychological parameters) and stairs data (biomechanical parameters).
Inspired by this study, this research uses two datasets, (i) Non-Stairs Data and (ii) Stairs + Non-Stairs data (subjects independent of hand rails), to predict stair falls for ascent stairs, descent stairs and a combination of the two, using a supervised machine-learning approach. SMOTE, random oversampling and GAN-based oversampling are employed to handle class imbalance. Multiple models including Logistic Regression, XGBoost and Random Forest are compared across multiple data inputs, with performance measured using ROC-AUC and Average Precision. The impact of the number of input features is then explored, and SHAP analysis is used to determine each feature's contribution to the final prediction.
Oversampling to handle imbalanced data was found to be significant for model performance. Adjusting the number of input features improved results. Non-stairs data was effective but constrained by the size of the test data.