Prospective elderly fall prediction by older-adult fall-risk modeling with feature selection. (May 2018)

• Record Type: Journal Article
• Authors: Howcroft, Jennifer; Lemaire, Edward D.; Kofman, Jonathan
• Is Part Of: Biomedical signal processing and control. Volume 43(2018)
• Page Start: 320
• Page End: 328
• Publication Date: 2018-05
• Subjects: AP anterior-posterior -- ASUFSR Arizona State University Feature Selection Repository -- AV all variables -- CFS correlation-based feature selection -- CI confidence intervals -- CoP center of pressure -- CoV coefficient of variation -- DT dual-task -- FCBF fast correlation based filter -- FFTFQ ast Fourier Transform first quartile -- FS feature selection -- H head accelerometer measures -- I pressure-sensing insole measures -- I1 impulse, foot-strike to first peak -- I2 impulse, first peak to minimum -- I3 impulse, minimum to second peak -- I4 impulse, second peak to foot-off -- I5 impulse, foot-strike to minimum -- I6 impulse, minimum to foot-off -- I7 impulse, foot-strike to foot-off -- LS left shank accelerometer measures -- MCC Matthew's Correlation Coefficient -- ML medial-lateral -- MLE maximum Lyapunov exponent -- NB naïve Bayes -- NB-L linear naïve Bayesian -- NN neural network -- NN-a NN with a, the number of nodes in hidden layer -- NPV negative predictive value -- P pelvis accelerometer measures -- PCA principal component analysis -- PPV positive predictive value -- REOH ratio of even to odd harmonics -- RS right shank accelerometer measures -- SD standard deviation -- SR summed rank -- ST single-task gait -- SVM support vector machine -- SVM-b SVM with polynomial degree b; Fall-risk -- Prediction -- Prospective falls -- Elderly -- Wearable sensors -- Feature selection -- Accelerometer -- Plantar pressure; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2018.03.005
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Feature selection improved elderly fall-risk prospective model predictive performance. The best multi-sensor model used insole and left shank accelerometer sensors. The best single-sensor model used the lower-back accelerometer. Optimal features were different between prospective and retrospective fall models. Abstract: This study implemented feature-selection methods for prospective elderly fall-risk prediction and compared modeling outcomes to retrospective fall occurrence classification. Seventy-five elderly adults without prior falls (75.2 (6.6) years; 47 non-fallers, 28 fallers based on 6 month prospective falls) walked 7.62 m while wearing F-Scan insoles and tri-axial accelerometers at the left and right shanks, pelvis, and head. Feature sets were reduced using Relief-F, correlation-based feature selection, and fast correlation based filter algorithms. Naïve Bayesian, multi-layer perceptron neural network, and support vector machine (SVM) classifiers were used for faller prediction. The top twenty feature selection-based models and top ten all-variable-based models determined by single 75:25 train:test stratified holdouts were further examined by training 10, 000 models with randomized 75:25 train:test stratified holdouts. Model evaluation parameters were averaged across all 10, 000 models and then ranked. Feature selection increased predictive accuracy by 9%, comparing models with and without feature selection. The increase in accuracy was greater than for retrospective faller classification using feature selection. Based on the 10, 000 randomized holdouts, the highest ranked model had 65% accuracy and 59% sensitivity, achieved by a pressure-sensing-insole and left-shank-accelerometer based feature subset. The best performance without feature selection was 56% accuracy and 42% sensitivity. Feature selection improved model performance and should therefore be included as a model development step for elderly fall-risk prediction. Some model performance results, such as Relief-F providing the best feature selection and high single-sensor performance using the lower-back accelerometer, were consistent for prospective faller prediction and retrospective classification. However, other contributions to high model performance, such as dual-task assessment for single-sensor models and use of SVM, were specific to older adult fall-risk prediction based on prospective fall occurrence.