JMIR Publications

ABSTRACT

Background:

Fall risk assessment is complex. Based on current scientific evidence, a multifactorial approach including the analysis of physical performance, gait parameters and extrinsic as well as intrinsic risk factors is highly recommended. Using these determinants, a smartphone-based application was designed to assess the individual risk of falling by a score that combines multiple fall risk factors into one comprehensive metric.

Objective:

This study aims to provide a descriptive evaluation of the designed fall risk score as well as an analysis of its discriminative ability based on real-world data.

Methods:

Anonymized data of 242 seniors were analyzed retrospectively. Data were collected between June 2018 and May 2019 using the fall risk assessment app. First, we provide a descriptive statistical analysis of the underlying dataset. Subsequently, multiple learning models (Logistic Regression, Gaussian Naive Bayes, Gradient Boosting, Support Vector Classification and Random Forrest Regression) are trained on the data set to obtain optimal decision boundary. To assess the ability of the fall risk score to discriminate fallers from non-fallers by the fall risk score, mainly the receiver operating curve with its corresponding area under the curve (AUC) and Sensitivity were utilized as performance metrics. For the sake of completeness, specificity, precision and overall accuracy were provided for each model.

Results:

Out of 242 participants with a mean age of 84.6 ± 6.7 years, 139 (57.4%) reported no previous falls (non-faller) versus 103 (43%) that reported a previous fall (faller). The average fall risk was 29.5 ± 12,4 points. The metrics of the Logistic Regression Model were AUC = 0.9; Sensitivity = 100%; Specificity = 52%; Accuracy = 73%. The metrics of the Gaussian Naive Bayes Model were AUC = 0.9; Sensitivity = 100%; Specificity = 52%; Accuracy = 73%. The Metrics of the Gradient Boosting Model were AUC = 0.85; Sensitivity = 88%; Specificity = 62%; Accuracy = 73%. The Metrics of the Support Vector Classification Model are AUC = 0.84; Sensitivity = 88%; Specificity = 67%; Accuracy = 76%. The Metrics of the Random Forrest Model were AUC = 0.84; Sensitivity = 88%; Specificity = 57%; Accuracy = 70%.

Conclusions:

Descriptive statistics of the data set were provided and can be used for comparison and reference values. The fall risk score showed a high discriminative ability to distinguish fallers from non-fallers, irrespective of the evaluated learning model. The models showed an average AUC of 0.8, an average sensitivity of 93% and an average specificity of 58%. Overall average accuracy is 73%. Hence, the investigated fall risk app has the potential to support care takers to easily conduct a valid fall risk assessment. The prospective accuracy of the fall risk score will be further validated in a prospective trial.