Artificial neural network models in the diagnosis of balance impairments

Falls in older population is a major public health issue and tripping is a major cause of falls. The main aim of this research is to explore effectiveness of artificial neural network (ANN) models for automated recognition of gait changes due to falling behaviour. Minimum foot clearance (MFC) during continuous walking on a treadmill was recorded on 10 healthy elderly and 10 elderly with reported balance problem and tripping falls history. MFC histogram characteristic features were used as inputs to a three-layer ANN model with backpropagation error corrections algorithm to build relationships between MFC features and healthy/balance-impaired category. A number of tests were performed to find out optimum architecture of the ANN system, i.e. number of hidden layers and neurons. Cross-validation and Jack-knife techniques were utilized for training the models and subsequently, testing model performance of the trained ANN models. Receiver operating characteristics (ROC) plots, sensitivity and specificity results as well as accuracy rates were used to evaluate performance of the diagnostic model. Test results indicate that ANN architecture with 1 hidden layer consisting of 8 neurons would be sufficient for providing optimal performance. The generalization performance of the diagnostic model was found to be >91.7% (ROC_area >0.94) with respect to its capacity to classify healthy and balance-impaired gait patterns, with sensitivity >0.95 and specificity >0.71. The results of this study demonstrate considerable potential by neural network models in the detection of gait changes in ageing population as a result of balance impairments and falling behavior. Such ANN model may prove to be useful in the clinical/rehabilitation context to categorize normal and pathological gait patterns for diagnosis and also for monitoring improvements.