Giansanti Daniele, Macellari Velio, Maccioni Giovanni
Dipartimento di Tecnologie e Salute, Istituto Superiore di Sanità, Rome, Italy.
Physiol Meas. 2008 Mar;29(3):N11-9. doi: 10.1088/0967-3334/29/3/N01. Epub 2008 Mar 7.
Fall prevention lacks easy, quantitative and wearable methods for the classification of fall-risk (FR). Efforts must be thus devoted to the choice of an ad hoc classifier both to reduce the size of the sample used to train the classifier and to improve performances. A new methodology that uses a neural network (NN) and a wearable device are hereby proposed for this purpose. The NN uses kinematic parameters assessed by a wearable device with accelerometers and rate gyroscopes during a posturography protocol. The training of the NN was based on the Mahalanobis distance and was carried out on two groups of 30 elderly subjects with varying fall-risk Tinetti scores. The validation was done on two groups of 100 subjects with different fall-risk Tinetti scores and showed that, both in terms of specificity and sensitivity, the NN performed better than other classifiers (naive Bayes, Bayes net, multilayer perceptron, support vector machines, statistical classifiers). In particular, (i) the proposed NN methodology improved the specificity and sensitivity by a mean of 3% when compared to the statistical classifier based on the Mahalanobis distance (SCMD) described in Giansanti (2006 Physiol. Meas. 27 1081-90); (ii) the assessed specificity was 97%, the assessed sensitivity was 98% and the area under receiver operator characteristics was 0.965.
预防跌倒缺乏简单、定量且可穿戴的方法来对跌倒风险（FR）进行分类。因此，必须致力于选择一种专门的分类器，以减少用于训练分类器的样本量并提高性能。为此，本文提出了一种使用神经网络（NN）和可穿戴设备的新方法。该神经网络在姿势描记法协议期间使用由带有加速度计和速率陀螺仪的可穿戴设备评估的运动学参数。神经网络的训练基于马氏距离，并在两组各30名具有不同跌倒风险Tinetti评分的老年受试者上进行。验证是在两组各100名具有不同跌倒风险Tinetti评分的受试者上进行的，结果表明，在特异性和敏感性方面，该神经网络的表现均优于其他分类器（朴素贝叶斯、贝叶斯网络、多层感知器、支持向量机、统计分类器）。特别是，（i）与Giansanti（2006年，《生理测量》27卷，1081 - 1090页）中描述的基于马氏距离的统计分类器（SCMD）相比，所提出的神经网络方法的特异性和敏感性平均提高了3%；（ii）评估的特异性为97%，评估的敏感性为98%，受试者操作特征曲线下面积为0.965。
