背景:帕金森病(PD)的核心临床体征是运动迟缓,标准测试是手指敲击:临床医生观察到一个人反复敲击手指和拇指。这需要专家的眼光,稀缺资源,甚至专家也显示出可变性和不准确性。现有的手指敲击技术的应用减少了敲击信号的一维措施,具有从这些度量中得出的研究人员定义的特征。
目的:(1)将深度学习神经网络直接应用于手指敲击视频,没有人类定义的度量/特征,并确定特发性PD与对照组的分类准确性。(2)可视化模型学习的特征。
方法:从40名PD患者和37名对照者中收集了152个10秒手指敲击的智能手机视频。我们对像素尺寸进行了下采样,并将视频分为1s片段。在这些剪辑上训练了3D卷积神经网络。
结果:为了区分PD和对照,我们的模型显示训练准确率为0.91,测试准确率为0.69,测试准确率为0.73,测试召回率为0.76,测试AUROC为0.76.我们还报告了五个最具预测性的功能的类激活图。这些显示了网络关注的视频的空间和时间部分,以进行预测,包括PD组明显不同的拇指下降运动。
结论:深度学习神经网络可以直接应用于手指敲击的标准视频,为了区分PD和对照组,不需要从视频中提取一维信号,或预定义的攻丝功能。
BACKGROUND: The core clinical sign of Parkinson\'s disease (PD) is bradykinesia, for which a standard test is finger tapping: the clinician observes a person repetitively tap finger and thumb together. That requires an expert eye, a scarce resource, and even experts show variability and inaccuracy. Existing applications of technology to finger tapping reduce the tapping signal to one-dimensional measures, with researcher-defined features derived from those measures.
OBJECTIVE: (1) To apply a deep learning neural network directly to video of finger tapping, without human-defined measures/features, and determine classification accuracy for idiopathic PD versus controls. (2) To visualise the features learned by the model.
METHODS: 152 smartphone videos of 10s finger tapping were collected from 40 people with PD and 37 controls. We down-sampled pixel dimensions and videos were split into 1 s clips. A 3D convolutional neural network was trained on these clips.
RESULTS: For discriminating PD from controls, our model showed training accuracy 0.91, and test accuracy 0.69, with test precision 0.73, test recall 0.76 and test AUROC 0.76. We also report class activation maps for the five most predictive features. These show the spatial and temporal sections of video upon which the network focuses attention to make a prediction, including an apparent dropping thumb movement distinct for the PD group.
CONCLUSIONS: A deep learning neural network can be applied directly to standard video of finger tapping, to distinguish PD from controls, without a requirement to extract a one-dimensional signal from the video, or pre-define tapping features.