Open video data sharing in developmental science and clinical practice

Abstract

In behavioral research and clinical practice video data has rarely been shared or pooled across sites due to ethical concerns of confidentiality, although the need of shared large-scaled datasets remains increasing. This demand is even more imperative when data-heavy computer-based approaches are involved. To share data while abiding by privacy protection rules, a critical question arises whether efforts at data de-identification reduce data utility? We addressed this question by showcasing an established and video-based diagnostic tool for detecting neurological deficits. We demonstrated for the first time that, for analyzing infant neuromotor functions, pseudonymization by face-blurring video recordings is a viable approach. The redaction did not affect classification accuracy for either human assessors or artificial intelligence methods, suggesting an adequate and easy-to-apply solution for sharing behavioral video data. Our work shall encourage more innovative solutions to share and merge stand-alone video datasets into large data pools to advance science and public health.

Keywords: Clinical neuroscience; Diagnostics; Pediatrics.

Graphical abstract

Figure 1

Classification accuracy Classification accuracy of FM without and with head key points when using network architectures with one and two FC layers, respectively. Error bars denote confidence intervals of mean (95%). Network parameters were as follows. One FC layer without head key points: 64 filters of size 7x1 in the convolutional layer, and 300 neurons in the FC layer; one FC layer with head key points: 64 filters of size 7x1 in the convolutional layer, and 100 neurons in the FC layer; two FC layers without head key points: 64 filters of size 7x1 in the convolutional layer, and 300 and 200 neurons in the FC layers; two FC layers with head key points: 64 filters of size 31x1 in the convolutional layer, and 200 and 100 neurons in the FC layers. No statistically significant differences between means of networks with one FC layer and two layers were found (two-sample t-test, p > 0.05). Also, no statistically significant difference was found when comparing classification accuracy with and without head key points: p = 0.088 and p = 0.6368 for the network with one FC layer and with two FC layers, respectively.

Figure 2

Body key points extraction We used key points 1–5 for face blurring, and key points 1–21 (with head key points) or key points 1–16 (without head key points) for movement classification. Key points 22–25 were not used in this study due to poor position estimation of these key points.

Figure 3

Pseudonymization - Face blurring Flow diagram of face blurring procedure. (A) original image. (B) extracted body key points using OpenPose. (C) obscured face with a blurring mask.

Figure 4

Network architecture Network architecture with one convolutional layer with 64 filters of size 7x1 and two FC layers with 200 and 100 neurons, respectively. 250 corresponds to the number of video frames (5s x 50 frames/s) and 42 corresponds to the number of features with head key points (21 key points with x and y coordinates for each key point). We also used a batch normalization and a dropout of 10% after convolutional and FC layers (not shown).