Leg agility is a key indicator of bradykinesia, which in turn is a cardinal manifestation of Parkinson’s disease (PD). In fact, automated video assessment of the leg-agility task is critically required for improving the efficiency and objectivity of PD diagnosis. Therefore, we propose a causality-informed graph convolutional network to extract discriminative clinically-meaningful motion features from human skeletons in videos, finally achieving stable leg-agility 5-point scoring. The proposed scheme systematically mines causal features of each skeleton graph from graph node, structure, and representation levels. Specifically, we firstly developed a causality-informed node selection mechanism to mine the graph nodes representing the discriminative features, and thus identify nodes causally correlated to the clinical assessment aspects and suppress the interference from other nodes. Afterwards, a causality-informed structure generation mechanism was designed to generate a graph structure encoding the connections between the discriminative nodes, hence maintaining the discriminability of features associated with these causality-informed nodes. Finally, we employed a clinically-driven self-supervised learning scheme to embed clinical prior knowledge into the proposed model and hence boost the clinical significance of the causality-informed graph nodes, structures, and representations. The proposed method achieved a 71.11% accuracy and a 98.93% acceptable accuracy on a large clinical video dataset. Its effectiveness was also confirmed on an independent test set, and the obtained results exhibited interpretability from modeling and clinical perspectives. In conclusion, our method provides a highly stable scheme for objective video quantification of bradykinesia. Our source code will be released athttps://github.com/SJTUBME-QianLab/PD-CIGCN.

To speed up the efficiency of software development, the ability to retrieve codes through natural language is fundamental. At present, the approach of code search based on deep learning has been extensively researched and achieved a lot of results. However, these models are much complex and the training relies on artificially extracted features. Different from other deep learning models, we simulate people's reading habit of expanding content first and then refining content when learning new knowledge and propose the concept of Extension-Compression Learning. The model can effectively express the features of code and natural language through Extension Learning and Compression Learning. We evaluate the effect of the approach on the code search task with a small dataset and a large dataset, and the results show that all indicators are better than those of other approaches that embed code and text into a joint vector space.

Code search technique has gradually become a key skill to accelerate software development.However, the current deep learning methods only use the encoded results and ignores the original content of the code.Besides, the feature expression of the code is too single, which makes the model's understanding insufficient.And the last problem is the lack of separate processing of core attributes, which will cause the model to lack differentiated learning of the attributes with different importance.Therefore, we propose a residual network based on Multi-Attention, so that the model can not only retain the original content of the code but also allow the code to perform a large number of combined learning in different aspects to obtain differentiated features.Then we treat three core attributes and specific implementation of the code differently so that the model can pay extra attention to the core attributes.We use 158,201 Java code-comment pairs for training.In our experimental results, our model is 9.5% higher than the existing method on the indicator of MRR and 12% higher on the SuccessRate@1.