MOTIVATION: The increasing availability of Electronic Health Record (EHR) systems has created enormous potential for translational research. Recent developments in representation learning techniques have led to effective large-scale representations of EHR concepts along with knowledge graphs that empower downstream EHR studies. However, most existing methods require training with patient-level data, limiting their abilities to expand the training with multi-institutional EHR data. On the other hand, scalable approaches that only require summary-level data do not incorporate temporal dependencies between concepts. METHODS: We introduce a DirectiOnal Medical Embedding (DOME) algorithm to encode temporally directional relationships between medical concepts, using summary-level EHR data. Specifically, DOME first aggregates patient-level EHR data into an asymmetric co-occurrence matrix. Then it computes two Positive Pointwise Mutual Information (PPMI) matrices to correspondingly encode the pairwise prior and posterior dependencies between medical concepts. Following that, a joint matrix factorization is performed on the two PPMI matrices, which results in three vectors for each concept: a semantic embedding and two directional context embeddings. They collectively provide a comprehensive depiction of the temporal relationship between EHR concepts. RESULTS: We highlight the advantages and translational potential of DOME through three sets of validation studies. First, DOME consistently improves existing direction-agnostic embedding vectors for disease risk prediction in several diseases, for example achieving a relative gain of 5.5% in the area under the receiver operating characteristic (AUROC) for lung cancer. Second, DOME excels in directional drug-disease relationship inference by successfully differentiating between drug side effects and indications, correspondingly achieving relative AUROC gain over the state-of-the-art methods by 10.8% and 6.6%. Finally, DOME effectively constructs directional knowledge graphs, which distinguish disease risk factors from comorbidities, thereby revealing disease progression trajectories. The source codes are provided at https://github.com/celehs/Directional-EHR-embedding.

Video-text retrieval is a crucial task in numerous computer vision applications. In this paper, we focus on video-text retrieval involving complex action compositions, where a single video encompasses multiple primitive actions such as "sitting up", "opening door", "cooking food", and "eating." Despite the common occurrences in real-world scenarios, such action-compositional videos have received limited research attention, often leading to significant performance degradations in existing retrieval methods. To address this challenge, we present Hyperbolic Video-tExt Retrieval (HOVER), which models the hierarchical semantic relationships between videos and texts by embedding them in a low-dimensional hyperbolic space. Since hyperbolic space provides a geometric prior that naturally aligns with hierarchical data, it allows for more efficient and generalizable representations of video-text semantic hierarchies. HOVER first longitudinally decomposes each video into a hierarchical action tree, where primitive mono-actions are represented as leaf nodes and increasingly complex action compositions as parent nodes. The semantic structures and temporal dependencies of videos/texts are then encoded in hyperbolic space by exploiting hyperbolic distance, norm, and relative cosine similarity. Experimental results show that HOVER significantly outperforms traditional Euclidean-based methods, particularly in scenarios with limited training labels, achieving a notable performance improvement of 28.83%. Additionally, the hyperbolic video-text embeddings learned by HOVER demonstrate strong generalization across new datasets containing videos with varying levels of action complexity. The source code is available at https://github.com/shi-rq/HOVER.

MOTIVATION: Predicting molecule-disease indications and side effects is important for drug development and pharmacovigilance. Comprehensively mining molecule-molecule, molecule-disease and disease-disease semantic dependencies can potentially improve prediction performance. METHODS: We introduce a Multi-Modal REpresentation Mapping Approach to Predicting molecular-disease relations (M2REMAP) by incorporating clinical semantics learned from electronic health records (EHR) of 12.6 million patients. Specifically, M2REMAP first learns a multimodal molecule representation that synthesizes chemical property and clinical semantic information by mapping molecule chemicals via a deep neural network onto the clinical semantic embedding space shared by drugs, diseases and other common clinical concepts. To infer molecule-disease relations, M2REMAP combines multimodal molecule representation and disease semantic embedding to jointly infer indications and side effects. RESULTS: We extensively evaluate M2REMAP on molecule indications, side effects and interactions. Results show that incorporating EHR embeddings improves performance significantly, for example, attaining an improvement over the baseline models by 23.6% in PRC-AUC on indications and 23.9% on side effects. Further, M2REMAP overcomes the limitation of existing methods and effectively predicts drugs for novel diseases and emerging pathogens. AVAILABILITY AND IMPLEMENTATION: The code is available at https://github.com/celehs/M2REMAP, and prediction results are provided at https://shiny.parse-health.org/drugs-diseases-dev/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Domain adaptation methods reduce domain shift typically by learning domain-invariant features. Most existing methods are built on distribution matching, e.g., adversarial domain adaptation, which tends to corrupt feature discriminability. In this paper, we propose Discriminative Radial Domain Adaptation (DRDA) which bridges source and target domains via a shared radial structure. It's motivated by the observation that as the model is trained to be progressively discriminative, features of different categories expand outwards in different directions, forming a radial structure. We show that transferring such an inherently discriminative structure would enable to enhance feature transferability and discriminability simultaneously. Specifically, we represent each domain with a global anchor and each category a local anchor to form a radial structure and reduce domain shift via structure matching. It consists of two parts, namely isometric transformation to align the structure globally and local refinement to match each category. To enhance the discriminability of the structure, we further encourage samples to cluster close to the corresponding local anchors based on optimal-transport assignment. Extensively experimenting on multiple benchmarks, our method is shown to consistently outperforms state-of-the-art approaches on varied tasks, including the typical unsupervised domain adaptation, multi-source domain adaptation, domain-agnostic learning, and domain generalization.

In this article, we present an analysis of unsupervised domain adaptation with a series of theoretical and algorithmic results. We derive a novel Rényi-$\alpha$divergence-based generalization bound, which is tailored to domain adaptation algorithms with arbitrary loss functions in a stochastic setting. Moreover, our theoretical results provide new insights into the assumptions for successful domain adaptation: the closeness between the conditional distributions of the domains and the Lipschitzness on the source domain. With these assumptions, we reveal the following: if their conditional generation distributions are close, the Lipschitzness property of the target domain can be transferred from the Lipschitzness on the source domain, without knowing the exact target distribution. Motivated by our analysis and assumptions, we further derive practical principles for deep domain adaptation: 1) Rényi-2 adversarial training for marginal distributions matching and 2) Lipschitz regularization for the classifier. Our experimental results on both synthetic and real-world datasets support our theoretical findings and the practical efficiency of the proposed principles.

With the aging of the population and the long-time wearing of earphones, hearing health has gradually emerged as a worldwide health issue. Early detection of hearing health conditions would greatly reduce potential risks with timely medical intervention. This study proposes an earphone-based ear condition monitoring system, named EarHealth, which is low-cost, non-invasive, and easily usable in daily life. It can detect three major hearing health conditions: ruptured eardrum, earwax buildup and blockage, and otitis media. By analyzing the recorded echoes evoked by a chirp sound stimulus, EarHealth recognizes the distinguishable characteristics from ear canal structure and eardrum mobility. EarHealth achieves an accuracy of 82.6% in 92 human subjects, including 27 normal subjects, 22 patients with ruptured eardrum, 25 patients with otitis media, and 18 patients with earwax blockage. EarHealth is the first earphone-based system capable of monitoring hearing health conditions by utilizing the ear canal geometry and eardrum mobility. It is anticipated that EarHealth would provide pervasive and proactive protection for hearing health.

OBJECTIVE: Electronic Health Record (EHR) based phenotyping is a crucial yet challenging problem in the biomedical field. Though clinicians typically determine patient-level diagnoses via manual chart review, the sheer volume and heterogeneity of EHR data renders such tasks challenging, time-consuming, and prohibitively expensive, thus leading to a scarcity of clinical annotations in EHRs. Weakly supervised learning algorithms have been successfully applied to various EHR phenotyping problems, due to their ability to leverage information from large quantities of unlabeled samples to better inform predictions based on a far smaller number of patients. However, most weakly supervised methods are subject to the challenge to choose the right cutoff value to generate an optimal classifier. Furthermore, since they only utilize the most informative features (i.e., main ICD and NLP counts) they may fail for episodic phenotypes that cannot be consistently detected via ICD and NLP data. In this paper, we propose a label-efficient, weakly semi-supervised deep learning algorithm for EHR phenotyping (WSS-DL), which overcomes the limitations above. MATERIALS AND METHODS: WSS-DL classifies patient-level disease status through a series of learning stages: 1) generating silver standard labels, 2) deriving enhanced-silver-standard labels by fitting a weakly supervised deep learning model to data with silver standard labels as outcomes and high dimensional EHR features as input, and 3) obtaining the final prediction score and classifier by fitting a supervised learning model to data with a minimal number of gold standard labels as the outcome, and the enhanced-silver-standard labels and a minimal set of most informative EHR features as input. To assess the generalizability of WSS-DL across different phenotypes and medical institutions, we apply WSS-DL to classify a total of 17 diseases, including both acute and chronic conditions, using EHR data from three healthcare systems. Additionally, we determine the minimum quantity of training labels required by WSS-DL to outperform existing supervised and semi-supervised phenotyping methods. RESULTS: The proposed method, in combining the strengths of deep learning and weakly semi-supervised learning, successfully leverages the crucial phenotyping information contained in EHR features from unlabeled samples. Indeed, the deep learning model's ability to handle high-dimensional EHR features allows it to generate strong phenotype status predictions from silver standard labels. These predictions, in turn, provide highly effective features in the final logistic regression stage, leading to high phenotyping accuracy in notably small subsets of labeled data (e.g. n = 40 labeled samples). CONCLUSION: Our method's high performance in EHR datasets with very small numbers of labels indicates its potential value in aiding doctors to diagnose rare diseases as well as conditions susceptible to misdiagnosis.

Skeleton-based action recognition is widely used in varied areas, e.g., surveillance and human-machine interaction. Existing models are mainly learned in a supervised manner, thus heavily depending on large-scale labeled data, which could be infeasible when labels are prohibitively expensive. In this paper, we propose a novel Contrast-Reconstruction Representation Learning network (CRRL) that simultaneously captures postures and motion dynamics for unsupervised skeleton-based action recognition. It consists of three parts: Sequence Reconstructor (SER), Contrastive Motion Learner (CML), and Information Fuser (INF). SER learns representation from skeleton coordinate sequence via reconstruction. However the learned representation tends to focus on trivial postural coordinates and be hesitant in motion learning. To enhance the learning of motions, CML performs contrastive learning between the representation learned from coordinate sequences and additional velocity sequences, respectively. Finally, in the INF module, we explore varied strategies to combine SER and CML, and propose to couple postures and motions via a knowledge-distillation based fusion strategy which transfers the motion learning from CML to SER. Experimental results on several benchmarks, i.e., NTU RGB+D 60/120, PKU-MMD, CMU, and NW-UCLA, demonstrate the promise of the our method by outperforming state-of-the-art approaches.

Although recent work has achieved great progress in human pose estimation (HPE), most methods show limitations in either inference speed or accuracy. In this paper, we propose a fast and accurate end-to-end HPE method, which is specifically designed to overcome the commonly encountered jitter box, defective box and ambiguous box problems of box-based methods, e.g. Mask R-CNN. Concretely, 1) we propose the ROIGuider to aggregate box instance features from all feature levels under the guidance of global context instance information. Further, 2) the proposed Center Line Branch is equipped with a Dichotomy Extended Area algorithm to adaptively expand each instance box area, and Ambiguity Alleviation strategy to eliminate duplicated keypoints. Finally, 3) to achieve efficient multi-scale feature fusion and real-time inference, we design a novel Trapezoidal Network (TNet) backbone. Experimenting on the COCO dataset, our method achieves 68.1 AP at 25.4 fps, and outperforms Mask-RCNN by 8.9 AP at a similar speed. The competitive performance on the HPE and person instance segmentation tasks over the state-of-the-art models show the promise of the proposed method. The source code will be made available at https://github.com/zlcnup/CGANet.

A variety of computer vision tasks benefit significantly from increasingly powerful deep convolutional neural networks. However, the inherently local property of convolution operations prevents most existing models from capturing long-range feature interactions for improved performances. In this paper, we propose a novel module, called Spatially-Aware Context (SAC) block, to learn spatially-aware contexts by capturing multi-mode global contextual semantics for sophisticated long-range dependencies modeling. We enable customized non-local feature interactions for each spatial position through re-weighted global context fusion in a non-normalized way. SAC is very lightweight and can be easily plugged into popular backbone models. Extensive experiments on COCO, ImageNet, and HICO-DET benchmarks show that our SAC block achieves significant performance improvements over existing baseline architectures while with a negligible computational burden increase. The results also demonstrate the exceptional effectiveness and scalability of the proposed approach on capturing long-range dependencies for object detection, segmentation, and image classification, outperforming a bank of state-of-the-art attention blocks.

Squeeze-and-Excitation (SE) block presents a channel attention mechanism for modeling global context via explicitly capturing dependencies across channels. However, we are still far from understanding how the SE block works. In this work, we first revisit the SE block, and then present a detailed empirical study of the relationship between global context and attention distribution, based on which we propose a simple yet effective module, called Linear Context Transform (LCT) block. We divide all channels into different groups and normalize the globally aggregated context features within each channel group, reducing the disturbance from irrelevant channels. Through linear transform of the normalized context features, we model global context for each channel independently. The LCT block is extremely lightweight and easy to be plugged into different backbone models while with negligible parameters and computational burden increase. Extensive experiments show that the LCT block outperforms the SE block in image classification task on the ImageNet and object detection/segmentation on the COCO dataset with different backbone models. Moreover, LCT yields consistent performance gains over existing state-of-the-art detection architectures, e.g., 1.5∼1.7% APbbox and 1.0%∼1.2% APmask improvements on the COCO benchmark, irrespective of different baseline models of varied capacities. We hope our simple yet effective approach will shed some light on future research of attention-based models.

Unsupervised domain adaptation methods aim to alleviate performance degradation caused by domain-shift by learning domain-invariant representations. Existing deep domain adaptation methods focus on holistic feature alignment by matching source and target holistic feature distributions, without considering local features and their multi-mode statistics. We show that the learned local feature patterns are more generic and transferable and a further local feature distribution matching enables fine-grained feature alignment. In this paper, we present a method for learning domain-invariant local feature patterns and jointly aligning holistic and local feature statistics. Comparisons to the state-of-the-art unsupervised domain adaptation methods on two popular benchmark datasets demonstrate the superiority of our approach and its effectiveness on alleviating negative transfer.

Domain adaptation is an important technique to alleviate performance degradation caused by domain shift, e.g., when training and test data come from different domains. Most existing deep adaptation methods focus on reducing domain shift by matching marginal feature distributions through deep transformations on the input features, due to the unavailability of target domain labels. We show that domain shift may still exist via label distribution shift at the classifier, thus deteriorating model performances. To alleviate this issue, we propose an approximate joint distribution matching scheme by exploiting prediction uncertainty. Specifically, we use a Bayesian neural network to quantify prediction uncertainty of a classifier. By imposing distribution matching on both features and labels (via uncertainty), label distribution mismatching in source and target data is effectively alleviated, encouraging the classifier to produce consistent predictions across domains. We also propose a few techniques to improve our method by adaptively reweighting domain adaptation loss to achieve nontrivial distribution matching and stable training. Comparisons with state of the art unsupervised domain adaptation methods on three popular benchmark datasets demonstrate the superiority of our approach, especially on the effectiveness of alleviating negative transfer.

In recent years, skeleton based action recognition is becoming an increasingly attractive alternative to existing video-based approaches, beneficial from its robust and comprehensive 3D information. In this paper, we explore an unsupervised representation learning approach for the first time to capture the long-term global motion dynamics in skeleton sequences. We design a conditional skeleton inpainting architecture for learning a fixed-dimensional representation, guided by additional adversarial training strategies. We quantitatively evaluate the effectiveness of our learning approach on three well-established action recognition datasets. Experimental results show that our learned representation is discriminative for classifying actions and can substantially reduce the sequence inpainting errors.