OBJECTIVE: While timely interventions can improve medication adherence, it is challenging to identify which patients are at risk of nonadherence at point-of-care. We aim to develop and validate flexible machine learning (ML) models to predict a continuous measure of adherence to guideline-directed medication therapies (GDMTs) for heart failure (HF). MATERIALS AND METHODS: We utilized a large electronic health record (EHR) cohort of 34,697 HF patients seen at NYU Langone Health with an active prescription for ≥1 GDMT between April 01, 2021 and October 31, 2022. The outcome was adherence to GDMT measured as proportion of days covered (PDC) at 6 months following a clinical encounter. Over 120 predictors included patient-, therapy-, healthcare-, and neighborhood-level factors guided by the World Health Organization's model of barriers to adherence. We compared performance of several ML models and their ensemble (superlearner) for predicting PDC with traditional regression model (OLS) using mean absolute error (MAE) averaged across 10-fold cross-validation, % increase in MAE relative to superlearner, and predictive-difference across deciles of predicted PDC. RESULTS: Superlearner, a flexible nonparametric prediction approach, demonstrated superior prediction performance. Superlearner and quantile random forest had the lowest MAE (mean [95% CI] = 18.9% [18.7%-19.1%] for both), followed by MAEs for quantile neural network (19.5% [19.3%-19.7%]) and kernel support vector regression (19.8% [19.6%-20.0%]). Gradient boosted trees and OLS were the 2 worst performing models with 17% and 14% higher MAEs, respectively, relative to superlearner. Superlearner demonstrated improved predictive difference. CONCLUSION: This development phase study suggests potential of linked EHR-pharmacy data and ML to identify HF patients who will benefit from medication adherence interventions. DISCUSSION: Fairness evaluation and external validation are needed prior to clinical integration.

Graph convolutional networks (GCNs) have shown success in many graph-based applications as they can combine node features and graph topology to obtain expressive embeddings. While there exist numerous GCN variants, a typical graph convolution layer uses neighborhood aggregation and fully-connected (FC) layers to extract topological and node-wise features, respectively. However, when the receptive field of GCNs becomes larger, the tight coupling between the number of neighborhood aggregation and FC layers can increase the risk of over-fitting. Also, the FC layer between two successive aggregation operations will mix and pollute features in different channels, bringing noise and making node features hard to converge at each channel. In this article, we explore graph convolution without FC layers. We propose scale graph convolution, a new graph convolution using channel-wise scale transformation to extract node features. We provide empirical evidence that our new method has lower over-fitting risk and needs fewer layers to converge. We show from both theoretical and empirical perspectives that models with scale graph convolution have lower computational and memory costs than traditional GCN models. Experimental results on various datasets show that our method can achieve state-of-the-art results, in a cost-effective fashion.

Video summarization is an essential problem in signal processing, which intends to produce a concise summary of the original video. Existing video summarization approaches regard the task as a keyframe selection problem and generally construct the frame-wise representation by combining the long-range temporal dependency with either unimodal or bimodal information. The optimal keyframe should offer the semantic summarization of the whole content by exploiting the multimodal and shot-level hierarchical natures of videos, however, such natures are not fully exploited in existing methods. In this paper, we propose to construct a more powerful and robust frame-wise representation and predict the frame-level importance score in a fair and comprehensive manner. Specifically, we propose a multimodal hierarchical shot-aware convolutional network, denoted as MHSCNet, to enhance the frame-wise representation via combining the comprehensive available multimodal information. We further design a hierarchical ShotConv network to incorporate the adaptive shot-aware frame-level representation by considering the short-range and long-range temporal dependencies. Based on the learned shot-aware representations, MHSCNet can predict the frame-level importance score in the local and global view of the video. Extensive experiments on two standard video summarization datasets demonstrate that our proposed method consistently outperforms state-of-the-arts.

Data sparsity has been a long-standing issue for accurate and trustworthy recommendation systems (RS). To alleviate the problem, many researchers pay much attention to cross-domain recommendation (CDR), which aims at transferring rich knowledge from related source domains to enhance the recommendation performance of sparse target domain. To reach the knowledge transferring purpose, recent CDR works always focus on designing different pairwise directed or undirected information transferring strategies between source and target domains. However, such pairwise transferring idea is difficult to adapt to multi-target CDR scenarios directly, e.g., transferring knowledge between multiple domains and improving their performance simultaneously, as such strategies may lead the following issues: (1) When the number of domains increases, the number of transferring modules will grow exponentially, which causes heavy computation complexity. (2) A single pairwise transferring module could only capture the relevant information of two domains, but ignores the correlated information of other domains, which may limit the transferring effectiveness. (3) When a sparse domain serves as the source domain during the pairwise transferring, it would easily leads the negative transfer problem, and the untrustworthy information may hurt the target domain recommendation performance. In this article, we consider the key challenge of the multi-target CDR task: How to identify the most valuable trustworthy information over multiple domains and transfer such information efficiently to avoid the negative transfer problem? To fulfill the above challenge, we propose a novel end-to-end model termed as DR-MTCDR , standing for D isentangled R epresentations learning for M ulti- T arget CDR . DR-MTCDR aims at transferring the trustworthy domain-shared information across domains, which has the two major advantages in both efficiency and effectiveness: (1) For efficiency, DR-MTCDR utilizes a unified module on all domains to capture disentangled domain-shared information and domain-specific information, which could support all domain recommendation and be insensitive to the number of domains. (2) For effectiveness, based on the disentangled domain-shared and domain-specific information, DR-MTCDR has the capability to lead positive effect and make trustworthy recommendation for each domain. Empirical evaluations on datasets from both public datasets and real-world large-scale financial datasets have shown that the proposed framework outperforms other state-of-the-art baselines.

BACKGROUND: Allergy involves a series of complex reactions and factors that contribute to the development of the disease and triggering of the symptoms, including rhinitis, asthma, atopic eczema, skin sensitivity, even acute and fatal anaphylactic shock. Prediction and evaluation of the potential allergenicity is of importance for safety evaluation of foods and other environment factors. Although several computational approaches for assessing the potential allergenicity of proteins have been developed, their performance and relative merits and shortcomings have not been compared systematically. RESULTS: To evaluate and improve the existing methods for allergen prediction, we collected an up-to-date definitive dataset consisting of 989 known allergens and massive putative non-allergens. The three most widely used allergen computational prediction approaches including sequence-, motif- and SVM-based (Support Vector Machine) methods were systematically compared using the defined parameters and we found that SVM-based method outperformed the other two methods with higher accuracy and specificity. The sequence-based method with the criteria defined by FAO/WHO (FAO: Food and Agriculture Organization of the United Nations; WHO: World Health Organization) has higher sensitivity of over 98%, but having a low specificity. The advantage of motif-based method is the ability to visualize the key motif within the allergen. Notably, the performances of the sequence-based method defined by FAO/WHO and motif eliciting strategy could be improved by the optimization of parameters. To facilitate the allergen prediction, we integrated these three methods in a web-based application proAP, which provides the global search of the known allergens and a powerful tool for allergen predication. Flexible parameter setting and batch prediction were also implemented. The proAP can be accessed at http://gmobl.sjtu.edu.cn/proAP/main.html. CONCLUSIONS: This study comprehensively evaluated sequence-, motif- and SVM-based computational prediction approaches for allergens and optimized their parameters to obtain better performance. These findings may provide helpful guidance for the researchers in allergen-prediction. Furthermore, we integrated these methods into a web application proAP, greatly facilitating users to do customizable allergen search and prediction.