Density estimation using deep generative neural networks

Density estimation is one of the fundamental problems in both statistics and machine learning. In this study, we propose Roundtrip, a computational framework for general-purpose density estimation based on deep generative neural networks. Roundtrip retains the generative power of deep generative models, such as generative adversarial networks (GANs) while it also provides estimates of density values, thus supporting both data generation and density estimation. Unlike previous neural density estimators that put stringent conditions on the transformation from the latent space to the data space, Roundtrip enables the use of much more general mappings where target density is modeled by learning a manifold induced from a base density (e.g., Gaussian distribution). Roundtrip provides a statistical framework for GAN models where an explicit evaluation of density values is feasible. In numerical experiments, Roundtrip exceeds state-of-the-art performance in a diverse range of density estimation tasks.

Let p(·) be a density on a n-dimensional Euclidean space χ. The task of density estimation is to estimate p(·) based on a set of independently and identically distributed data points {xi}i=1Ndrawn from this density.Traditional density estimators such as histograms (1, 2) and kernel density estimators (KDEs) (3, 4) typically perform well only in low dimension. Recently, neural network-based approaches were proposed for density estimation and yielded promising results in problems with high-dimensional data points such as images. There are mainly two families of such neural density estimators: autoregressive models (5–7) and normalizing flows (8–11). Autoregression-based neural density estimators decompose the density into the product of conditional densities based on probability chain rule p(x)=∏ip(xi|x1:i−1). Each conditional probability p(xi|x1:i−1) is modeled by a parametric density (e.g., Gaussian or mixture of Gaussian), of which the parameters are learned by neural networks. Density estimators based on normalizing flows represent x as an invertible transformation of a latent variable z with known density, where the invertible transformation is a composition of a series of simple functions whose Jacobian is easy to compute. The parameters of these component functions are then learned by neural networks.As suggested in ref. 12, both of these are special cases of the following general framework. Given a differentiable and invertible mapping G:Rn→Rn and a base density pzz, the density of x=G(z) can be represented using the change of variable rule as follows:pxx=pzz|detJz|−1,[1]where Jz=∂Gz/∂zT is the Jacobian matrix of function G(·) at point z. Density estimation at x can be solved if the base density pzz is known and the determinant of Jacobian matrix is feasible to calculate. To achieve this, previous neural density estimators have to impose heavy constraints on the model architecture. For example, refs. 7, 10, and 12 require the Jacobian to be triangular, ref. 13 constructed low rank perturbations of a diagonal matrix as the Jacobian, and ref. 14 proposed a circular convolution where the Jacobian is a circulant matrix. These strong constraints diminish the expressiveness of neural networks, which may lead to poor performance. For example, autoregressive neural density estimators based on learning p(xi|x1:i−1) are naturally sensitive to the order of the features. Moreover, the change of variable rule is not applicable when the domain dimension in base density differs from target density. However, experiences from deep generative models [e.g., GAN (15) and VAE (16)] suggested that it is often desirable to use a latent space of smaller dimension than the data space.To overcome the limitations above, we propose a neural density estimator called Roundtrip. Our approach is motivated by recent advances in deep generative neural networks (15, 17, 18). Roundtrip differs from previous neural density estimators in two ways. 1) It allows the direct use of a deep generative network to model the transformation from the latent variable space to the data space, while previous neural density estimators use neural networks only to learn the parameters in the component functions that are used for building up an invertible transformation. 2) It can efficiently model data densities that are concentrated near learned manifolds, which is difficult to achieve by previous approaches as they require the latent space to have the same dimension as the data space. Importantly, we also provide methods, based on either importance sampling and Laplace approximation, for the pointwise evaluation of the density estimate. We summarize our major contributions in this study as follows: 1) We propose a general-purpose neural density estimator based on deep generative models, which requires less restrictive model assumptions compared to previous neural density estimators. 2) We show that the principle in previous neural density estimators can be regarded as a special case in our Roundtrip framework. 3) We demonstrate state-of-the-art performance of Roundtrip model through a series of experiments, including density estimation tasks in simulations as well as in real data applications ranging from image generation to outlier detection.     

Prevalence of, and risk factors for, hepatitis C virus infection among recent initiates to injecting in London and Glasgow: cross sectional analysis

Our aim was to compare the prevalence of antibody to hepatitis C virus (anti-HCV) among recently initiated injecting drug users (IDUs) in London and Glasgow, and to identify risk factors which could explain differences in prevalence between the cities. Complementary studies of community recruited IDUs who had initiated injection drug use since 1996 were conducted during 2001-2002. Data on HCV risk behaviours were gathered using structured questionnaires with identical core questions and respondents were asked to provide an oral fluid specimen which was tested anonymously for anti-HCV but was linked to the questionnaire. Sensitivities of the anti-HCV assays for oral fluid were 92-96%. Prevalence of anti-HCV was 35% (122/354) in London and 57% (207/366) in Glasgow (P < 0.001). Multifactorially, factors significantly associated with raised odds of anti-HCV positivity were increasing length of injecting career, daily injection, polydrug use, having had a needlestick injury, and having served a prison sentence. In addition lower odds of anti-HCV positivity were associated with non-injection use of crack cocaine and recruitment from drug agencies. After adjustment for these factors, the increased odds of anti-HCV associated with being a Glasgow IDU were diminished but remained significant. HCV continues to be transmitted among the IDU population of both cities at high rates despite the availability of syringe exchange and methadone maintenance. Effectiveness of harm reduction interventions may be compromised by inadequate coverage and failure to reduce sufficiently the frequency of sharing different types of injecting equipment, as well as the high background prevalence of HCV, and its high infectivity. Comprehensive action is urgently required to reduce the incidence of HCV among injectors.     

The use of coroner's autopsy reports to validate the use of targeted swabbing rather than tissue collection for rapid confirmation of virological causes of sudden death in the community

Background and objectivesIn this study, coroner's autopsy reports were used to validate results obtained from respiratory virus screening of swabs rather than tissue collected during autopsy in cases of adult death of unknown cause.Study designCoroner's autopsy samples collected for respiratory virus screening between October 2010 and February 2011, were identified. Autopsy reports were requested from cases positive for a virus. Each report was reviewed to correlate findings at autopsy with the virology result and to determine whether the virus found was listed as a contributing factor in the death.ResultsSixty-four coroner's autopsy cases were identified and a respiratory virus was found in 25 cases. Influenza A(H1N1)pdm09 virus was found most frequently, then RSV and influenza B with a dual influenza A and B infection and a parainfluenza type 1. Where multiple sites were swabbed, the virus was detected in all sites. Autopsy reports for 12 cases were obtained each reporting findings consistent with respiratory infection. Influenza A was always listed as a contributing factor in the death whereas RSV was listed once and influenza B was omitted in one case. The quality of the reports was variable and full histology was less likely to be performed in the elderly.ConclusionsWhile coroner's reports supported the use of swabbing rather than tissue collection, the lack of consistency and omission of the virology findings as contributing factors to death means that the burden of viruses on mortality statistics will remain under-estimated particularly in the elderly.     

彩色多普勒超声帧差伪像的成因与特征探讨

本文目的在于探讨超声帧差伪像的形成原因与声像图特征。通过分析医用超声诊断仪器所采用的设计模型与人体组织声学特性之间存在的矛盾、探头振元数与物理通道数之间的匹配关系、2D图像与彩色多普勒图像在显示屏上的“实时”复合方式,揭示出在图像复合显示过程中,因帧频不同而导致的人体解剖结构与血流信号在时间、空间上“错位”而导致的伪像,阐明它具有的必然性、复合性、隐蔽性、低关注度特点,为仪器设计开发和超声医生在工作中及时识别帧差伪像,减少误诊,提供思路。     

Spatially robust estimates of biological nitrogen (N) fixation imply substantial human alteration of the tropical N cycle

Biological nitrogen fixation (BNF) is the largest natural source of exogenous nitrogen (N) to unmanaged ecosystems and also the primary baseline against which anthropogenic changes to the N cycle are measured. Rates of BNF in tropical rainforest are thought to be among the highest on Earth, but they are notoriously difficult to quantify and are based on little empirical data. We adapted a sampling strategy from community ecology to generate spatial estimates of symbiotic and free-living BNF in secondary and primary forest sites that span a typical range of tropical forest legume abundance. Although total BNF was higher in secondary than primary forest, overall rates were roughly five times lower than previous estimates for the tropical forest biome. We found strong correlations between symbiotic BNF and legume abundance, but we also show that spatially free-living BNF often exceeds symbiotic inputs. Our results suggest that BNF in tropical forest has been overestimated, and our data are consistent with a recent top-down estimate of global BNF that implied but did not measure low tropical BNF rates. Finally, comparing tropical BNF within the historical area of tropical rainforest with current anthropogenic N inputs indicates that humans have already at least doubled reactive N inputs to the tropical forest biome, a far greater change than previously thought. Because N inputs are increasing faster in the tropics than anywhere on Earth, both the proportion and the effects of human N enrichment are likely to grow in the future.Over the last few decades, humans have dramatically altered the global nitrogen (N) cycle (1–3). Three main processes—Haber–Bosch fixation of atmospheric N₂, widespread cultivation of leguminous N-fixing crops, and incidental N fixation during fossil fuel combustion—collectively add more reactive N to the biosphere each year than all natural processes combined (2). Although human perturbation of the N cycle has brought substantial benefits to society (most notably, an increase in crop production) (4), it has also had a number of negative effects on both ecosystems (5, 6) and people (7).Although humanity’s large imprint on the global N cycle is clear, quantifying the extent of anthropogenic changes depends, in large part, on establishing baseline estimates of nonanthropogenic N inputs (1, 8, 9). Before recent human activities, biological N fixation (BNF) was the largest source of new N to the biosphere (9). Terrestrial BNF has been particularly challenging to quantify, because it displays high spatial and temporal heterogeneity at local scales, it arises from both symbiotic associations between bacteria and plants as well as free-living microorganisms (e.g., in leaf litter and soil) (10), and high atmospheric concentrations of N₂ make direct flux measurements unfeasible. Consequently, spatial estimates of BNF have always been highly uncertain (11), and global rate estimates have fallen precipitously in the last 15 y (from 100–290 to ∼44 Tg N y⁻¹) (9). This decline in BNF implies an increase in the relative magnitude of anthropogenic N inputs from 100–150% to 190–470% of BNF (9).Historically, the largest anthropogenic changes to the N cycle have occurred in the northern temperate zone: first throughout the United States and western Europe and more recently, in China (12, 13). Large-scale estimates of BNF in natural ecosystems in these regions are consistently low (11), leading some to conclude that anthropogenic N inputs in the northern temperate zone exceed naturally occurring BNF and preindustrial atmospheric N deposition by an order of magnitude or more (1, 14). By contrast, the highest rates of naturally occurring BNF have been thought to occur in the evergreen lowland tropical rainforest biome (11), implying that, on a regional basis, human alteration of the tropical N cycle has been comparatively modest. However, in recent years, the tropics have seen some of the most dramatic increases in anthropogenic N inputs of any region on Earth—a trend that is likely to continue (2, 6, 13). Anthropogenic N inputs are increasing in tropical regions, primarily because of increasing fossil fuel combustion (13) and expanding high-N-input agriculture for both food and biofuels (6). These anthropogenic N inputs are having a measurable effect on tropical ecosystems (15). However, understanding and forecasting the effects of anthropogenic N depend, in part, on accurate estimates of BNF in lowland tropical rainforest.Unfortunately, the paradigm that the tropics have high rates of BNF is based on a paucity of evidence and several tenuous assumptions. For example, an early global synthesis of terrestrial BNF (11)—which included contributions from both symbiotic and free-living sources—included only one measured estimate of symbiotic BNF from tropical forest (16 kg N ha⁻¹ y⁻¹) (16). That single estimate, scaled over thousands of square kilometers, represented the only direct evidence of high tropical BNF rates available at that time (Fig. 1). Subsequent modeled estimates (17) that indirectly estimated BNF have reinforced the notion that tropical BNF rates are high and dominated by the symbiotic form of fixation (Fig. 1). Such high estimates of symbiotic BNF are consistent with the large number of leguminous trees in tropical forest (18–20). However, many legume species do not form N-fixing nodules (21), and of those species that do, nodulation in individuals varies with soil nutrient status, N demand, and tree age (22). Several recent analyses (10, 22–24) indicate lower tropical forest BNF and suggest that symbiotic BNF may not be as important to total BNF as previously thought (Fig. 1), although few studies have simultaneously measured symbiotic and free-living BNF.Open in a separate windowFig. 1.Previous estimates of BNF in tropical rainforest and BNF measured in this study. Percentages indicate the proportion of total BNF from symbiotic BNF. Cleveland et al. 1999 A (11) is a literature database-derived estimate of tropical forest BNF; Cleveland et al. 1999 B (11) is a modeled estimate of BNF based on the correlation between net primary productivity (NPP) and BNF derived with remotely sensed NPP and evergreen broadleaved forest (EBF) land cover classification. Central estimates and variance for Cleveland et al., 1999 A (11) and Reed et al. 2011 (10) represent the low, central, and high data-based estimates of BNF assuming 5%, 15%, and 15% legume cover, respectively. Central estimates and variance for Wang and Houlton 2009 (17) represent the modeled mean and SD of BNF predicted for the EBF biome. Central estimates and variance for Cleveland et al. 2010 (23) represent the low, central, and high estimates of symbiotic BNF plus free-living BNF or modeled BNF plus free-living BNF. Central estimates and variance for BNF in the four forest ages measured here (primary, 5–15 y, 15–30 y, and 30–50 y) represent means ± 1 SD (n = 3). Our estimate of BNF in a dynamic primary forest (gap dynamics) lacks SD, because it consisted of only two measurements: low and high estimates of forest turnover times equal to 150 and 75 y, respectively.There is also a sound theoretical basis for questioning high estimates of BNF in tropical forest. Namely, high concentrations of soil N in the legume-rich tropics create something of a paradox. Although BNF could create N-rich conditions, the substantial energetic cost of BNF means—and some data show—that BNF should be suppressed under high N availability in primary forests (25). Because of high rates of net primary productivity and high N demand in secondary forests (26, 27), regenerating canopy gaps or abandoned agricultural land may have higher rates of BNF than late-successional forest ecosystems (26).Resolving the uncertainty in the tropical (and global) N cycle requires that we overcome the enduring challenge of quantifying BNF in any ecosystem. How do we estimate large-scale rates of a process that displays extreme spatial heterogeneity at local scales? Whether using acetylene reduction assays, ¹⁵N tracer incubations, or the ¹⁵N natural abundance method, most past approaches to empirically estimate symbiotic BNF have relied on spatial extrapolations of BNF rates measured at the level of individual trees. Typically, such extrapolations are based on legume abundance (e.g., percent cover) and make species- or genera-level assumptions about nodulation status of putative N fixers. Here, we applied a method commonly used by community ecologists to measure rare species abundances—stratified adaptive cluster sampling (SACS) (28)—to measure symbiotic BNF. This approach could be used in any ecosystem, and in contrast to other methods, SACS generates unbiased estimates of mean symbiotic BNF (independent of legume abundance) and can more robustly capture the irregular distribution of nodules on the landscape. We simultaneously measured symbiotic and free-living BNF multiple times over the course of 1 y to generate spatially explicit rates of BNF inputs in primary and secondary (5–50 y old) lowland tropical forest in Costa Rica and then used the understanding gained from those estimates to revisit estimates of BNF and anthropogenic N inputs in the tropical forest biome.     

新形势下医疗器械抽查检验管理规定的修订

目的：探讨医疗器械抽检有关规定的修订要点。方法：对医疗器械监管特别是抽检所面临的新形势进行整理,对现行医疗器械抽检有关规定的不足和问题进行归纳,对修订要点的可行性和影响进行分析。结果：提出了修订《医疗器械抽查检验管理办法》的若干建议。结论：监管机构应当充分认识当前医疗器械监管形势的变化,及时总结经验教训,调整定位并主动作为,将修订医疗器械抽检有关规定提上日程。     

2017-2018年国家医疗器械抽检产品质量状况分析

目的：从国家医疗器械抽检结果出发,总结抽检品种质量状况特点,探讨如何进一步加强医疗器械监督管理。方法：通过对监督抽检发现的医疗器械质量状况存在的问题进行分析研究,提出相应的措施和意见建议。结果：有源医疗器械不合格检出率较高,无源医疗器械的原料选购和工艺存在一定问题,医疗器械领域贯彻执行国家标准、行业标准存在隐忧及一些问题企业。结论：应从落实企业主体责任并强化质量体系建设、严格处理问题产品和企业、提高审核查验针对性和力度、加大审评指导、加强国家标准和行业标准宣贯等角度入手,加强医疗器械监管。     

A seven day actigraphy-based study of rumination and sleep disturbance among young adults with depressive symptoms

Objectives

Trait ruminators exhibit significantly higher levels of sleep disturbance than those without this cognitive vulnerability. However, support for the sleep disruptive effects of state rumination, especially in the pre-sleep period, is rare, and hindered by methodological drawbacks such as self-report and single night assays of sleep. Finally, despite the pervasiveness of the ruminative response style among individuals with depression, the association between rumination and sleep disturbance has not been explored in this population. The present study employed a week-long daily sampling approach to examine the effects of naturally occurring pre-sleep rumination on self-reported and actigraphy-based sleep among individuals with high depressive symptomatology.

Methods

Forty-two university students (19.6 ± 3.2 yo;73.8% female), all of whom reported at least moderate levels of depressive symptoms, completed a short questionnaire after waking each morning for seven days. On this questionnaire, they self-reported sleep indices from the previous night and levels of engagement in pre-sleep rumination. Sleep was also monitored throughout this period via wrist actigraphy. Hierarchical-linear-modeling was used to examine the association between nightly rumination and sleep.

Results

Nightly variations in pre-sleep rumination were predictive of significantly longer actigraphy- and diary-based sleep onset latency (SOL). Notably, a 1 SD increase on the pre-sleep rumination scale was associated with an approximately 7 minute increase in actigraphy-based SOL, even after controlling for baseline sleep disturbance and depressive symptoms.

Conclusions

These data offer compelling evidence for the impact of pre-sleep rumination on sleep onset, providing insight into one potential mechanism that triggers sleep disturbance among individuals with depressive symptoms.     

Non-clinical pharmacokinetics and pharmacodynamics of rVIII-SingleChain,a novel recombinant single-chain factor VIII

Introduction

rVIII-SingleChain (CSL627), a novel recombinant coagulation factor VIII (FVIII), is under investigation in a phase I/III clinical programme (AFFINITY) for the treatment of haemophilia A. Non-clinical studies were conducted to investigate the pharmacokinetic/pharmacodynamic profile of rVIII-SingleChain in comparison with full-length recombinant FVIII.

Materials and Methods

Binding affinity of rVIII-SingleChain for von Willebrand factor was investigated by surface plasmon resonance analysis. The pharmacokinetic profile of rVIII-SingleChain was compared with a marketed full-length recombinant FVIII concentrate (Advate^®) in haemophilia A mice, von Willebrand factor knock-out mice, Crl:CD (SD) rats, rabbits and cynomolgus monkeys. Systemic FVIII activity or antigen levels were recorded. Procoagulant activity was measured in an FeCl₃-induced arterial occlusion model and by recording thrombin generation activity (ex vivo) after administration of 200–250 IU/kg rVIII-SingleChain or full-length FVIII to haemophilia A mice.

Results

rVIII-SingleChain displayed a high affinity for von Willebrand factor (K_D = 44 pM vs. 139 pM for full-length recombinant FVIII). In all animal species tested, rVIII-SingleChain had more favourable pharmacokinetic properties than full-length recombinant FVIII: clearance was decreased and area under the curve and terminal half-life were enhanced vs. full-length recombinant FVIII, while in vivo recovery and volume of distribution were equivalent. rVIII-SingleChain showed a prolonged thrombin generation potential and prolonged procoagulant activity vs. full-length recombinant FVIII in an FeCl₃-induced arterial occlusion model.

Conclusions

rVIII-SingleChain had a higher affinity for von Willebrand factor than full-length recombinant FVIII and displayed favourable pharmacokinetic/pharmacodynamic properties in non-clinical models.