Uncertainty-aware probabilistic graph neural networks for road-level traffic crash prediction.

Traffic crashes present substantial challenges to human safety and socio-economic development in urban areas. Developing a reliable and responsible traffic crash prediction model is crucial to address growing public safety concerns and improve the safety of urban mobility systems. Traditional methods face limitations at fine spatiotemporal scales due to the sporadic nature of high-risk crashes and the predominance of non-crash characteristics. Furthermore, while most current models show promising occurrence prediction, they overlook the uncertainties arising from the inherent nature of crashes, and then fail to adequately map the hierarchical ranking of crash risk values for more precise insights. To address these issues, we introduce the Spatiotemporal Zero-Inflated Tweedie Graph Neural Networks (STZITD-GNN), the first uncertainty-aware probabilistic graph deep learning model in road-level daily-basis traffic crash prediction for multi-steps. Our model combines the interpretability of the statistical Tweedie family with the predictive power of graph neural networks, excelling in predicting a comprehensive range of crash risks. The decoder employs a compound Tweedie model, handling the non-Gaussian distribution inherent in crash data, with a zero-inflated component for accurately identifying non-crash cases and low-risk roads. The model accurately predicts and differentiates between high-risk, low-risk, and no-risk scenarios, providing a holistic view of road safety that accounts for the full spectrum of probability and severity of crashes. Empirical tests using real-world traffic data from London, UK, demonstrate that the STZITD-GNN surpasses other baseline models across multiple benchmarks, including a reduction in regression error of up to 34.60% in point estimation metrics and an improvement of above 47% in interval-based uncertainty metrics.

Investigating the Heterogeneity of "Study Twins".

Meta-analyses are commonly performed based on random-effects models, while in certain cases one might also argue in favor of a common-effect model. One such case may be given by the example of two "study twins" that are performed according to a common (or at least very similar) protocol. Here we investigate the particular case of meta-analysis of a pair of studies, for example, summarizing the results of two confirmatory clinical trials in phase III of a clinical development program. Thereby, we focus on the question of to what extent homogeneity or heterogeneity may be discernible and include an empirical investigation of published ("twin") pairs of studies. A pair of estimates from two studies only provide very little evidence of homogeneity or heterogeneity of effects, and ad hoc decision criteria may often be misleading.

Sensitivity analysis for publication bias in meta-analysis of sparse data based on exact likelihood.

Meta-analysis is a powerful tool to synthesize findings from multiple studies. The normal-normal random-effects model is widely used to account for between-study heterogeneity. However, meta-analyses of sparse data, which may arise when the event rate is low for binary or count outcomes, pose a challenge to the normal-normal random-effects model in the accuracy and stability in inference since the normal approximation in the within-study model may not be good. To reduce bias arising from data sparsity, the generalized linear mixed model can be used by replacing the approximate normal within-study model with an exact model. Publication bias is one of the most serious threats in meta-analysis. Several quantitative sensitivity analysis methods for evaluating the potential impacts of selective publication are available for the normal-normal random-effects model. We propose a sensitivity analysis method by extending the likelihood-based sensitivity analysis with the $t$-statistic selection function of Copas to several generalized linear mixed-effects models. Through applications of our proposed method to several real-world meta-analyses and simulation studies, the proposed method was proven to outperform the likelihood-based sensitivity analysis based on the normal-normal model. The proposed method would give useful guidance to address publication bias in the meta-analysis of sparse data.

A comparison of machine learning methods for recovering noisy and missing 4D flow MRI data.

Experimental blood flow measurement techniques are invaluable for a better understanding of cardiovascular disease formation, progression, and treatment. One of the emerging methods is time-resolved three-dimensional phase-contrast magnetic resonance imaging (4D flow MRI), which enables noninvasive time-dependent velocity measurements within large vessels. However, several limitations hinder the usability of 4D flow MRI and other experimental methods for quantitative hemodynamics analysis. These mainly include measurement noise, corrupt or missing data, low spatiotemporal resolution, and other artifacts. Traditional filtering is routinely applied for denoising experimental blood flow data without any detailed discussion on why it is preferred over other methods. In this study, filtering is compared to different singular value decomposition (SVD)-based machine learning and autoencoder-type deep learning methods for denoising and filling in missing data (imputation). An artificially corrupted and voxelized computational fluid dynamics (CFD) simulation as well as in vitro 4D flow MRI data are used to test the methods. SVD-based algorithms achieve excellent results for the idealized case but severely struggle when applied to in vitro data. The autoencoders are shown to be versatile and applicable to all investigated cases. For denoising, the in vitro 4D flow MRI data, the denoising autoencoder (DAE), and the Noise2Noise (N2N) autoencoder produced better reconstructions than filtering both qualitatively and quantitatively. Deep learning methods such as N2N can result in noise-free velocity fields even though they did not use clean data during training. This work presents one of the first comprehensive assessments and comparisons of various classical and modern machine-learning methods for enhancing corrupt cardiovascular flow data in diseased arteries for both synthetic and experimental test cases.

Comparing regularized Kelvinlet functions and the finite element method for registration of medical images to sparse organ data.

Image-guided surgery collocates patient-specific data with the physical environment to facilitate surgical decision making. Unfortunately, these guidance systems commonly become compromised by intraoperative soft-tissue deformations. Nonrigid image-to-physical registration methods have been proposed to compensate for deformations, but clinical utility requires compatibility of these techniques with data sparsity and temporal constraints in the operating room. While finite element models can be effective in sparse data scenarios, computation time remains a limitation to widespread deployment. This paper proposes a registration algorithm that uses regularized Kelvinlets, which are analytical solutions to linear elasticity in an infinite domain, to overcome these barriers. This algorithm is demonstrated and compared to finite element-based registration on two datasets: a phantom liver deformation dataset and an in vivo breast deformation dataset. The regularized Kelvinlets algorithm resulted in a significant reduction in computation time compared to the finite element method. Accuracy as evaluated by target registration error was comparable between methods. Average target registration errors were 4.6 ± 1.0 and 3.2 ± 0.8 mm on the liver dataset and 5.4 ± 1.4 and 6.4 ± 1.5 mm on the breast dataset for the regularized Kelvinlets and finite element method, respectively. Limitations of regularized Kelvinlets include the lack of organ-specific geometry and the assumptions of linear elasticity and infinitesimal strain. Despite limitations, this work demonstrates the generalizability of regularized Kelvinlets registration on two soft-tissue elastic organs. This method may improve and accelerate registration for image-guided surgery, and it shows the potential of using regularized Kelvinlets on medical imaging data.

Advancing urban traffic accident forecasting through sparse spatio-temporal dynamic learning.

Traffic accidents have emerged as one of the most public health safety matters, raising concerns from both the public and urban administrators. The ability to accurately predict traffic accident not only supports the governmental decision-making in advance but also enhances public confidence in safety measures. However, the efficacy of traditional spatio-temporal prediction models are compromised by the skewed distributions and sparse labeling of accident data. To this end, we propose a Sparse Spatio-Temporal Dynamic Hypergraph Learning (SST-DHL) framework that captures higher-order dependencies in sparse traffic accidents by combining hypergraph learning and self-supervised learning. The SST-DHL model incorporates a multi-view spatiotemporal convolution block to capture local correlations and semantics of traffic accidents, a cross-regional dynamic hypergraph learning model to identify global spatiotemporal dependencies, and a two-supervised self-learning paradigm to capture both local and global spatiotemporal patterns. Through experimentation on New York City and London accident datasets, we demonstrate that our proposed SST-DHL exhibits significant improvements compared to optimal baseline models at different sparsity levels. Additionally, it offers enhanced interpretability of results by elucidating complex spatio-temporal dependencies among various traffic accident instances. Our study demonstrates the effectiveness of the SST-DHL framework in accurately predicting traffic accidents, thereby enhancing public safety and trust.

Sharing information across patient subgroups to draw conclusions from sparse treatment networks.

Network meta-analysis (NMA) usually provides estimates of the relative effects with the highest possible precision. However, sparse networks with few available studies and limited direct evidence can arise, threatening the robustness and reliability of NMA estimates. In these cases, the limited amount of available information can hamper the formal evaluation of the underlying NMA assumptions of transitivity and consistency. In addition, NMA estimates from sparse networks are expected to be imprecise and possibly biased as they rely on large-sample approximations that are invalid in the absence of sufficient data. We propose a Bayesian framework that allows sharing of information between two networks that pertain to different population subgroups. Specifically, we use the results from a subgroup with a lot of direct evidence (a dense network) to construct informative priors for the relative effects in the target subgroup (a sparse network). This is a two-stage approach where at the first stage, we extrapolate the results of the dense network to those expected from the sparse network. This takes place by using a modified hierarchical NMA model where we add a location parameter that shifts the distribution of the relative effects to make them applicable to the target population. At the second stage, these extrapolated results are used as prior information for the sparse network. We illustrate our approach through a motivating example of psychiatric patients. Our approach results in more precise and robust estimates of the relative effects and can adequately inform clinical practice in presence of sparse networks.

A transformer-based diffusion probabilistic model for heart rate and blood pressure forecasting in Intensive Care Unit.

BACKGROUND AND OBJECTIVE: Vital sign monitoring in the Intensive Care Unit (ICU) is crucial for enabling prompt interventions for patients. This underscores the need for an accurate predictive system. Therefore, this study proposes a novel deep learning approach for forecasting Heart Rate (HR), Systolic Blood Pressure (SBP), and Diastolic Blood Pressure (DBP) in the ICU. METHODS: We extracted 24,886 ICU stays from the MIMIC-III database which contains data from over 46 thousand patients, to train and test the model. The model proposed in this study, Transformer-based Diffusion Probabilistic Model for Sparse Time Series Forecasting (TDSTF), merges Transformer and diffusion models to forecast vital signs. The TDSTF model showed state-of-the-art performance in predicting vital signs in the ICU, outperforming other models' ability to predict distributions of vital signs and being more computationally efficient. The code is available at https://github.com/PingChang818/TDSTF. RESULTS: The results of the study showed that TDSTF achieved a Standardized Average Continuous Ranked Probability Score (SACRPS) of 0.4438 and a Mean Squared Error (MSE) of 0.4168, an improvement of 18.9% and 34.3% over the best baseline model, respectively. The inference speed of TDSTF is more than 17 times faster than the best baseline model. CONCLUSION: TDSTF is an effective and efficient solution for forecasting vital signs in the ICU, and it shows a significant improvement compared to other models in the field.

The Image-to-Physical Liver Registration Sparse Data Challenge: comparison of state-of-the-art using a common dataset.

Purpose: Computational methods for image-to-physical registration during surgical guidance frequently rely on sparse point clouds obtained over a limited region of the organ surface. However, soft tissue deformations complicate the ability to accurately infer anatomical alignments from sparse descriptors of the organ surface. The Image-to-Physical Liver Registration Sparse Data Challenge introduced at SPIE Medical Imaging 2019 seeks to characterize the performance of sparse data registration methods on a common dataset to benchmark and identify effective tactics and limitations that will continue to inform the evolution of image-to-physical registration algorithms. Approach: Three rigid and five deformable registration methods were contributed to the challenge. The deformable approaches consisted of two deep learning and three biomechanical boundary condition reconstruction methods. These algorithms were compared on a common dataset of 112 registration scenarios derived from a tissue-mimicking phantom with 159 subsurface validation targets. Target registration errors (TRE) were evaluated under varying conditions of data extent, target location, and measurement noise. Jacobian determinants and strain magnitudes were compared to assess displacement field consistency. Results: Rigid registration algorithms produced significant differences in TRE ranging from 3.8±2.4 mm to 7.7±4.5 mm, depending on the choice of technique. Two biomechanical methods yielded TRE of 3.1±1.8 mm and 3.3±1.9 mm, which outperformed optimal rigid registration of targets. These methods demonstrated good performance under varying degrees of surface data coverage and across all anatomical segments of the liver. Deep learning methods exhibited TRE ranging from 4.3±3.3 mm to 7.6±5.3 mm but are likely to improve with continued development. TRE was weakly correlated among methods, with greatest agreement and field consistency observed among the biomechanical approaches. Conclusions: The choice of registration algorithm significantly impacts registration accuracy and variability of deformation fields. Among current sparse data driven image-to-physical registration algorithms, biomechanical simulations that incorporate task-specific insight into boundary conditions seem to offer best performance.

Predicting Medical Event Occurrence Using Medical Insurance Claims Big Data.

Medical events are often infrequent, thus becomes hard to predict. In this paper, we focus on predictor that forecasts whether a medical event would occur in the next year, and analyzes the impact of event's frequency and data size via predictor's performance. In the experiment, we made 1572 predictors for medical events using Medical Insurance Claims (MICs) data from 800,000 participants and 205.8 million claims over 8 years. The result revealed that (a) forecasting error will be increased when predicting low-frequency events, and (b) increasing the number of training dataset reduces errors. This result suggests that increasing data size is a key to solve low frequency problems. However, we still need additional methods to cope with sparse and imbalanced data.

TextNetTopics Pro, a topic model-based text classification for short text by integration of semantic and document-topic distribution information.

With the exponential growth in the daily publication of scientific articles, automatic classification and categorization can assist in assigning articles to a predefined category. Article titles are concise descriptions of the articles' content with valuable information that can be useful in document classification and categorization. However, shortness, data sparseness, limited word occurrences, and the inadequate contextual information of scientific document titles hinder the direct application of conventional text mining and machine learning algorithms on these short texts, making their classification a challenging task. This study firstly explores the performance of our earlier study, TextNetTopics on the short text. Secondly, here we propose an advanced version called TextNetTopics Pro, which is a novel short-text classification framework that utilizes a promising combination of lexical features organized in topics of words and topic distribution extracted by a topic model to alleviate the data-sparseness problem when classifying short texts. We evaluate our proposed approach using nine state-of-the-art short-text topic models on two publicly available datasets of scientific article titles as short-text documents. The first dataset is related to the Biomedical field, and the other one is related to Computer Science publications. Additionally, we comparatively evaluate the predictive performance of the models generated with and without using the abstracts. Finally, we demonstrate the robustness and effectiveness of the proposed approach in handling the imbalanced data, particularly in the classification of Drug-Induced Liver Injury articles as part of the CAMDA challenge. Taking advantage of the semantic information detected by topic models proved to be a reliable way to improve the overall performance of ML classifiers.

Modelling 3D supramolecular structure from sparse single-molecule localisation microscopy data.

Single-molecule localisation microscopy (SMLM) has the potential to reveal the underlying organisation of specific molecules within supramolecular complexes and their conformations, which is not possible with conventional microscope resolution. However, the detection efficiency for fluorescent molecules in cells can be limited in SMLM, even to below 1% in thick and dense samples. Segmentation of individual complexes can also be challenging. To overcome these problems, we have developed a software package termed PERPL: Pattern Extraction from Relative Positions of Localisations. This software assesses the relative likelihoods of models for underlying patterns behind incomplete SMLM data, based on the relative positions of pairs of localisations. We review its principles and demonstrate its use on the 3D lattice of Z-disk proteins in mammalian cardiomyocytes. We find known and novel features at ~20 nm with localisations of less than 1% of the target proteins, using mEos fluorescent protein constructs.

Bayesian heterogeneity in a meta-analysis with two studies and binary data.

The meta-analysis of two trials is valuable in many practical situations, such as studies of rare and/or orphan diseases focussed on a single intervention. In this context, additional concerns, like small sample size and/or heterogeneity in the results obtained, might make standard frequentist and Bayesian techniques inappropriate. In a meta-analysis, moreover, the presence of between-sample heterogeneity adds model uncertainty, which must be taken into consideration when drawing inferences. We suggest that the most appropriate way to measure this heterogeneity is by clustering the samples and then determining the posterior probability of the cluster models. The meta-inference is obtained as a mixture of all the meta-inferences for the cluster models, where the mixing distribution is the posterior model probability. We present a simple two-component form of Bayesian model averaging that is unaffected by characteristics such as small study size or zero-cell counts, and which is capable of incorporating uncertainties into the estimation process. Illustrative examples are given and analysed, using real sparse binomial data.

Triangle-Mesh-Rasterization-Projection (TMRP): An Algorithm to Project a Point Cloud onto a Consistent, Dense and Accurate 2D Raster Image.

The projection of a point cloud onto a 2D camera image is relevant in the case of various image analysis and enhancement tasks, e.g., (i) in multimodal image processing for data fusion, (ii) in robotic applications and in scene analysis, and (iii) for deep neural networks to generate real datasets with ground truth. The challenges of the current single-shot projection methods, such as simple state-of-the-art projection, conventional, polygon, and deep learning-based upsampling methods or closed source SDK functions of low-cost depth cameras, have been identified. We developed a new way to project point clouds onto a dense, accurate 2D raster image, called Triangle-Mesh-Rasterization-Projection (TMRP). The only gaps that the 2D image still contains with our method are valid gaps that result from the physical limits of the capturing cameras. Dense accuracy is achieved by simultaneously using the 2D neighborhood information (rx,ry) of the 3D coordinates in addition to the points P(X,Y,V). In this way, a fast triangulation interpolation can be performed. The interpolation weights are determined using sub-triangles. Compared to single-shot methods, our algorithm is able to solve the following challenges. This means that: (1) no false gaps or false neighborhoods are generated, (2) the density is XYZ independent, and (3) ambiguities are eliminated. Our TMRP method is also open source, freely available on GitHub, and can be applied to almost any sensor or modality. We also demonstrate the usefulness of our method with four use cases by using the KITTI-2012 dataset or sensors with different modalities. Our goal is to improve recognition tasks and processing optimization in the perception of transparent objects for robotic manufacturing processes.

Prediction of fluid flow in porous media by sparse observations and physics-informed PointNet.

We predict steady-state Stokes flow of fluids within porous media at pore scales using sparse point observations and a novel class of physics-informed neural networks, called "physics-informed PointNet" (PIPN). Taking the advantages of PIPN into account, three new features become available compared to physics-informed convolutional neural networks for porous medium applications. First, the input of PIPN is exclusively the pore spaces of porous media (rather than both the pore and grain spaces). This feature diminishes required computer memory. Second, PIPN represents the boundary of pore spaces smoothly and realistically (rather than pixel-wise representations). Third, spatial resolution can vary over the physical domain (rather than equally spaced resolutions). This feature enables users to reach an optimal resolution with a minimum computational cost. The performance of our framework is evaluated by the study of the influence of noisy sensor data, pressure observations, and spatial correlation length.

Using source data to aid and build variational state-space autoencoders with sparse target data for process monitoring.

In industrial processes, different operating conditions and ratios of ingredients are used to produce multi-grade products in the same production line. Yet, the production grade changes so quickly as the demand from customers varies from time to time. As a result, the process data collected in certain operating regions are often scarce. Process dynamics, nonlinearity, and process uncertainty increase the hardship in developing a reliable model to monitor the process status. In this paper, the source-aided variational state-space autoencoder (SA-VSSAE) is proposed. It integrates variational state-space autoencoder with the Gaussian mixture. With the additional information from the source grades, SA-VSSAE can be used for monitoring processes with sparse target data by performing information sharing to enhance the reliability of the target model. Unlike the past works which perform information sharing and modeling in a two-step procedure, the proposed model is designed for information sharing and modeling in a one-step procedure without causing information loss. In contrast to the traditional state-space model, which is linear and deterministic, the variational state-space autoencoder (VSSAE) extracts the dynamic and nonlinear features in the process variables using neural networks. Also, by taking process uncertainty into consideration, VSSAE describes the features in a probabilistic form. Probability density estimates of the residual and latent variables are given to design the monitoring indices for fault detection. A numerical example and an industrial polyvinyl chloride drying process are presented to show the advantages of the proposed method over the comparative methods.

LOCOM: A logistic regression model for testing differential abundance in compositional microbiome data with false discovery rate control.

Compositional analysis is based on the premise that a relatively small proportion of taxa are differentially abundant, while the ratios of the relative abundances of the remaining taxa remain unchanged. Most existing methods use log-transformed data, but log-transformation of data with pervasive zero counts is problematic, and these methods cannot always control the false discovery rate (FDR). Further, high-throughput microbiome data such as 16S amplicon or metagenomic sequencing are subject to experimental biases that are introduced in every step of the experimental workflow. McLaren et al. [eLife 8, e46923 (2019)] have recently proposed a model for how these biases affect relative abundance data. Motivated by this model, we show that the odds ratios in a logistic regression comparing counts in two taxa are invariant to experimental biases. With this motivation, we propose logistic compositional analysis (LOCOM), a robust logistic regression approach to compositional analysis, that does not require pseudocounts. Inference is based on permutation to account for overdispersion and small sample sizes. Traits can be either binary or continuous, and adjustment for confounders is supported. Our simulations indicate that LOCOM always preserved FDR and had much improved sensitivity over existing methods. In contrast, analysis of composition of microbiomes (ANCOM) and ANCOM with bias correction (ANCOM-BC)/ANOVA-Like Differential Expression tool (ALDEx2) had inflated FDR when the effect sizes were small and large, respectively. Only LOCOM was robust to experimental biases in every situation. The flexibility of our method for a variety of microbiome studies is illustrated by the analysis of data from two microbiome studies. Our R package LOCOM is publicly available.

An Automated Scanning Transmission Electron Microscope Guided by Sparse Data Analytics.

Artificial intelligence (AI) promises to reshape scientific inquiry and enable breakthrough discoveries in areas such as energy storage, quantum computing, and biomedicine. Scanning transmission electron microscopy (STEM), a cornerstone of the study of chemical and materials systems, stands to benefit greatly from AI-driven automation. However, present barriers to low-level instrument control, as well as generalizable and interpretable feature detection, make truly automated microscopy impractical. Here, we discuss the design of a closed-loop instrument control platform guided by emerging sparse data analytics. We hypothesize that a centralized controller, informed by machine learning combining limited a priori knowledge and task-based discrimination, could drive on-the-fly experimental decision-making. This platform may unlock practical, automated analysis of a variety of material features, enabling new high-throughput and statistical studies.

Bootstrap aggregated classification for sparse functional data.

Sparse functional data are commonly observed in real-data analyzes. For such data, we propose a new classification method based on functional principal component analysis (FPCA) and bootstrap aggregating. Bootstrap aggregating is believed to improve the single classifier. In this paper, we apply this belief to an FPCA based classification, and compare the classification performance with that of the single classifiers. The simulation results show that the proposed method performs better than the conventional single classifiers. We then conduct two real-data analyzes.

MarZIC: A Marginal Mediation Model for Zero-Inflated Compositional Mediators with Applications to Microbiome Data.

BACKGROUND: The human microbiome can contribute to pathogeneses of many complex diseases by mediating disease-leading causal pathways. However, standard mediation analysis methods are not adequate to analyze the microbiome as a mediator due to the excessive number of zero-valued sequencing reads in the data and that the relative abundances have to sum to one. The two main challenges raised by the zero-inflated data structure are: (a) disentangling the mediation effect induced by the point mass at zero; and (b) identifying the observed zero-valued data points that are not zero (i.e., false zeros). METHODS: We develop a novel marginal mediation analysis method under the potential-outcomes framework to address the issues. We also show that the marginal model can account for the compositional structure of microbiome data. RESULTS: The mediation effect can be decomposed into two components that are inherent to the two-part nature of zero-inflated distributions. With probabilistic models to account for observing zeros, we also address the challenge with false zeros. A comprehensive simulation study and the application in a real microbiome study showcase our approach in comparison with existing approaches. CONCLUSIONS: When analyzing the zero-inflated microbiome composition as the mediators, MarZIC approach has better performance than standard causal mediation analysis approaches and existing competing approach.