2024 年的期刊文章 · 第 25 页

共 1515 篇,按发表时间排,新的在前,每页 40 篇。这一页是 2024-01-31 到 2024-01-12。

按期刊、按主题看 →

第 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 页

SiM Latent Variables & Mixtures · 发表于 2024-01-31

Safety signal detection with control of latent factors

Xianming Tan, William Wang, Donglin Zeng, Guanghan F. Liu, Guoqing Diao, Niusha Jafari, Ethan M. Alt, Joseph G. Ibrahim

摘要

Postmarket drug safety database like vaccine adverse event reporting system (VAERS) collect thousands of spontaneous reports annually, with each report recording occurrences of any adverse events (AEs) and use of vaccines. We hope to identify signal vaccine‐AE pairs, for which certain vaccines are statistically associated with certain adverse events (AE), using such data. Thus, the outcomes of interest are multiple AEs, which are binary outcomes and could be correlated because they might share certain latent factors; and the primary covariates are vaccines. Appropriately accounting for the complex correlation among AEs could improve the sensitivity and specificity of identifying signal vaccine‐AE pairs. We propose a two‐step approach in which we first estimate the shared latent factors among AEs using a working multivariate logistic regression model, and then use univariate logistic regression model to examine the vaccine‐AE associations after controlling for the latent factors. Our simulation studies show that this approach outperforms current approaches in terms of sensitivity and specificity. We apply our approach in analyzing VAERS data and report our findings.

JASA Genetics & Omics · 发表于 2024-01-31

Quantitative Methods for Precision Medicine: Pharmacogenomics in Action

Arthur Berg

摘要

This book, a valuable addition to the Chapman & Hall/CRC Biostatistics Series, caters to researchers and graduate students with an interest in advanced statistical genetics models and methods. The ... *(abstract via OpenAlex)*

SMMR Diagnostics & Prediction · 发表于 2024-01-31

Optimal design for inference on the threshold of a biomarker

Alessandro Baldi Antognini, Rosamarie Frieri, William F Rosenberger, Maroussa Zagoraiou

摘要

Enrichment designs with a continuous biomarker require the estimation of a threshold to determine the subpopulation benefitting from the treatment. This article provides the optimal allocation for inference in a two-stage enrichment design for treatment comparisons when a continuous biomarker is suspected to affect patient response. Several design criteria, associated with different trial objectives, are optimized under balanced or Neyman allocation and under equality of the first two empirical biomarker’s moments. Moreover, we propose a new covariate-adaptive randomization procedure that converges to the optimum with the fastest available rate. Theoretical and simulation results show that this strategy improves the efficiency of a two-stage enrichment clinical trial, especially with smaller sample sizes and under heterogeneous responses.

JASA Optimization & Learning Theory · 发表于 2024-01-31

Online Statistical Inference for Stochastic Optimization via Kiefer-Wolfowitz Methods

Xi Chen, Zehua Lai, He Li, Yichen Zhang

摘要

This article investigates the problem of online statistical inference of model parameters in stochastic optimization problems via the Kiefer-Wolfowitz algorithm with random search directions. We first present the asymptotic distribution for the Polyak-Ruppert-averaging type Kiefer-Wolfowitz (AKW) estimators, whose asymptotic covariance matrices depend on the distribution of search directions and the function-value query complexity. The distributional result reflects the tradeoff between statistical efficiency and function query complexity. We further analyze the choice of random search directions to minimize certain summary statistics of the asymptotic covariance matrix. Based on the asymptotic distribution, we conduct online statistical inference by providing two construction procedures of valid confidence intervals. Supplementary materials for this article are available online. *(abstract via OpenAlex)*

JRSSB Latent Variables & Mixtures · 发表于 2024-01-31

Full-model estimation for non-parametric multivariate finite mixture models

Marie Du Roy de Chaumaray, Matthieu Marbac

摘要

This paper addresses the problem of full-model estimation for non-parametric finite mixture models. It presents an approach for selecting the number of components and the subset of discriminative variables (i.e. the subset of variables having different distributions among the mixture components) by considering an upper bound on the number of components (this number being allowed to increase with the sample size). The proposed approach considers a discretization of each variable into B bins and a penalization of the resulting log-likelihood. Considering that the number of bins tends to infinity as the sample size tends to infinity, we prove that our estimator of the model (number of components and subset of relevant variables for clustering) is consistent under a suitable choice of the penalty term. The relevance of our proposal is illustrated on simulated and benchmark data.

JASA Networks & Graphs High-Dimensional · 发表于 2024-01-31

Estimating Cell-Type-Specific Gene Co-Expression Networks from Bulk Gene Expression Data with an Application to Alzheimer’s Disease

Chang Su, Jingfei Zhang, Hongyu Zhao

摘要

Inferring and characterizing gene co-expression networks has led to important insights on the molecular mechanisms of complex diseases. Most co-expression analyses to date have been performed on gene expression data collected from bulk tissues with different cell type compositions across samples. As a result, the co-expression estimates only offer an aggregated view of the underlying gene regulations and can be confounded by heterogeneity in cell type compositions, failing to reveal gene coordination that may be distinct across different cell types. In this paper, we introduce a flexible framework for estimating cell-type-specific gene co-expression networks from bulk sample data, without making specific assumptions on the distributions of gene expression profiles in different cell types. We develop a novel sparse least squares estimator, referred to as CSNet, that is efficient to implement and has good theoretical properties. Using CSNet, we analyzed the bulk gene expression data from a cohort study on Alzheimer's disease and identified previously unknown cell-type-specific co-expressions among Alzheimer's disease risk genes, suggesting cell-type-specific disease mechanisms. *(abstract via PubMed)*

SiM Clinical Trials Diagnostics & Prediction · 发表于 2024-01-31

Effective sample size: A measure of individual uncertainty in predictions

Doranne Thomassen, Saskia le Cessie, Hans C. van Houwelingen, Ewout W. Steyerberg

摘要

Clinical prediction models are estimated using a sample of limited size from the target population, leading to uncertainty in predictions, even when the model is correctly specified. Generally, not all patient profiles are observed uniformly in model development. As a result, sampling uncertainty varies between individual patients' predictions. We aimed to develop an intuitive measure of individual prediction uncertainty. The variance of a patient's prediction can be equated to the variance of the sample mean outcome in hypothetical patients with the same predictor values. This hypothetical sample size can be interpreted as the number of similar patients that the prediction is effectively based on, given that the model is correct. For generalized linear models, we derived analytical expressions for the effective sample size. In addition, we illustrated the concept in patients with acute myocardial infarction. In model development, can be used to balance accuracy versus uncertainty of predictions. In a validation sample, the distribution of indicates which patients were more and less represented in the development data, and whether predictions might be too uncertain for some to be practically meaningful. In a clinical setting, the effective sample size may facilitate communication of uncertainty about predictions. We propose the effective sample size as a clinically interpretable measure of uncertainty in individual predictions. Its implications should be explored further for the development, validation and clinical implementation of prediction models.

SMMR Bayesian Methods · 发表于 2024-01-31

Determination of correlations in multivariate count data with informative observation times

Chia-Hui Huang

摘要

We consider there are various types of recurrent events and the total number of occurrences are collected at the random observation times. It has concerned that the observation process may not be independent to the multivariate event processes, hence the total counts and observation times may be correlated and the dependence may exist among different types of the event processes as well. Many methods have developed nonparametric models to accommodate such unknown structures; however, it is difficult to assess and directly quantify their correlation relationships. A multivariate frailty model is proposed to this study, in which the event and observation processes are linked by frailty variables whose joint distribution can be implicitly specified through the multivariate normal distribution with some unknown covariance matrix. The Bayesian inference method is conducted to obtain the estimates of the regression coefficients and correlation parameters. We use a form of trigonometric functions to represent the covariance matrix, so that it meets the positive-definiteness condition efficiently during the estimation schemes. The simulation studies demonstrate the utility of the proposed models. We apply the model to a skin cancer prevention study, and aim to determine the covariate and association effects. We found treatment is significant in determining the duration of examination times; prior-counts, age and gender are significant variables on the occurrence rates of tumor counts. Using the covariance matrix to access the underlying dependent structure, the mutual correlations among them are all positive, and the basal cell counts are more related to the examination times.

SMMR Survival Analysis · 发表于 2024-01-31

Cure models with adaptive activation for modeling cancer survival

Qi Jiang, Sanjib Basu

摘要

We propose a class of cure rate models motivated by analysis of colon cancer and triple-negative breast cancer survival data. This class is indexed by an adaptive activation parameter and a function. We establish that the class is stochastically ordered in the activation parameter and also establish two identifiability results for this class. The first- and last-activation models are members of this class whereas many cure rate models proposed in the literature are also part of this class. We illustrate that while first- and last-activation models may perform poorly under model misspecifications, the proposed model with adaptive activation provides appropriate inference in these cases. We apply the proposed approach to assess treatment-sex interaction on cure rate in a colon cancer study and to assess role of tumor heterogeneity and ethnic disparity in breast cancer.

JASA Survey & Sampling · 发表于 2024-01-31

Controlled Epidemiological Studies

Kaushik Ghosh

摘要

This book presents a variety of sampling designs used to collect data for epidemiological studies. The designs range from well-known commonly used techniques, to the less familiar specialized techn... *(abstract via OpenAlex)*

SMMR Diagnostics & Prediction Clinical Trials Longitudinal Data · 发表于 2024-01-31

A statistical framework for planning and analysing test–retest studies of repeatability

Moritz Fabian Danzer, Maria Eveslage, Dennis Görlich, Benjamin Noto

摘要

There is an increasing number of potential quantitative biomarkers that could allow for early assessment of treatment response or disease progression. However, measurements of such biomarkers are subject to random variability. Hence, differences of a biomarker in longitudinal measurements do not necessarily represent real change but might be caused by this random measurement variability. Before utilizing a quantitative biomarker in longitudinal studies, it is therefore essential to assess the measurement repeatability. Measurement repeatability obtained from test–retest studies can be quantified by the repeatability coefficient, which is then used in the subsequent longitudinal study to determine if a measured difference represents real change or is within the range of expected random measurement variability. The quality of the point estimate of the repeatability coefficient, therefore, directly governs the assessment quality of the longitudinal study. Repeatability coefficient estimation accuracy depends on the case number in the test–retest study, but despite its pivotal role, no comprehensive framework for sample size calculation of test–retest studies exists. To address this issue, we have established such a framework, which allows for flexible sample size calculation of test–retest studies, based upon newly introduced criteria concerning assessment quality in the longitudinal study. This also permits retrospective assessment of prior test–retest studies.

SiM Diagnostics & Prediction Nonparametric & Semiparametric · 发表于 2024-01-30

Likelihood ratio combination of multiple biomarkers via smoothing spline estimated densities

Zhiyuan Du, Pang Du, Aiyi Liu

摘要

The diagnostic accuracy of multiple biomarkers in medical research is crucial for detecting diseases and predicting patient outcomes. An optimal method for combining these biomarkers is essential to maximize the Area Under the Receiver Operating Characteristic (ROC) Curve (AUC). Although the optimality of the likelihood ratio has been proven by Neyman and Pearson, challenges persist in estimating the likelihood ratio, primarily due to the estimation of multivariate density functions. In this study, we propose a non‐parametric approach for estimating multivariate density functions by utilizing Smoothing Spline density estimation to approximate the full likelihood function for both diseased and non‐diseased groups, which compose the likelihood ratio. Simulation results demonstrate the efficiency of our method compared to other biomarker combination techniques under various settings for generated biomarker values. Additionally, we apply the proposed method to a real‐world study aimed at detecting childhood autism spectrum disorder (ASD), showcasing its practical relevance and potential for future applications in medical research.

Biometrics Survival Analysis · 发表于 2024-01-29

Principal stratification analysis of noncompliance with time-to-event outcomes

Bo Liu, Lisa Wruck, Fan Li

SiM Diagnostics & Prediction Nonparametric & Semiparametric · 发表于 2024-01-29

Nonparametric estimation of linear personalized diagnostics rules via efficient grid algorithm

Yaliang Zhang, Yunro Chung

摘要

Many diseases are heterogeneous, comprised of multiple disease subgroups. It is of great interest but highly unlikely to find a single biomarker that can accurately detect such heterogeneous diseases across different subgroups. In this article, we propose to estimate a personalized diagnostic rule (PDR) to tailor more effective biomarkers to each individual according to a linear combination of his or her profiles. A standard grid search algorithm can be used to estimate the optimal linear PDR that maximizes the area under the receiver operating characteristics curve (AUC) among all the linear PDRs, but it is time‐consuming especially when the number of variables is large. Alternatively, we developed an efficient grid rotation algorithm that provides a nearly suboptimal solution and studied its variation to find the optimal solution. We implemented the cross‐validated forward variable selection method to find a subset of useful variables while avoid overfitting. Extensive simulations show that our proposed method reduces bias and variance. Analysis of a gastric cancer biomarker study and censored survival outcome data illustrates the practical utility of our proposed method. The proposed method is implemented in the open‐source R package persDx.

Biometrics Missing Data & Measurement Error Genetics & Omics Optimization & Learning Theory · 发表于 2024-01-29

Multiple augmented reduced rank regression for pan-cancer analysis

Jiuzhou Wang, Eric F Lock

摘要

Statistical approaches that successfully combine multiple datasets are more powerful, efficient, and scientifically informative than separate analyses. To address variation architectures correctly and comprehensively for high-dimensional data across multiple sample sets (ie, cohorts), we propose multiple augmented reduced rank regression (maRRR), a flexible matrix regression and factorization method to concurrently learn both covariate-driven and auxiliary structured variations. We consider a structured nuclear norm objective that is motivated by random matrix theory, in which the regression or factorization terms may be shared or specific to any number of cohorts. Our framework subsumes several existing methods, such as reduced rank regression and unsupervised multimatrix factorization approaches, and includes a promising novel approach to regression and factorization of a single dataset (aRRR) as a special case. Simulations demonstrate substantial gains in power from combining multiple datasets, and from parsimoniously accounting for all structured variations. We apply maRRR to gene expression data from multiple cancer types (ie, pan-cancer) from The Cancer Genome Atlas, with somatic mutations as covariates. The method performs well with respect to prediction and imputation of held-out data, and provides new insights into mutation-driven and auxiliary variations that are shared or specific to certain cancer types.

Biometrics Networks & Graphs Latent Variables & Mixtures Bayesian Methods · 发表于 2024-01-29

Incorporating graph information in Bayesian factor analysis with robust and adaptive shrinkage priors

Qiyiwen Zhang, Changgee Chang, Li Shen, Qi Long

摘要

There has been an increasing interest in decomposing high-dimensional multi-omics data into a product of low-rank and sparse matrices for the purpose of dimension reduction and feature engineering. Bayesian factor models achieve such low-dimensional representation of the original data through different sparsity-inducing priors. However, few of these models can efficiently incorporate the information encoded by the biological graphs, which has been already proven to be useful in many analysis tasks. In this work, we propose a Bayesian factor model with novel hierarchical priors, which incorporate the biological graph knowledge as a tool of identifying a group of genes functioning collaboratively. The proposed model therefore enables sparsity within networks by allowing each factor loading to be shrunk adaptively and by considering additional layers to relate individual shrinkage parameters to the underlying graph information, both of which yield a more accurate structure recovery of factor loadings. Further, this new priors overcome the phase transition phenomenon, in contrast to existing graph-incorporated approaches, so that it is robust to noisy edges that are inconsistent with the actual sparsity structure of the factor loadings. Finally, our model can handle both continuous and discrete data types. The proposed method is shown to outperform several existing factor analysis methods through simulation experiments and real data analyses.

Biometrics Survival Analysis · 发表于 2024-01-29

Efficient estimation for left-truncated competing risks regression for case-cohort studies

Xi Fang, Kwang Woo Ahn, Jianwen Cai, Soyoung Kim

摘要

The case-cohort study design provides a cost-effective study design for a large cohort study with competing risk outcomes. The proportional subdistribution hazards model is widely used to estimate direct covariate effects on the cumulative incidence function for competing risk data. In biomedical studies, left truncation often occurs and brings extra challenges to the analysis. Existing inverse probability weighting methods for case-cohort studies with competing risk data not only have not addressed left truncation, but also are inefficient in regression parameter estimation for fully observed covariates. We propose an augmented inverse probability-weighted estimating equation for left-truncated competing risk data to address these limitations of the current literature. We further propose a more efficient estimator when extra information from the other causes is available. The proposed estimators are consistent and asymptotically normally distributed. Simulation studies show that the proposed estimator is unbiased and leads to estimation efficiency gain in the regression parameter estimation. We analyze the Atherosclerosis Risk in Communities study data using the proposed methods.

JRSSB Spatial & Time Series Machine Learning Networks & Graphs · 发表于 2024-01-29

Authors’ reply to the Discussion of ‘Automatic change-point detection in time series via deep learning’ at the Discussion Meeting on ‘Probabilistic and statistical aspects of machine learning’

Jie Li, Paul Fearnhead, Piotr Fryzlewicz, Tengyao Wang

摘要

We would like to thank the proposer, seconder, and all discussants for their time in reading our article and their thought-provoking comments. We are glad to find a broad consensus that neural-network-based approach offers a flexible framework for automatic change-point analysis. There are a number of common themes to the comments, and we have therefore structured our response around the topics of the theory, training, the importance of standardization and possible extensions, before addressing some of the remaining individual comments. Both Wilkinson and Zhang compare the theoretical bound on the generalization error in Theorem 4.2 with our empirical results in Figure 2 of main text, Figures S2 and S3 of supplement. Our experience has been that the empirical generalization error has been substantially lower than the theoretical bounds suggest—with good performance of our fitted neural network with training sample sizes that are orders of magnitude smaller than one may expect given the number of parameters within the neural network. Related to this is that how the bound on the generalization error depends on e.g. the number of layers, is not particularly informative about how these factors affect the error in practice. We agree with both Wilkinson and Zhang that there is a need for a new theoretical framework for bounding the generalization error of neural networks that is more meaningful in practice. We thank Zhang for pointing out how our theoretical analysis can be extended to more general data generating mechanisms, including heavier-than-Gaussian noise distributions and data with weak temporal correlation (a concern of Hong et al., 2024). Indeed, as Zhang comments, the same procedure still works in such settings and the current proof will go through, with minor modifications to the choice of λ in Corollary 4.1. As λ is adaptively chosen by the neural network (which is one of the main attractions of our procedure), the results in Theorems 4.2 and 4.3 will be essentially unchanged. In a similar vein, Schmidt-Hieber mentions that our theory could be modified to prove local change-point detection error rates using convolutional neural networks (CNN). Indeed, by combining existing results on moving sum (MOSUM) (Eichinger & Kirch, 2018) together with Vapnik-Chervonenkis (VC) dimension results of CNN, we could arrive at a similar result to Theorems 4.2 and 4.3 in the paper. Gavioli-Akilagun points out that the current architecture cannot directly exactly represent the likelihood-ratio test statistic for detection of piecewise affine changes. We agree with this observation. However, by including squared observations as inputs, or if we include squaring in the set of nonlinearities permitted by the neural network, we can directly express the likelihood-ratio test as a function in the neural network class (VC-dimension results concerning neural networks with piecewise polynomial activation functions are given in Theorem 7 of Bartlett et al., 2019). Moreover, by including multiple layers within our architecture, the neural network can accurately approximate the square function. This is related to the nice results of Gavioli-Akilagun that show test statistics based on linear functions, which can be represented simply by a neural network, can have statistical performance comparably to those based on the likelihood-ratio test. Regarding Zhang’s observation that the first term in Theorem 4.3 does not decrease as training sample size N increases, our intuition is that the first term represents the Bayes risk of the classification task on the test set, which is achieved by the cumulative sum based classifier. Therefore, it will not be affected by increasing training sample size and only depends on the test sample. Both Wilkinson and Nemeth raise the question of how to choose a suitable neural network architecture for different settings. This is a well-studied problem in machine learning. A few possible Neural Architecture Search approaches are mentioned in Paaß and Giesselbach (2023, Section 2.4.3) (also discussed in Section 6 of the main text). Nemeth also asks the question of whether it is easier to choose the neural network architecture than to choose stochastic models to represent the data. Given these Neural Architecture Search methods mentioned above, we believe that the former is at least a more structured problem that can be solved algorithmically. Schmidt-Hieber discusses the effect of the depth of the network on its ability to detect various structures in the signal. Indeed, we agree that in general the less we know about the data generating mechanism, the more layers we need in the architecture. Cribben and Anastasiou raise a number of important practical considerations with training. First, as our approach requires labelled data, there is the challenge of obtaining such data. We agree that for manually labelled data, the labelling of changes is subjective, and this means that our method will only aim to replicate the manual classification. If we use simulation, then we can avoid this, but at the expense of needing to model what changes would look like. They also point out that often changes are rare—so training data may be imbalanced, and we would also want to account for this imbalance when detecting new changes. If we believe the frequency of changes will be different in the training than in the test data, or if we wish to account differently for different types of errors (false detection versus missing a true change), this is possible by including different weights for each error when training. Both Bodenham and Adams, and Hong et al., ask how our approach could be applied to a single long time series, for example from finance. First, our method requires labelled training data, so we would require part of the time series to have labelled changes, or known to have no changes. In this case, we can divide such historical data into time series of a fixed length, leading to a set of labelled training data. One can then fit a neural network classifier to this data and use the fitted neural network to classify windows of new data using the same moving window idea as in Section 6. When not enough training data is available, we proposed to simulate artificial data to train our neural network. Wilkinson points out that this has a close link to the simulation-based inference. We agree that it is worth investigating the links with this literature further. We agree with Chen and Chen that the simple neural network classifier is not automatically invariant to the shifting and scaling because it may not have learned an exactly invariant statistic. Differences in the results for standardized versus nonstandardized data show that the algorithm (not unexpectedly) learns to perform classification for the problem at hand, whether or not it aligns with the analyst’s perceived best way of distinguishing the two classes. More specifically, the algorithm’s task is to solve a binary classification problem, distinguishing between two groups of sequences. While the analyst may suspect that it is the presence of the change point that is the main feature separating the two classes, the learning algorithm may take a different view given the training data. For example, in the nonstandardized case, if all the input data starts with μL=0⁠, what the analyst regards as a change-point problem the algorithm may construe as a testing problem of departure from a zero mean. Furthermore, as pointed out by Wilkinson, the min–max scaling used in our algorithm may not be appropriate for very heavy-tailed data. In such contexts, scaling by empirical quantiles other than 0 and 1 would be more appropriate. In many applications, one would also like the test to be invariant to the reversal of the time direction. Hence, Chen and Chen’s proposal of adding reversed sequence Xn,…,X1 into the training data, which has the additional benefit of enlarging the sample size, makes sense. As Bodenham and Adams point out, often one is interested in localizing changes, rather than just detecting them. As a first work in the area of using neural networks to automatically construct change-point detectors, we deliberately focused on the problem of detection rather than localization. However frameworks similar to that of MOSUM (Eichinger & Kirch, 2018), where we apply a detector to different windows of data, can use a change-point detector to also estimate the location of any changes. We used this idea, i.e. Algorithm 1, in the analysis of the Human Activity Sensing Consortium data in Section 6. However, we believe that there may be more attractive ways of extending our idea to the problem of change-point localization. A number of discussants (Anastasiou and Cribben, Schoenberg and Wong, Bodenham and Adams) ask whether our method can be applied in an online setting. This is possible provided that we have a pretrained neural network that can classify data of a given window length, say h. When we receive a new observation, we can apply the neural network classifier to the most recent h time points. Computationally this is possible in an online setting, as once trained, the cost of running the classifier is fixed. There are challenges in terms of how to tune the classifier so that the resulting online change-point detector has an appropriate average run length. Also, how to extend this idea so that we also update the classifier online as we get new data, is an interesting open question. Schoenberg and Wong, Anastasiou and Cribben, Mateu, van Lieshout and Lu ask whether our ideas could be extended to multivariate data, and in particular spatial data or point process data. We are interested to hear about the recent developments in this area and look forward to seeing future works in this direction. One challenge of dealing with point process data is that the number of points is random and cannot be easily interpreted as input of a neural network. As a first order approximation, we could bin the data into a (multivariate) time series of counts and a similar method to the one proposed in our article could then be applied. We are interested to read about other research at the interface between neural networks and change point and related areas in statistics. This includes the using change-point detection to improve the fitting of deep neural network models to time-series data (Jungbluth & Lederer, 2023) and the possibility of using ideas from our article to improve existing change-point detection methods as suggested by Ombao. Schmidt-Hieber also points out the possibility of using a convolutional neural network-based approach for detecting local change points or two-dimensional edge detection. Finally, we agree with MacKenzie that one disadvantage with automated procedures like the one in our article is that the final test statistic for a change is hard to interpret. We welcome work on improving interpretability of artificial intelligence (AI) and believe ideas in this area will be important as AI methods are increasingly using within statistics. *(abstract via OpenAlex)*

SiM Diagnostics & Prediction · 发表于 2024-01-29

Addressing subject heterogeneity in time‐dependent discrimination for biomarker evaluation

Xinyang Jiang, Wen Li, Ruosha Li, Jing Ning

摘要

Accurate discrimination has been the central goal in identifying biomarkers for monitoring disease progression and early detection. Acknowledging the fact that discrimination accuracy of biomarkers for a time‐to‐event outcome often changes over time, local measures such as the time‐dependent receiver operating characteristic curve and its area under the curve (AUC) are used to assess time‐dependent predictive discrimination. However, such measures do not address subject heterogeneity, although the impact of covariates including demographics, disease‐related characteristics, and other clinical information on the discriminatory performance of biomarkers needs to be investigated before their clinical use. We propose the covariate‐specific time‐dependent AUC, a measure for covariate‐adjusted discrimination. We develop a regression model on the covariate‐specific time‐dependent AUC to understand how and in what magnitude the covariates influence biomarker performance. Then we construct a pseudo partial‐likelihood for estimation and inference. This is followed by our establishing the asymptotic properties of the proposed estimators and provide variance estimation. The simulation studies and application to the AIDS Clinical Trials Group 175 data demonstrate that the proposed method offers an informative tool for inferring covariate‐specific and time‐dependent predictive discrimination.

Biometrics Networks & Graphs Causal Inference · 发表于 2024-01-29

Adaptive sequential surveillance with network and temporal dependence

Ivana Malenica, Jeremy R Coyle, Mark J van der Laan, Maya L Petersen

摘要

Strategic test allocation is important for control of both emerging and existing pandemics (eg, COVID-19, HIV). It supports effective epidemic control by (1) reducing transmission via identifying cases and (2) tracking outbreak dynamics to inform targeted interventions. However, infectious disease surveillance presents unique statistical challenges. For instance, the true outcome of interest (positive infection status) is often a latent variable. In addition, presence of both network and temporal dependence reduces data to a single observation. In this work, we study an adaptive sequential design, which allows for unspecified dependence among individuals and across time. Our causal parameter is the mean latent outcome we would have obtained, if, starting at time t given the observed past, we had carried out a stochastic intervention that maximizes the outcome under a resource constraint. The key strength of the method is that we do not have to model network and time dependence: a short-term performance Online Super Learner is used to select among dependence models and randomization schemes. The proposed strategy learns the optimal choice of testing over time while adapting to the current state of the outbreak and learning across samples, through time, or both. We demonstrate the superior performance of the proposed strategy in an agent-based simulation modeling a residential university environment during the COVID-19 pandemic.

SiM Longitudinal Data · 发表于 2024-01-27

Classified functional mixed effects model prediction

Xiaoyan Liu, Jiming Jiang

摘要

In nowadays biomedical research, there has been a growing demand for making accurate prediction at subject levels. In many of these situations, data are collected as longitudinal curves and display distinct individual characteristics. Thus, prediction mechanisms accommodated with functional mixed effects models (FMEM) are useful. In this paper, we developed a classified functional mixed model prediction (CFMMP) method, which adapts classified mixed model prediction (CMMP) to the framework of FMEM. Performance of CFMMP against functional regression prediction based on simulation studies and the consistency property of CFMMP estimators are explored. Real‐world applications of CFMMP are illustrated using real world examples including data from the hormone research menstrual cycles and the diffusion tensor imaging.

SiM Longitudinal Data Diagnostics & Prediction Survival Analysis · 发表于 2024-01-25

Optimizing dynamic predictions from joint models using super learning

Dimitris Rizopoulos, Jeremy M. G. Taylor

摘要

Joint models for longitudinal and time‐to‐event data are often employed to calculate dynamic individualized predictions used in numerous applications of precision medicine. Two components of joint models that influence the accuracy of these predictions are the shape of the longitudinal trajectories and the functional form linking the longitudinal outcome history to the hazard of the event. Finding a single well‐specified model that produces accurate predictions for all subjects and follow‐up times can be challenging, especially when considering multiple longitudinal outcomes. In this work, we use the concept of super learning and avoid selecting a single model. In particular, we specify a weighted combination of the dynamic predictions calculated from a library of joint models with different specifications. The weights are selected to optimize a predictive accuracy metric using V‐fold cross‐validation. We use as predictive accuracy measures the expected quadratic prediction error and the expected predictive cross‐entropy. In a simulation study, we found that the super learning approach produces results very similar to the Oracle model, which was the model with the best performance in the test datasets. All proposed methodology is implemented in the freely available R package JMbayes2.

SiM Causal Inference · 发表于 2024-01-25

Comparison of methods that combine multiple randomized trials to estimate heterogeneous treatment effects

Carly Lupton Brantner, Trang Quynh Nguyen, Tengjie Tang, Congwen Zhao, Hwanhee Hong, Elizabeth A. Stuart

摘要

Individualized treatment decisions can improve health outcomes, but using data to make these decisions in a reliable, precise, and generalizable way is challenging with a single dataset. Leveraging multiple randomized controlled trials allows for the combination of datasets with unconfounded treatment assignment to better estimate heterogeneous treatment effects. This article discusses several nonparametric approaches for estimating heterogeneous treatment effects using data from multiple trials. We extend single‐study methods to a scenario with multiple trials and explore their performance through a simulation study, with data generation scenarios that have differing levels of cross‐trial heterogeneity. The simulations demonstrate that methods that directly allow for heterogeneity of the treatment effect across trials perform better than methods that do not, and that the choice of single‐study method matters based on the functional form of the treatment effect. Finally, we discuss which methods perform well in each setting and then apply them to four randomized controlled trials to examine effect heterogeneity of treatments for major depressive disorder.

SMMR Clinical Trials · 发表于 2024-01-24

A comparison of model-free phase I dose escalation designs for dual-agent combination therapies

Helen Barnett, Matthew George, Donia Skanji, Gaelle Saint-Hilary, Thomas Jaki, Pavel Mozgunov

摘要

It is increasingly common for therapies in oncology to be given in combination. In some cases, patients can benefit from the interaction between two drugs, although often at the risk of higher toxicity. A large number of designs to conduct phase I trials in this setting are available, where the objective is to select the maximum tolerated dose combination. Recently, a number of model-free (also called model-assisted) designs have provoked interest, providing several practical advantages over the more conventional approaches of rule-based or model-based designs. In this paper, we demonstrate a novel calibration procedure for model-free designs to determine their most desirable parameters. Under the calibration procedure, we compare the behaviour of model-free designs to model-based designs in a comprehensive simulation study, covering a number of clinically plausible scenarios. It is found that model-free designs are competitive with the model-based designs in terms of the proportion of correct selections of the maximum tolerated dose combination. However, there are a number of scenarios in which model-free designs offer a safer alternative. This is also illustrated in the application of the designs to a case study using data from a phase I oncology trial.

SMMR Survival Analysis · 发表于 2024-01-23

Regression analysis of multivariate recurrent event data allowing time-varying dependence with application to stroke registry data

Wen Li, Mohammad H. Rahbar, Sean I. Savitz, Jing Zhang, Sori Kim Lundin, Amirali Tahanan, Jing Ning

摘要

In multivariate recurrent event data, each patient may repeatedly experience more than one type of event. Analysis of such data gets further complicated by the time-varying dependence structure among different types of recurrent events. The available literature regarding the joint modeling of multivariate recurrent events assumes a constant dependency over time, which is strict and often violated in practice. To close the knowledge gap, we propose a class of flexible shared random effects models for multivariate recurrent event data that allow for time-varying dependence to adequately capture complex correlation structures among different types of recurrent events. We developed an expectation–maximization algorithm for stable and efficient model fitting. Extensive simulation studies demonstrated that the estimators of the proposed approach have satisfactory finite sample performance. We applied the proposed model and the estimating method to data from a cohort of stroke patients identified in the University of Texas Houston Stroke Registry and evaluated the effects of risk factors and the dependence structure of different types of post-stroke readmission events.

SMMR Missing Data & Measurement Error Survival Analysis · 发表于 2024-01-23

Comparison between inverse-probability weighting and multiple imputation in Cox model with missing failure subtype

Fuyu Guo, Benjamin Langworthy, Shuji Ogino, Molin Wang

摘要

Identifying and distinguishing risk factors for heterogeneous disease subtypes has been of great interest. However, missingness in disease subtypes is a common problem in those data analyses. Several methods have been proposed to deal with the missing data, including complete-case analysis, inverse-probability weighting, and multiple imputation. Although extant literature has compared these methods in missing problems, none has focused on the competing risk setting. In this paper, we discuss the assumptions required when complete-case analysis, inverse-probability weighting, and multiple imputation are used to deal with the missing failure subtype problem, focusing on how to implement these methods under various realistic scenarios in competing risk settings. Besides, we compare these three methods regarding their biases, efficiency, and robustness to model misspecifications using simulation studies. Our results show that complete-case analysis can be seriously biased when the missing completely at random assumption does not hold. Inverse-probability weighting and multiple imputation estimators are valid when we correctly specify the corresponding models for missingness and for imputation, and multiple imputation typically shows higher efficiency than inverse-probability weighting. However, in real-world studies, building imputation models for the missing subtypes can be more challenging than building missingness models. In that case, inverse-probability weighting could be preferred for its easy usage. We also propose two automated model selection procedures and demonstrate their usage in a study of the association between smoking and colorectal cancer subtypes in the Nurses’ Health Study and Health Professional Follow-Up Study.

SiM Bayesian Methods · 发表于 2024-01-23

Bayesian response adaptive randomization design with a composite endpoint of mortality and morbidity

Zhongying Xu, Tianzhou Ma, Lu Tang, Victor B. Talisa, Chung‐Chou H. Chang

摘要

Allocating patients to treatment arms during a trial based on the observed responses accumulated up to the decision point, and sequential adaptation of this allocation, could minimize the expected number of failures or maximize total benefits to patients. In this study, we developed a Bayesian response‐adaptive randomization (RAR) design targeting the endpoint of organ support‐free days (OSFD) for patients admitted to the intensive care units. The OSFD is a mixture of mortality and morbidity assessed by the number of days of free of organ support within a predetermined post‐randomization time‐window. In the past, researchers treated OSFD as an ordinal outcome variable where the lowest category is death. We propose a novel RAR design for a composite endpoint of mortality and morbidity, for example, OSFD, by using a Bayesian mixture model with a Markov chain Monte Carlo sampling to estimate the posterior probability distribution of OSFD and determine treatment allocation ratios at each interim. Simulations were conducted to compare the performance of our proposed design under various randomization rules and different alpha spending functions. The results show that our RAR design using Bayesian inference allocated more patients to the better performing arm(s) compared to other existing adaptive rules while assuring adequate power and type I error rate control across a range of plausible clinical scenarios.

SiM Missing Data & Measurement Error · 发表于 2024-01-22

Multiple imputation strategies for missing event times in a multi‐state model analysis

Elinor Curnow, Rachael A. Hughes, Kate Birnie, Kate Tilling, Michael J. Crowther

摘要

In clinical studies, multi‐state model (MSM) analysis is often used to describe the sequence of events that patients experience, enabling better understanding of disease progression. A complicating factor in many MSM studies is that the exact event times may not be known. Motivated by a real dataset of patients who received stem cell transplants, we considered the setting in which some event times were exactly observed and some were missing. In our setting, there was little information about the time intervals in which the missing event times occurred and missingness depended on the event type, given the analysis model covariates. These additional challenges limited the usefulness of some missing data methods (maximum likelihood, complete case analysis, and inverse probability weighting). We show that multiple imputation (MI) of event times can perform well in this setting. MI is a flexible method that can be used with any complete data analysis model. Through an extensive simulation study, we show that MI by predictive mean matching (PMM), in which sampling is from a set of observed times without reliance on a specific parametric distribution, has little bias when event times are missing at random, conditional on the observed data. Applying PMM separately for each sub‐group of patients with a different pathway through the MSM tends to further reduce bias and improve precision. We recommend MI using PMM methods when performing MSM analysis with Markov models and partially observed event times.

JRSSB Functional Data · 发表于 2024-01-22

Gradient synchronization for multivariate functional data, with application to brain connectivity

Yaqing Chen, Shu-Chin Lin, Yang Zhou, Owen Carmichael, Hans-Georg Müller, Jane-Ling Wang

摘要

Quantifying the association between components of multivariate random curves is of general interest and is a ubiquitous and basic problem that can be addressed with functional data analysis. An important application is the problem of assessing functional connectivity based on functional magnetic resonance imaging (fMRI), where one aims to determine the similarity of fMRI time courses that are recorded on anatomically separated brain regions. In the functional brain connectivity literature, the static temporal Pearson correlation has been the prevailing measure for functional connectivity. However, recent research has revealed temporally changing patterns of functional connectivity, leading to the study of dynamic functional connectivity. This motivates new similarity measures for pairs of random curves that reflect the dynamic features of functional similarity. Specifically, we introduce gradient synchronization measures in a general setting. These similarity measures are based on the concordance and discordance of the gradients between paired smooth random functions. Asymptotic normality of the proposed estimates is obtained under regularity conditions. We illustrate the proposed synchronization measures via simulations and an application to resting-state fMRI signals from the Alzheimer’s Disease Neuroimaging Initiative and they are found to improve discrimination between subjects with different disease status.

SiM Machine Learning · 发表于 2024-01-21

Statistical performance review on diagnosis of leukemia, glaucoma and diabetes mellitus using AI

Rengaraju Perumalraja, B. Felcia Logan's Deshna, N. Swetha

摘要

The growth of artificial intelligence (AI) in the healthcare industry tremendously increases the patient outcomes by reshaping the way we diagnose, treat and monitor patients. AI‐based innovation in healthcare include exploration of drugs, personalized medicine, clinical diagnosis investigations, robotic‐assisted surgery, verified prescriptions, pregnancy care for women, radiology, and reviewed patient information analytics. However, prediction of AI‐based solutions are depends mainly on the implementation of statistical algorithms and input data set. In this article, statistical performance review on various algorithms, Accuracy, Precision, Recall and F1‐Score used to predict the diagnosis of leukemia, glaucoma, and diabetes mellitus is presented. Review on statistical algorithms' performance, used for individual disease diagnosis gives a complete picture of various research efforts during the last two decades. At the end of statistical review on each disease diagnosis, we have discussed our inferences that will give future directions for the new researchers on selection of AI statistical algorithm as well as the input data set.

SiM Spatial & Time Series Survival Analysis · 发表于 2024-01-21

Efficient estimation of Cox model with random change point

Xuerong Chen, Yalu Ping, Jianguo Sun

摘要

In clinical studies, the risk of a disease may dramatically change when some biological indexes of the human body exceed some thresholds. Furthermore, the differences in individual characteristics of patients such as physical and psychological experience may lead to subject‐specific thresholds or change points. Although a large literature has been established for regression analysis of failure time data with change points, most of the existing methods assume the same, fixed change point for all study subjects. In this paper, we consider the situation where there exists a subject‐specific change point and two Cox type models are presented. The proposed models also offer a framework for subgroup analysis. For inference, a sieve maximum likelihood estimation procedure is proposed and the asymptotic properties of the resulting estimators are established. An extensive simulation study is conducted to assess the empirical performance of the proposed method and indicates that it works well in practical situations. Finally the proposed approach is applied to a set of breast cancer data.

SiM Survival Analysis Missing Data & Measurement Error · 发表于 2024-01-20

τ$$ \tau $$‐Inflated beta regression model for censored recurrent events

Yizhuo Wang, Susan Murray

摘要

This research introduces a multivariate ‐inflated beta regression (‐IBR) modeling approach for the analysis of censored recurrent event data that is particularly useful when there is a mixture of (a) individuals who are generally less susceptible to recurrent events and (b) heterogeneity in duration of event‐free periods amongst those who experience events. The modeling approach is applied to a restructured version of the recurrent event data that consists of censored longitudinal times‐to‐first‐event in length follow‐up windows that potentially overlap. Multiple imputation (MI) and expectation‐solution (ES) approaches appropriate for censored data are developed as part of the model fitting process. A suite of useful analysis outputs are provided from the ‐IBR model that include parameter estimates to help interpret the (a) and (b) mixture of event times in the data, estimates of mean ‐restricted event‐free duration in a ‐length follow‐up window based on a patient's covariate profile, and heat maps of raw ‐restricted event‐free durations observed in the data with censored observations augmented via averages across MI datasets. Simulations indicate good statistical performance of the proposed ‐IBR approach to modeling censored recurrent event data. An example is given based on the Azithromycin for Prevention of COPD Exacerbations Trial.

SiM Clinical Trials · 发表于 2024-01-20

Steady‐state statistical properties and implementation of randomization designs with maximum tolerated imbalance restriction for two‐arm equal allocation clinical trials

Wenle Zhao, Kerstine Carter, Oleksandr Sverdlov, Annika Scheffold, Yevgen Ryeznik, Christy Cassarly, Vance W. Berger

摘要

In recent decades, several randomization designs have been proposed in the literature as better alternatives to the traditional permuted block design (PBD), providing higher allocation randomness under the same restriction of the maximum tolerated imbalance (MTI). However, PBD remains the most frequently used method for randomizing subjects in clinical trials. This status quo may reflect an inadequate awareness and appreciation of the statistical properties of these randomization designs, and a lack of simple methods for their implementation. This manuscript presents the analytic results of statistical properties for five randomization designs with MTI restriction based on their steady‐state probabilities of the treatment imbalance Markov chain and compares them to those of the PBD. A unified framework for randomization sequence generation and real‐time on‐demand treatment assignment is proposed for the straightforward implementation of randomization algorithms with explicit formulas of conditional allocation probabilities. Topics associated with the evaluation, selection, and implementation of randomization designs are discussed. It is concluded that for two‐arm equal allocation trials, several randomization designs offer stronger protection against selection bias than the PBD does, and their implementation is not necessarily more difficult than the implementation of the PBD.

Biometrika Nonparametric & Semiparametric · 发表于 2024-01-20

Efficient nonparametric estimation of Toeplitz covariance matrices

K Klockmann, T Krivobokova

摘要

A new efficient nonparametric estimator for Toeplitz covariance matrices is proposed. This estimator is based on a data transformation that translates the problem of Toeplitz covariance matrix estimation to the problem of mean estimation in an approximate Gaussian regression. The resulting Toeplitz covariance matrix estimator is positive definite by construction, fully data driven and computationally very fast. Moreover, this estimator is shown to be minimax optimal under the spectral norm for a large class of Toeplitz matrices. These results are readily extended to estimation of inverses of Toeplitz covariance matrices. Also, an alternative version of the Whittle likelihood for the spectral density based on the discrete cosine transform is proposed.

Biometrika Other · 发表于 2024-01-18

A note on minimax robustness of designs against correlated or heteroscedastic responses

D P Wiens

摘要

We present a result according to which certain functions of covariance matrices are maximized at scalar multiples of the identity matrix. This is used to show that experimental designs that are optimal under an assumption of independent, homoscedastic responses can be minimax robust, in broad classes of alternate covariance structures. In particular, it can justify the common practice of disregarding possible dependence, or heteroscedasticity, at the design stage of an experiment.

JRSSB Machine Learning 讨论 / 评论 · 发表于 2024-01-17

Tamás P. Papp, Paul Fearnhead, and Chris Sherlock’s contribution to the Discussion of ‘the Discussion Meeting on Probabilistic and statistical aspects of machine learning’

Tamás P Papp, Paul Fearnhead, Chris Sherlock

摘要

We provide comments only on the read paper “From Denoising Diffusions to Denoising Markov Models” by Benton, Shi, De Bortoli, Deligiannidis and Doucet. We congratulate the authors on an important contribution to the exciting area of generative models: providing a simple framework for applying the idea of diffusion generative models to data defined on general spaces and for general noising processes. Diffusion models have shown remarkable empirical performance for several applications, most notably as a way of defining implicit models for images. However it is unclear, to us at least, why these methods are so successful. They typically involve fitting highly parameterized models for the score, and it is unclear why one would expect to be able to find good parameter estimates for such models. Furthermore, such models may be sufficiently flexible to learn the exact denoising process for the samples they are trained on—thereby collapsing the diffusion model to a categorical sampler from the training data. Could the authors provide insight into why these methods work? Is there intuition as to what types of data their denoising Markov models will work well for? Diffusion models are often used when the data are supported on a low-dimensional manifold of the state space. However, the Kullback-Leibler (KL) divergence which these models implicitly minimize can be infinite when the supports of the distributions are disjoint. Can the proposed framework extend to objectives which do not have this issue? Or is the use of KL beneficial, in that it actually helps force the model to sample from close to the manifold? The proposed framework parameterizes the generative process in terms of a function β(x;t)⁠, whose optimal value is the marginal density of the noised data at time t. However, in practice the authors use different parameterizations, mainly based on differences in logβ(x;t)⁠. Can the authors give more intuition as to why this is the best choice? And do they think this would be the case in general? Finally, both parameterizations of the continuous-time Markov chain model revolve around logr(x,y;t)=logβ(x;t)−logβ(y;t). As far as we can tell, the authors do not impose the anti-symmetry logr(x,y;t)=−logr(y,x;t) when fitting r, which seems like some form of relaxation of the optimization problem. Are there advantages in fitting a more general functional form? Even if you fit a general function r, you can apply a post-hoc anti-symmetric correction: With the denoising formulation, this does not require additional neural network evaluations. We found that this correction produced almost identical results for the MNIST inpainting example (see Table 1). Image quality metrics for MNIST 14×14 inpainting, with standard errors in brackets *(abstract via OpenAlex)*

JRSSB Other · 发表于 2024-01-17

From denoising diffusions to denoising Markov models

Joe Benton, Yuyang Shi, Valentin De Bortoli, George Deligiannidis, Arnaud Doucet

摘要

Denoising diffusions are state-of-the-art generative models exhibiting remarkable empirical performance. They work by diffusing the data distribution into a Gaussian distribution and then learning to reverse this noising process to obtain synthetic datapoints. The denoising diffusion relies on approximations of the logarithmic derivatives of the noised data densities using score matching. Such models can also be used to perform approximate posterior simulation when one can only sample from the prior and likelihood. We propose a unifying framework generalizing this approach to a wide class of spaces and leading to an original extension of score matching. We illustrate the resulting models on various applications.

JASA Latent Variables & Mixtures High-Dimensional · 发表于 2024-01-16

Clustering High-Dimensional Noisy Categorical Data

Zhiyi Tian, Jiaming Xu, Jen Tang

摘要

Clustering is a widely used unsupervised learning technique that groups data into homogeneous clusters. However, when dealing with real-world data that contain categorical values, existing algorithms can be computationally costly in high dimensions and can struggle with noisy data that has missing values. Furthermore, except for one algorithm, no others provide theoretical guarantees of clustering accuracy. In this article, we propose a general categorical data encoding method and a computationally efficient spectral-based algorithm to cluster high-dimensional noisy categorical data (nominal or ordinal). Under a statistical model for data on m attributes from n subjects in r clusters with missing probability ϵ, we show that our algorithm exactly recovers the true clusters with high probability when mn(1−ϵ)≥CMr2 log 3M, with M=max(n,m) and a fixed constant C. In addition, we show that mn(1−ϵ)2≥rδ/2 with 0<δ<1 is necessary for any algorithm to succeed with probability at least (1+δ)/2. In cases where m = n and r are fixed, the sufficient condition matches with the necessary condition up to a polylog(n) factor. In numerical studies our algorithm outperforms several existing algorithms in both clustering accuracy and computational efficiency. Supplementary materials for this article are available online. *(abstract via OpenAlex)*

SiM Bayesian Methods Longitudinal Data Functional Data · 发表于 2024-01-14

A Bayesian framework for modeling COVID‐19 case numbers through longitudinal monitoring of SARS‐CoV‐2 RNA in wastewater

Xiaotian Dai, Nicole Acosta, Xuewen Lu, Casey R. J. Hubert, Jangwoo Lee, Kevin Frankowski, Maria A. Bautista, Barbara J. Waddell, Kristine Du, Janine McCalder, Jon Meddings, Norma Ruecker, Tyler Williamson, Danielle A. Southern, Jordan Hollman, Gopal Achari, M. Cathryn Ryan, Steve E. Hrudey, Bonita E. Lee, Xiaoli Pang, Rhonda G. Clark, Michael D. Parkins, Thierry Chekouo

摘要

Wastewater‐based surveillance has become an important tool for research groups and public health agencies investigating and monitoring the COVID‐19 pandemic and other public health emergencies including other pathogens and drug abuse. While there is an emerging body of evidence exploring the possibility of predicting COVID‐19 infections from wastewater signals, there remain significant challenges for statistical modeling. Longitudinal observations of viral copies in municipal wastewater can be influenced by noisy datasets and missing values with irregular and sparse samplings. We propose an integrative Bayesian framework to predict daily positive cases from weekly wastewater observations with missing values via functional data analysis techniques. In a unified procedure, the proposed analysis models severe acute respiratory syndrome coronavirus‐2 RNA wastewater signals as a realization of a smooth process with error and combines the smooth process with COVID‐19 cases to evaluate the prediction of positive cases. We demonstrate that the proposed framework can achieve these objectives with high predictive accuracies through simulated and observed real data.

JRSSB Machine Learning Spatial & Time Series Networks & Graphs 讨论 / 评论 · 发表于 2024-01-12

Johannes Schmidt-Hieber’s contribution to the Discussion of ‘the Discussion Meeting on Probabilistic and statistical aspects of machine learning’

Johannes Schmidt-Hieber

摘要

I would like to congratulate the authors to this important contribution linking change-point detection and machine learning. My first comment is related to convolutional neural networks (CNNs) that are considered in Section 6. Empirical studies for image classification have shown that in the first hidden layer, CNNs learn, among other features, to detect edges in various directions. This can be viewed as change-point detection in two dimensions. For one change point, the article uses that the cumulative sum (CUSUM) transformation C(x)=(v1⊤x,…,vn−1⊤x)⊤ can be realized by the first layer in a neural network with n−1 units in the hidden layer and input vector x. Since with ri:=n−ini, all weight vectors v1,…,vn−1 have a similar structure. If the ri would not depend on i, the CUSUM transformation could even be realized by a feature map in a CNN. But always a small discrepancy remains between the CUSUM approach and feature learning in CNNs. An interesting follow-up question is whether one could derive an analogue of Theorem 4.2 for CNNs instead of fully connected neural networks. To answer this question, the right setting might be (as in Section 6) to assume that there are several change points present and that any two change points could be close. This requires a more local approach and CNNs might then be close to optimal. My second comment relates to depth. While shallow networks are sufficient for the change-point detection problem considered here, one can wonder what type of relevant higher order structures become learnable by adding layers. This question might be even more relevant for edge detection in the two-dimensional case. It seems conceivable that adding another convolutional layer allows to learn patterns within the detected edges. Consequently, a neural network with two hidden layers could potentially recognize far more complex shapes. *(abstract via OpenAlex)*

第 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 页