204 research outputs found
Editorial: Explainable artificial intelligence models and methods in finance and healthcare
No full textThis article is a foreword to a special issue on "Explainable artificial intelligence models and methods in finance and healthcare" and introduces the main articles of the collection. The core topic of this special issue is explainability and trusting algorithmic output
Combining hidden Markov models for comparing the dynamics of multiple sleep electroencephalograms
No full textLangrock R, Swihart BJ, Caffo BS, Punjabi NM, Crainiceanu CM. Combining hidden Markov models for comparing the dynamics of multiple sleep electroencephalograms. Statistics in Medicine. 2013;32(19):3342-3356
FACE TOUCH DETECTION TO REDUCE DISEASE TRANSMISSION
No full textIn the last several years, multiple research teams have investigated the face-touch detection problem for the purpose of developing technology that can detect and subsequently prevent face-touching. IMU-based datasets have been cultivated under controlled settings of trial participants performing different face-touches. This thesis seeks to use one well-outlined dataset in particular, to build a “pre-face touch detection model” and a “completed face-touch detection model”. These two Human Activity Recognition (HAR) models are designed to be situated in a free-living study of participants “in the wild”. The pre-face touch detection model is the focus of this thesis, as we primarily seek to build a face-touch prevention tool that potentially can reduce disease transmission for its user. The need for researchers to collect free-living face-touch trials such that a pre-face touch detection model can be built that is generalizable to users “in the wild” is what motivated the development of the completed face-touch detection model. The main contribution of this thesis to the sub-problem of face-touch detection within the HAR space is our investigation of utilizing Discrete Wavelet Transform (DWT) on IMU sensor signals in order to extract localized time-frequency features relevant to the tasks of pre-face touch detection and completed face-touch detection. Additionally, we identify constraints on Precision and Specificity metrics in order to minimize the invasiveness of the pre-face touch detection model, and we maximize Recall score with these constraints considered
FACE TOUCH DETECTION TO REDUCE DISEASE TRANSMISSION
No full textIn the last several years, multiple research teams have investigated the face-touch detection problem for the purpose of developing technology that can detect and subsequently prevent face-touching. IMU-based datasets have been cultivated under controlled settings of trial participants performing different face-touches. This thesis seeks to use one well-outlined dataset in particular, to build a “pre-face touch detection model” and a “completed face-touch detection model”. These two Human Activity Recognition (HAR) models are designed to be situated in a free-living study of participants “in the wild”. The pre-face touch detection model is the focus of this thesis, as we primarily seek to build a face-touch prevention tool that potentially can reduce disease transmission for its user. The need for researchers to collect free-living face-touch trials such that a pre-face touch detection model can be built that is generalizable to users “in the wild” is what motivated the development of the completed face-touch detection model. The main contribution of this thesis to the sub-problem of face-touch detection within the HAR space is our investigation of utilizing Discrete Wavelet Transform (DWT) on IMU sensor signals in order to extract localized time-frequency features relevant to the tasks of pre-face touch detection and completed face-touch detection. Additionally, we identify constraints on Precision and Specificity metrics in order to minimize the invasiveness of the pre-face touch detection model, and we maximize Recall score with these constraints considered
Autoencoder-based deep learning methods for single cell and spatial transcriptomics
No full textThe advent of single-cell and spatial transcriptomics has provided unprecedented resolution into cellular heterogeneity but introduced major computational challenges for data integration and interpretation. In this thesis, we develop and explore deep learning-based methods to address these challenges through the lens of autoencoder architectures.
In Chapter 1, we benchmark Variational Autoencoders (VAEs) against traditional linear decomposition methods, demonstrating that VAEs more effectively capture nonlinear biological variation in single-cell RNA-seq datasets. However, we also highlight inherent limitations in interpretability of VAE latent spaces and show through simulation studies that interpretability cannot be guaranteed without additional model constraints.
Building on these findings, Chapter 2 addresses the critical problem of cross-sample integration in spatial transcriptomics. Using the human dorsolateral prefrontal cortex dataset, we evaluate the performance of adversarial domain adaptation methods and show that standard domain classifier approaches may falter when scaling to multiple samples, motivating the need for more stable integration frameworks.
Finally, in Chapter 3, we introduce a novel autoencoder-based model that explicitly disentangles latent representations into common and dataset-specific components. We propose using the Sliced Wasserstein Distance as a stable and interpretable alternative to adversarial losses for multi-sample integration. Proof-of-concept experiments on synthetic datasets show that our approach successfully recovers and disentangles shared and residual signals, laying the groundwork for future applications to real-world biological data.
Overall, this thesis advances the application of deep learning for high-dimensional biomedical data analysis by proposing interpretable, stable, and scalable representation learning frameworks
Turning the R Console Into an Interactive Learning Environment with swirl
No full textInteractive platforms for learning computer programming have exploded in recent years. Sadly, the R programming language is overlooked by most of these tools. swirl is an open source software package for the R programming language that allows users to learn R and statistics interactively, right in the R console. swirl supports multiple questions types, but emphasis is on having the user interact directly with the R prompt, while receiving helpful and immediate feedback. We (swirl’s creators) have pioneered early content creation, but swirl is constructed in such a way that anyone can create his or her own content and share it freely with others. While swirl has been downloaded over 70,000 times in its first nine months, development of the software and instructional content is ongoing. We hope that the R and statistics communities will continue to rally around the platform and truly embrace it as their own
Autoencoder-based deep learning methods for single cell and spatial transcriptomics
No full textThe advent of single-cell and spatial transcriptomics has provided unprecedented resolution into cellular heterogeneity but introduced major computational challenges for data integration and interpretation. In this thesis, we develop and explore deep learning-based methods to address these challenges through the lens of autoencoder architectures.
In Chapter 1, we benchmark Variational Autoencoders (VAEs) against traditional linear decomposition methods, demonstrating that VAEs more effectively capture nonlinear biological variation in single-cell RNA-seq datasets. However, we also highlight inherent limitations in interpretability of VAE latent spaces and show through simulation studies that interpretability cannot be guaranteed without additional model constraints.
Building on these findings, Chapter 2 addresses the critical problem of cross-sample integration in spatial transcriptomics. Using the human dorsolateral prefrontal cortex dataset, we evaluate the performance of adversarial domain adaptation methods and show that standard domain classifier approaches may falter when scaling to multiple samples, motivating the need for more stable integration frameworks.
Finally, in Chapter 3, we introduce a novel autoencoder-based model that explicitly disentangles latent representations into common and dataset-specific components. We propose using the Sliced Wasserstein Distance as a stable and interpretable alternative to adversarial losses for multi-sample integration. Proof-of-concept experiments on synthetic datasets show that our approach successfully recovers and disentangles shared and residual signals, laying the groundwork for future applications to real-world biological data.
Overall, this thesis advances the application of deep learning for high-dimensional biomedical data analysis by proposing interpretable, stable, and scalable representation learning frameworks
A User-Friendly Introduction to Link-Probit-Normal Models
No full textProbit-normal models have attractive properties compared to logit-normal models. In particular, they allow for easy specification of marginal links of interest while permitting a conditional random effects structure. Moreover, programming fitting algorithms for probit-normal models can be trivial with the use of well-developed algorithms for approximating multivariate normal quantiles. In typical settings, the data cannot distinguish between probit and logit conditional link functions. Therefore, if marginal interpretations are desired, the default conditional link should be the most convenient one. We refer to models with a probit conditional link an arbitrary marginal link and a normal random effect distribution as link-probit-normal models. In this manuscript we outline these models and discuss appropriate situations for using multivariate normal approximations. Unlike other manuscripts in this area that focus on very general situations and implement Markov chain or MCEM algorithms, we focus on simpler, random intercept settings and give a collection of user-friendly examples and reproducible code. Marginally, the link-probit-normal model is obtained by a non-linear model on a discretized multivariate normal distribution, and thus can be thought of as a special case of discretizing a multivariate T distribution (as the degrees of freedom go to infinity). We also consider the larger class of multivariate T marginal models and illustrate how these models can be used to closely approximate a logit link
Exact Hypothesis Tests for Log-linear Models with exactLoglinTest
No full textThis manuscript overviews exact testing of goodness of fit for log-linear models using the R package exactLoglinTest. This package evaluates model fit for Poisson log-linear models by conditioning on minimal sufficient statistics to remove nuisance parameters. A Monte Carlo algorithm is proposed to estimate P values from the resulting conditional distribution. In particular, this package implements a sequentially rounded normal approximation and importance sampling to approximate probabilities from the conditional distribution. Usually, this results in a high percentage of valid samples. However, in instances where this is not the case, a Metropolis Hastings algorithm can be implemented that makes more localized jumps within the reference set. The manuscript details how some conditional tests for binomial logit models can also be viewed as conditional Poisson log-linear models and hence can be performed via exactLoglinTest. A diverse battery of examples is considered to highlight use, features and extensions of the software. Notably, potential extensions to evaluating disclosure risk are also considered.
Methods in Biostatistics I
No full textThe field of biostatistics, which combines a number of different disciplines, is one that more people seek to enter. The Johns Hopkins School of Public Health's Brian Caffo created these course materials for his Methods in Biostatistics I class. As the site notes, these materials present "fundamental concepts in applied probability, exploratory data analysis, and statistical inference, focusing on probability and analysis of one and two samples." Visitors can look over the syllabus here, check out the original course schedule, peruse the lecture materials, and look over the readings. The lecture notes cover set theory basics and probability, expected values, random vectors, distribution, and confidence intervals. The site is rounded out by the Other Resources area, which includes links to free statistical software programs and other supplemental items
- …
