Search

Objectives We developed a machine-learning model-based algorithm (MBA) for smoking in Administrative Health Data (AHD). The validity of this MBA was compared to a rule-based algorithm (RBA). Approaches The study included adults (≥18 years) from a clinical registry containing self-reported current smoking from 2017 to 2020 in Manitoba, Canada. Clinical data were linked to up to five years of hospitalization, physician billing claims, and prescription medication records. The RBA was based on diagnosis codes for tobacco use and nicotine dependence medication. MBAs, constructed using random forest (RF) models, included these indicators in addition to comorbid condition diagnoses and sociodemographic factors. Sensitivity, specificity, positive and negative predictive values (PPV, NPV), and 95% confidence intervals (CIs) were estimated. Results The cohort comprised 24,718 individuals (88.6% female); prevalence of current smokers was 10.0%. The RBA had sensitivity of 27.3% (95% CI: 24.2-30.7), specificity of 96.6% (95% CI: 96.1-97.0), and PPV of 47.2% (95% CI: 42.9-51.5). The MBA had sensitivity of 68.6% (95% CI: 65.1-71.9), specificity of 76.3% (95% CI: 75.2-77.3), and PPV of 24.3% (95% CI: 23.2-25.6). NPV was high irrespective of algorithms. Stratified analyses revealed similar estimates for males and females, and the number of years of AHD did not affect the MBA results. Conclusions An RF-based MBA for smoking ascertainment in linked AHD sources improved sensitivity compared to the RBA. However, the RBA excelled in specificity and PPV. Implication Balancing accurate smoker identification with the risk of false positives is crucial when choosing an algorithm to ascertain current smokers using AHD.

Unsupervised anomaly detection (UAD) is a diverse research area explored across various application domains. Over time, numerous anomaly detection techniques, including clustering, generative, and variational inference-based methods, are developed to address specific drawbacks and advance state-of-the-art techniques. Deep learning and generative models recently played a significant role in identifying unique challenges and devising advanced approaches. Auto-encoders (AEs) represent one such powerful technique that combines generative and probabilistic variational modeling with deep architecture. Auto-Encoder aims to learn the underlying data distribution to generate consequential sample data. This concept of data generation and the adoption of generative modeling have emerged in extensive research and variations in Auto-Encoder design, particularly in unsupervised representation learning. This study systematically reviews 11 Auto-Encoder architectures categorized into three groups, aiming to differentiate their reconstruction ability, sample generation, latent space visualization, and accuracy in classifying anomalous data using the Fashion-MNIST (FMNIST) and MNIST datasets. Additionally, we closely observed the reproducibility scope under different training parameters. We conducted reproducibility experiments utilizing similar model setups and hyperparameters and attempted to generate comparative results to address the scope of improvements for each Auto-Encoder. We conclude this study by analyzing the experimental results, which guide us in identifying the efficiency and trade-offs among auto-encoders, providing valuable insights into their performance and applicability in unsupervised anomaly detection techniques.

In the context of survival analysis, data-driven neural network-based methods have been developed to model complex covariate effects. While these methods may provide better predictive performance than regression-based approaches, not all can model time-varying interactions and complex baseline hazards. To address this, we propose Case-Base Neural Networks (CBNNs) as a new approach that combines the case-base sampling framework with flexible neural network architectures. Using a novel sampling scheme and data augmentation to naturally account for censoring, we construct a feed-forward neural network that includes time as an input. CBNNs predict the probability of an event occurring at a given moment to estimate the full hazard function. We compare the performance of CBNNs to regression and neural network-based survival methods in a simulation and three case studies using two time-dependent metrics. First, we examine performance on a simulation involving a complex baseline hazard and time-varying interactions to assess all methods, with CBNN outperforming competitors. Then, we apply all methods to three real data applications, with CBNNs outperforming the competing models in two studies and showing similar performance in the third. Our results highlight the benefit of combining case-base sampling with deep learning to provide a simple and flexible framework for data-driven modeling of single event survival outcomes that estimates time-varying effects and a complex baseline hazard by design. An R package is available at https://github.com/Jesse-Islam/cbnn.

Abstract Background Smoking is a risk factor for many chronic diseases. Multiple smoking status ascertainment algorithms have been developed for population-based electronic health databases such as administrative databases and electronic medical records (EMRs). Evidence syntheses of algorithm validation studies have often focused on chronic diseases rather than risk factors. We conducted a systematic review and meta-analysis of smoking status ascertainment algorithms to describe the characteristics and validity of these algorithms. Methods The Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines were followed. We searched articles published from 1990 to 2022 in EMBASE, MEDLINE, Scopus, and Web of Science with key terms such as validity, administrative data, electronic health records, smoking, and tobacco use. The extracted information, including article characteristics, algorithm characteristics, and validity measures, was descriptively analyzed. Sources of heterogeneity in validity measures were estimated using a meta-regression model. Risk of bias (ROB) in the reviewed articles was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 tool. Results The initial search yielded 2086 articles; 57 were selected for review and 116 algorithms were identified. Almost three-quarters (71.6%) of algorithms were based on EMR data. The algorithms were primarily constructed using diagnosis codes for smoking-related conditions, although prescription medication codes for smoking treatments were also adopted. About half of the algorithms were developed using machine-learning models. The pooled estimates of positive predictive value, sensitivity, and specificity were 0.843, 0.672, and 0.918 respectively. Algorithm sensitivity and specificity were highly variable and ranged from 3 to 100% and 36 to 100%, respectively. Model-based algorithms had significantly greater sensitivity (p = 0.006) than rule-based algorithms. Algorithms for EMR data had higher sensitivity than algorithms for administrative data (p = 0.001). The ROB was low in most of the articles (76.3%) that underwent the assessment. Conclusions Multiple algorithms using different data sources and methods have been proposed to ascertain smoking status in electronic health data. Many algorithms had low sensitivity and positive predictive value, but the data source influenced their validity. Algorithms based on machine-learning models for multiple linked data sources have improved validity.

Clustering techniques are used to group observations and discover interesting patterns within data. Model-based clustering is one such method that is often an attractive choice due to the specification of a generative model for the given data and the ability to calculate model-selection criteria, which is in turn used to select the number of clusters. However, when only distances between observations are available, model-based clustering can no longer be used, and heuristic algorithms without the aforementioned advantages are usually used instead. As a solution, Oh and Raftery (2007) suggest a Bayesian model-based clustering method (named BMCD) that only requires a dissimilarity matrix as input, while also accounting for the measurement error that may be present within the observed data. In this paper, we extend the BMCD framework by proposing several additional models, alternative model selection criteria, and strategies for reducing computing time of the algorithm. These extensions ensure that the algorithm is effective even in high-dimensional spaces and provides a wide range of choices to the practitioner that can be used with a variety of data. Additionally, a publicly available software implementation of the algorithm is provided as a package in the R programming language.

An unprecedented amount of SARS-CoV-2 sequencing has been performed, however, novel bioinformatic tools to cope with and process these large datasets is needed. Here, we have devised a bioinformatic pipeline that inputs SARS-CoV-2 genome sequencing in FASTA/FASTQ format and outputs a single Variant Calling Format file that can be processed to obtain variant annotations and perform downstream population genetic testing. As proof of concept, we have analyzed over 229,000 SARS-CoV-2 viral sequences up until November 30, 2020. We have identified over 39,000 variants worldwide with increased polymorphisms, spanning the ORF3a gene as well as the 3′ untranslated (UTR) regions, specifically in the conserved stem loop region of SARS-CoV-2 which is accumulating greater observed viral diversity relative to chance variation. Our analysis pipeline has also discovered the existence of SARS-CoV-2 hypermutation with low frequency (less than in 2% of genomes) likely arising through host immune responses and not due to sequencing errors. Among annotated non-sense variants with a population frequency over 1%, recurrent inactivation of the ORF8 gene was found. This was found to be present in the newly identified B.1.1.7 SARS-CoV-2 lineage that originated in the United Kingdom. Almost all VOC-containing genomes possess one stop codon in ORF8 gene (Q27∗), however, 13% of these genomes also contains another stop codon (K68∗), suggesting that ORF8 loss does not interfere with SARS-CoV-2 spread and may play a role in its increased virulence. We have developed this computational pipeline to assist researchers in the rapid analysis and characterization of SARS-CoV-2 variation.