Search

The aim of survival analysis in healthcare is to estimate the probability of occurrence of an event, such as a patient's death in an intensive care unit (ICU). Recent developments in deep neural networks (DNNs) for survival analysis show the superiority of these models in comparison with other well-known models in survival analysis applications. Ensuring the reliability and explainability of deep survival models deployed in healthcare is a necessity. Since DNN models often behave like a black box, their predictions might not be easily trusted by clinicians, especially when predictions are contrary to a physician's opinion. A deep survival model that explains and justifies its decision-making process could potentially gain the trust of clinicians. In this research, we propose the reverse survival model (RSM) framework that provides detailed insights into the decision-making process of survival models. For each patient of interest, RSM can extract similar patients from a dataset and rank them based on the most relevant features that deep survival models rely on for their predictions. RSM acts as an add-on to a deep survival model and offers three functionalities: 1) Finding the most relevant clinical measurements for the probability density functions (PDFs) of events. 2) Categorizing patients into disjoint clusters based on the similarity of their survival PDFs. 3) Ranking similar patients based on the similarity of survival outcomes and relevant clinical measurements. The explainability of deep survival models is rarely addressed in the literature. Therefore, the RSM pipeline is a unique approach to explaining the predictions of deep survival models. We validated the RSM pipeline by testing it on a synthetic dataset and MIMIC-IV, a dataset of intensive care unit (ICU) clinical observations. Our experiments showed that given a deep survival model and a patient of interest, RSM can successfully detect similar patient records from historical data and rank them based on the similarities between their survival PDFs and the most relevant patient observations. [ABSTRACT FROM AUTHOR]

Uncertainty in renewable energy generation, energy consumption, and electricity prices, as well as transmission congestion, pose a number of problems in modern power grids, necessitating stability on the supply, grid, and demand sides. Grid‐side stability can be achieved by dynamic line rating (DLR) forecasting, which reliably predicts the overall current carrying potential of overhead transmission lines. Long short‐term memory proved beneficiary in this field, owing to its ability to learn highly variable and uncertain data. To empower this network to tackle the non‐stationary nature of meteorological parameters, a novel machine learning (ML) architecture based on Dagging technique is proposed and tested on the data collected from a 400 kV overhead transmission line. Simulation results corroborate that the proposed Dagging‐based stacked LSTM can successfully handle the non‐stationary issue and outperform the decomposition‐based technique, as the state‐of‐the‐art algorithm, for various forecasting horizons. The results confirm the generalizability of models with an application in forecasting DLR over the line without utilizing additional sensors and communication networks. Moreover, the proposed model is compared to several ML architectures, including support vector machines (SVM), random forest (RF), and multi‐layer perceptron (MLP) in a comprehensive benchmark study. The introduced algorithm outperforms MLP by 3.4%, RF by 9.4%, and SVM by 6.7% in terms of average prediction accuracy. [ABSTRACT FROM AUTHOR]

Background: Ulcerative colitis (UC), a complex polygenic disorder, is one of the main subphenotypes of inflammatory bowel disease. A comprehensive dissection of the genetic etiology of UC needs to assess the contribution of rare genetic variants including copy number variations (CNVs) to disease risk. In this study, we performed a multi-step genome-wide case-control analysis to interrogate the presence of disease-relevant rare copy number variants. Methods: One thousand one hundred twenty-one German UC patients and 1770 healthy controls were initially screened for rare deletions and duplications employing SNP-array data. Quantitative PCR and high density custom array-CGH were used for validation of identified CNVs and fine mapping. Two main follow-up panels consisted of an independent cohort of 451 cases and 1274 controls, in which CNVs were assayed through quantitative PCR, and a British cohort of 2396 cases versus 4886 controls with CNV genotypes based on array data. Additional sample sets were assessed for targeted and in silico replication. Results: Twenty-four rare copy number variants (14 deletions and 10 duplications), overrepresented in UC patients were identified in the initial screening panel. Follow-up of these CNV regions in four independent case-control series as well as an additional public in silico control group (totaling 4439 UC patients and 15,961 healthy controls) revealed three copy number variants enriched in UC patients; a 15.8 kb deletion upstream of ABCC4 and CLDN10 at13q32.1 (0.43 % cases, 0.11 % controls), a 119 kb duplication at 7p22.1, overlapping RNF216, ZNF815, OCM and CCZ1 (0.13 % cases, 0.01 % controls) and a 134 kb large duplication upstream of the KCNK9 gene at 8q24.3 (0.22 % carriers among cases, 0.03 % carriers among controls). The trend of association with UC was present after the P-values were corrected for combining data from different subpopulations. Break-point mapping of the deleted region suggested non-allelic homologous recombination as the mechanism underlying its formation. Conclusion: Our study presents a pragmatic approach for effective rare CNV screening of SNP-array data sets and implicates the potential contribution of rare structural variants in the pathogenesis of UC. [ABSTRACT FROM AUTHOR]