Comparison of Machine Learning Methods for Predicting Outcomes After In-Hospital Cardiac Arrest.

OBJECTIVES: Prognostication of neurologic status among survivors of in-hospital cardiac arrests remains a challenging task for physicians. Although models such as the Cardiac Arrest Survival Post-Resuscitation In-hospital score are useful for predicting neurologic outcomes, they were developed using traditional statistical techniques. In this study, we derive and compare the performance of several machine learning models with each other and with the Cardiac Arrest Survival Post-Resuscitation In-hospital score for predicting the likelihood of favorable neurologic outcomes among survivors of resuscitation. DESIGN: Analysis of the Get With The Guidelines-Resuscitation registry. SETTING: Seven-hundred fifty-five hospitals participating in Get With The Guidelines-Resuscitation from January 1, 2001, to January 28, 2017. PATIENTS: Adult in-hospital cardiac arrest survivors. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: Of 117,674 patients in our cohort, 28,409 (24%) had a favorable neurologic outcome, as defined as survival with a Cerebral Performance Category score of less than or equal to 2 at discharge. Using patient characteristics, pre-existing conditions, prearrest interventions, and periarrest variables, we constructed logistic regression, support vector machines, random forests, gradient boosted machines, and neural network machine learning models to predict favorable neurologic outcome. Events prior to October 20, 2009, were used for model derivation, and all subsequent events were used for validation. The gradient boosted machine predicted favorable neurologic status at discharge significantly better than the Cardiac Arrest Survival Post-Resuscitation In-hospital score (C-statistic: 0.81 vs 0.73; p < 0.001) and outperformed all other machine learning models in terms of discrimination, calibration, and accuracy measures. Variables that were consistently most important for prediction across all models were duration of arrest, initial cardiac arrest rhythm, admission Cerebral Performance Category score, and age. CONCLUSIONS: The gradient boosted machine algorithm was the most accurate for predicting favorable neurologic outcomes in in-hospital cardiac arrest survivors. Our results highlight the utility of machine learning for predicting neurologic outcomes in resuscitated patients.

Evaluation of Hybridization Capture Versus Amplicon-Based Methods for Whole-Exome Sequencing.

Next-generation sequencing has aided characterization of genomic variation. While whole-genome sequencing may capture all possible mutations, whole-exome sequencing remains cost-effective and captures most phenotype-altering mutations. Initial strategies for exome enrichment utilized a hybridization-based capture approach. Recently, amplicon-based methods were designed to simplify preparation and utilize smaller DNA inputs. We evaluated two hybridization capture-based and two amplicon-based whole-exome sequencing approaches, utilizing both Illumina and Ion Torrent sequencers, comparing on-target alignment, uniformity, and variant calling. While the amplicon methods had higher on-target rates, the hybridization capture-based approaches demonstrated better uniformity. All methods identified many of the same single-nucleotide variants, but each amplicon-based method missed variants detected by the other three methods and reported additional variants discordant with all three other technologies. Many of these potential false positives or negatives appear to result from limited coverage, low variant frequency, vicinity to read starts/ends, or the need for platform-specific variant calling algorithms. All methods demonstrated effective copy-number variant calling when evaluated against a single-nucleotide polymorphism array. This study illustrates some differences between whole-exome sequencing approaches, highlights the need for selecting appropriate variant calling based on capture method, and will aid laboratories in selecting their preferred approach.

Regulation of NF-κB signaling by oxidized glycerophospholipid and IL-1ß induced miRs-21-3p and -27a-5p in human aortic endothelial cells.

Exposure of endothelial cells (ECs) to agents such as oxidized glycerophospholipids (oxGPs) and cytokines, known to accumulate in atherosclerotic lesions, perturbs the expression of hundreds of genes in ECs involved in inflammatory and other biological processes. We hypothesized that microRNAs (miRNAs) are involved in regulating the inflammatory response in human aortic endothelial cells (HAECs) in response to oxGPs and interleukin 1ß (IL-1ß). Using next-generation sequencing and RT-quantitative PCR, we characterized the profile of expressed miRNAs in HAECs pre- and postexposure to oxGPs. Using this data, we identified miR-21-3p and miR-27a-5p to be induced 3- to 4-fold in response to oxGP and IL-1ß treatment compared with control treatment. Transient overexpression of miR-21-3p and miR-27a-5p resulted in the downregulation of 1,253 genes with 922 genes overlapping between the two miRNAs. Gene Ontology functional enrichment analysis predicted that the two miRNAs were involved in the regulation of nuclear factor κB (NF-κB) signaling. Overexpression of these two miRNAs leads to changes in p65 nuclear translocation. Using 3' untranslated region luciferase assay, we identified 20 genes within the NF-κB signaling cascade as putative targets of miRs-21-3p and -27a-5p, implicating these two miRNAs as modulators of NF-κB signaling in ECs.

Genetic regulation of human adipose microRNA expression and its consequences for metabolic traits.

The genetics of messenger RNA (mRNA) expression has been extensively studied in humans and other organisms, but little is known about genetic factors contributing to microRNA (miRNA) expression. We examined natural variation of miRNA expression in adipose tissue in a population of 200 men who have been carefully characterized for metabolic syndrome (MetSyn) phenotypes as part of the Metabolic Syndrome in Men (METSIM) study. We genotyped the subjects using high-density single-nucleotide polymorphism microarrays and quantified the mRNA abundance using genome-wide expression arrays and miRNA abundance using next-generation sequencing. We reliably quantified 356 miRNA species that were expressed in human adipose tissue, a limited number of which made up most of the expressed miRNAs. We mapped the miRNA abundance as an expression quantitative trait and determined cis regulation of expression for nine of the miRNAs and of the processing of one miRNA (miR-28). The degree of genetic variation of miRNA expression was substantially less than that of mRNAs. For the majority of the miRNAs, genetic regulation of expression was independent of the expression of mRNA from which the miRNA is transcribed. We also showed that for 108 miRNAs, mapped reads displayed widespread variation from the canonical sequence. We found a total of 24 miRNAs to be significantly associated with MetSyn traits. We suggest a regulatory role for miR-204-5p which was predicted to inhibit acetyl coenzyme A carboxylase ß, a key fatty acid oxidation enzyme that has been shown to play a role in regulating body fat and insulin resistance in adipose tissue.

Genetic regulation of mouse liver metabolite levels.

We profiled and analyzed 283 metabolites representing eight major classes of molecules including Lipids, Carbohydrates, Amino Acids, Peptides, Xenobiotics, Vitamins and Cofactors, Energy Metabolism, and Nucleotides in mouse liver of 104 inbred and recombinant inbred strains. We find that metabolites exhibit a wide range of variation, as has been previously observed with metabolites in blood serum. Using genome-wide association analysis, we mapped 40% of the quantified metabolites to at least one locus in the genome and for 75% of the loci mapped we identified at least one candidate gene by local expression QTL analysis of the transcripts. Moreover, we validated 2 of 3 of the significant loci examined by adenoviral overexpression of the genes in mice. In our GWAS results, we find that at significant loci the peak markers explained on average between 20 and 40% of variation in the metabolites. Moreover, 39% of loci found to be regulating liver metabolites in mice were also found in human GWAS results for serum metabolites, providing support for similarity in genetic regulation of metabolites between mice and human. We also integrated the metabolomic data with transcriptomic and clinical phenotypic data to evaluate the extent of co-variation across various biological scales.

Comparative analysis of proteome and transcriptome variation in mouse.

The relationships between the levels of transcripts and the levels of the proteins they encode have not been examined comprehensively in mammals, although previous work in plants and yeast suggest a surprisingly modest correlation. We have examined this issue using a genetic approach in which natural variations were used to perturb both transcript levels and protein levels among inbred strains of mice. We quantified over 5,000 peptides and over 22,000 transcripts in livers of 97 inbred and recombinant inbred strains and focused on the 7,185 most heritable transcripts and 486 most reliable proteins. The transcript levels were quantified by microarray analysis in three replicates and the proteins were quantified by Liquid Chromatography-Mass Spectrometry using O(18)-reference-based isotope labeling approach. We show that the levels of transcripts and proteins correlate significantly for only about half of the genes tested, with an average correlation of 0.27, and the correlations of transcripts and proteins varied depending on the cellular location and biological function of the gene. We examined technical and biological factors that could contribute to the modest correlation. For example, differential splicing clearly affects the analyses for certain genes; but, based on deep sequencing, this does not substantially contribute to the overall estimate of the correlation. We also employed genome-wide association analyses to map loci controlling both transcript and protein levels. Surprisingly, little overlap was observed between the protein- and transcript-mapped loci. We have typed numerous clinically relevant traits among the strains, including adiposity, lipoprotein levels, and tissue parameters. Using correlation analysis, we found that a low number of clinical trait relationships are preserved between the protein and mRNA gene products and that the majority of such relationships are specific to either the protein levels or transcript levels. Surprisingly, transcript levels were more strongly correlated with clinical traits than protein levels. In light of the widespread use of high-throughput technologies in both clinical and basic research, the results presented have practical as well as basic implications.

Identification of CAD candidate genes in GWAS loci and their expression in vascular cells.

Recent genome-wide association studies (GWAS) have identified 35 loci that significantly associate with coronary artery disease (CAD) susceptibility. The majority of the genes represented in these loci have not previously been studied in the context of atherosclerosis. To characterize the roles of these candidate genes in the vessel wall, we determined their expression levels in endothelial, smooth muscle, and macrophage cells isolated from healthy, prelesioned, and lesioned mouse aortas. We also performed expression quantitative locus (eQTL) mapping of these genes in human endothelial cells under control and proatherogenic conditions. Of the 57 genes studied, 31 were differentially expressed in one or more cell types in disease state in mice, and the expression levels of 8 were significantly associated with the CAD SNPs in human cells, 7 of which were also differentially expressed in mice. By integrating human and mouse results, we predict that PPAP2B, GALNT4, MAPKAPK5, TCTN1, SRR, SNF8, and ICAM1 play a causal role in the susceptibility to atherosclerosis through a role in the vasculature. Additionally, we highlight the genetic complexity of a subset of CAD loci through the differential expression of multiple candidate genes per locus and the involvement of genes that lie outside linkage disequilibrium blocks.

SATCHMO-JS: a webserver for simultaneous protein multiple sequence alignment and phylogenetic tree construction.

We present the jump-start simultaneous alignment and tree construction using hidden Markov models (SATCHMO-JS) web server for simultaneous estimation of protein multiple sequence alignments (MSAs) and phylogenetic trees. The server takes as input a set of sequences in FASTA format, and outputs a phylogenetic tree and MSA; these can be viewed online or downloaded from the website. SATCHMO-JS is an extension of the SATCHMO algorithm, and employs a divide-and-conquer strategy to jump-start SATCHMO at a higher point in the phylogenetic tree, reducing the computational complexity of the progressive all-versus-all HMM-HMM scoring and alignment. Results on a benchmark dataset of 983 structurally aligned pairs from the PREFAB benchmark dataset show that SATCHMO-JS provides a statistically significant improvement in alignment accuracy over MUSCLE, Multiple Alignment using Fast Fourier Transform (MAFFT), ClustalW and the original SATCHMO algorithm. The SATCHMO-JS webserver is available at http://phylogenomics.berkeley.edu/satchmo-js. The datasets used in these experiments are available for download at http://phylogenomics.berkeley.edu/satchmo-js/supplementary/.

Predicting neurological outcomes after in-hospital cardiac arrests for patients with Coronavirus Disease 2019.

BACKGROUND: Machine learning models are more accurate than standard tools for predicting neurological outcomes in patients resuscitated after cardiac arrest. However, their accuracy in patients with Coronavirus Disease 2019 (COVID-19) is unknown. Therefore, we compared their performance in a cohort of cardiac arrest patients with COVID-19. METHODS: We conducted a retrospective analysis of resuscitation survivors in the Get With The Guidelines®-Resuscitation (GWTG-R) COVID-19 registry between February 2020 and May 2021. The primary outcome was a favorable neurological outcome, indicated by a discharge Cerebral Performance Category score ≤ 2. Pre- and peri-arrest variables were used as predictors. We applied our published logistic regression, neural network, and gradient boosted machine models developed in patients without COVID-19 to the COVID-19 cohort. We also updated the neural network model using transfer learning. Performance was compared between models and the Cardiac Arrest Survival Post-Resuscitation In-Hospital (CASPRI) score. RESULTS: Among the 4,125 patients with COVID-19 included in the analysis, 484 (12 %) patients survived with favorable neurological outcomes. The gradient boosted machine, trained on non-COVID-19 patients was the best performing model for predicting neurological outcomes in COVID-19 patients, significantly better than the CASPRI score (c-statistic: 0.75 vs 0.67, P < 0.001). While calibration improved for the neural network with transfer learning, it did not surpass the gradient boosted machine in terms of discrimination. CONCLUSION: Our gradient boosted machine model developed in non-COVID patients had high discrimination and adequate calibration in COVID-19 resuscitation survivors and may provide clinicians with important information for these patients.

Comparison of custom capture for targeted next-generation DNA sequencing.

Targeted, capture-based DNA sequencing is a cost-effective method to focus sequencing on a coding region or other customized region of the genome. There are multiple targeted sequencing methods available, but none has been systematically investigated and compared. We evaluated four commercially available custom-targeted DNA technologies for next-generation sequencing with respect to on-target sequencing, uniformity, and ability to detect single-nucleotide variations (SNVs) and copy number variations. The technologies that used sonication for DNA fragmentation displayed impressive uniformity of capture, whereas the others had shorter preparation times, but sacrificed uniformity. One of those technologies, which uses transposase for DNA fragmentation, has a drawback requiring sample pooling, and the last one, which uses restriction enzymes, has a limitation depending on restriction enzyme digest sites. Although all technologies displayed some level of concordance for calling SNVs, the technologies that require restriction enzymes or transposase missed several SNVs largely because of the lack of coverage. All technologies performed well for copy number variation calling when compared to single-nucleotide polymorphism arrays. These results enable laboratories to compare these methods to make informed decisions for their intended applications.

Limited RNA editing in exons of mouse liver and adipose.

Several studies have investigated RNA-DNA differences (RDD), presumably due to RNA editing, with conflicting results. We report a rigorous analysis of RDD in exonic regions in mice, taking into account critical biases in RNA-Seq analysis. Using deep-sequenced F1 reciprocal inbred mice, we mapped 40 million RNA-Seq reads per liver sample and 180 million reads per adipose sample. We found 7300 apparent hepatic RDDs using a multiple-site mapping procedure, compared with 293 RDD found using a unique-site mapping procedure. After filtering for repeat sequence, splice junction proximity, undirectional strand, and extremity read bias, 63 RDD remained. In adipose tissue unique-site mapping identified 1667 RDD, and after applying the same four filters, 188 RDDs remained. In both tissues, the filtering procedure increased the proportion of canonical (A-to-I and C-to-U) editing events. The genomic DNA of 12 RDD sites among the potential 63 hepatic RDD was tested by Sanger sequencing, three of which proved to be due to unreferenced SNPs. We validated seven liver RDD with Sequenom technology, including two noncanonical, Gm5424 C-to-I(G) and Pisd I(G)-to-A RDD. Differences in diet, sex, or genetic background had very modest effects on RDD occurrence. Only a small number of apparent RDD sites overlapped between liver and adipose, indicating a high degree of tissue specificity. Our findings underscore the importance of properly filtering for bias in RNA-Seq investigations, including the necessity of confirming the DNA sequence to eliminate unreferenced SNPs. Based on our results, we conclude that RNA editing is likely limited to hundreds of events in exonic RNA in liver and adipose.