-Omic and Electronic Health Records Big Data Analytics for Precision Medicine

Abstract

Objective

Rapid advances of high-throughput technologies and wide adoption of electronic health records (EHRs) have led to fast accumulation of -omic and EHR data. These voluminous complex data contain abundant information for precision medicine, and big data analytics can extract such knowledge to improve the quality of health care.

Methods

In this article, we present -omic and EHR data characteristics, associated challenges, and data analytics including data pre-processing, mining, and modeling.

Results

To demonstrate how big data analytics enables precision medicine, we provide two case studies, including identifying disease biomarkers from multi-omic data and incorporating -omic information into EHR.

Conclusion

Big data analytics is able to address –omic and EHR data challenges for paradigm shift towards precision medicine.

Significance

Big data analytics makes sense of –omic and EHR data to improve healthcare outcome. It has long lasting societal impact.

I. Introduction

TO achieve the best care for patients, many models have been proposed over the years to improve the healthcare system. The goal of the early “personalized medicine” model is to customize healthcare delivery for each individual and to maximize the effectiveness of each patient’s treatment [1]. In 2009, Hood et al. propose the “personalized, predictive, preventive, and participatory medicine” (a.k.a. P4 medicine) model that aims to transform current reactive care to future proactive medicine, and ultimately to reduce healthcare expenditure and improve patients’ health outcome [2]. Recently, the new “precision medicine” model is proposed to precisely classify patients into subgroups sharing a common biological basis of diseases for more effective treatment and improved care outcome [3, 4]. Precision medicine requires data utility ranging from collection and management (i.e. data storage, sharing, and privacy) to analytics (i.e. data mining, integration, and visualization) [5]. Because of rapid advances in biotechnologies, highly complex biomedical data are becoming available in huge volumes [6]. To make sense of these heterogeneous data, big data analytics, including data quality control, analysis, modeling, interpretation, and validation, is needed to cover application areas such as bioinformatics [7–9], health informatics [10–12], imaging informatics [13, 14], and sensor informatics [15, 16].

As presented in 2015 US Precision Medicine Initiative [17], incorporating -omic data and knowledge into electronic health record (EHR) (Fig. 1) is viewed as a necessary step for delivering precision medicine [3, 5, 17, 18]. Thus, this article reviews big -omic and EHR data analytics for precision medicine with key terms summarized in Tables I and II. Section II presents -omic and EHR data characteristics, challenges, and big data analytics; Section III uses two case studies to illustrate the impact of big data analytics in precision medicine; Section IV enumerates several well-known biomedical big data initiatives; Section V discusses current opportunities in big data analytics for precision medicine; and finally, Section VI concludes this article.

Fig. 1.

The key types of biomedical big data for precision medicine.

TABLE I.

Biomedical Big Data Keywords

Topics	Keywords
-Omic Data	Genomics, transcriptomics, epigenomics, proteomics, metabolomics, etc.
EHR Data	Big data in EHR, next-generation EHR, clinical data management, medical coding systems, etc.
Data Challenges	Biomedical big data challenges, -omic data challenges, EHR data challenges, etc.
-Omic Data Analytics	NGS sequence mapping, NGS variant detection, RNA-seq computation, ChIP-seq computation, MS pre-processing, NGS biomarker identification, NGS differential analysis, -omic network analysis, -omic dynamic modeling, etc.
EHR Data Analytics	Temporal medical data mining, irregular time series analysis in EHR, clinical decision support, unsupervised/ supervised learning in EHR, waveform analysis in EHR, etc.
Big Data Analytics Enablers	Big data harmonization, big data platform, big data framework

TABLE II.

-Omic and EHR Data Concept Glossary

Term	Definition	Ref.
Genome	“An organism’s complete set of DNA”	[20]
Transcriptome	“A collection of all the gene readouts present in an organism’s cell”	[20]
Epigenome	“A multitude of chemical compounds that can tell the genome what to do”	[20]
Proteome	“An entire set of proteins encoded by the genome”	[21]
Metabolome	“A comprehensive catalogue of metabolites in an organism’s cell”	[21]
-Omics	“The study of the -ome”	[21]
Single Nucleotide Polymorphism	“A variation at a single position in a DNA sequence among individuals”	[22]
Frameshift Mutation (Indels)	“A genetic mutation caused by a deletion or insertion in a DNA sequence that shifts the way the sequence is read”	[22]
Copy Number Variation	“The number of copies of a particular genetic sequence differs between individuals”	[22]
Structural Variation	“Genomic alterations that involve segments of DNA that are larger than 1 kb, and can be microscopic or submicroscopic”	[23]
Fusion Gene	“A new gene formed by the breakage and re-joining of two different genes”	[24]
Spanning Read Pair	“paired reads that harbor a fusion boundary in the insert sequence”	[25]
Split Read	“A read that harbors a fusion boundary in the read itself ”	[25]
Alternative Splicing	A process that includes or excludes certain exons when forming mature mRNAs	[26]
Protein-DNA Binding Site	A segment of DNA sequences where targeted proteins may bind	[27]
Histone Modification	“A covalent post-translational modification to histone proteins that can impact gene expression”	[27]
DNA Methylation	“The addition of methyl (CH3) group to DNA that modifies the function of the genes”	[27]
Antecedent (Ant.) in ARL	A set of conditions which the outcome variable depends on	[28]
Consequent (Cons.) in ARL	A set of conditions serving as the outcome variable	[28]
Confidence in ARL	$\mathit{\text{Confidence}}(\mathit{\text{Ant}}.\Rightarrow \mathit{\text{Cons}}.)=\frac{\text{count}(\mathit{\text{Ant}}.\cup \mathit{\text{Cons}}.)}{\text{count}(\mathit{\text{Ant}}.)}$	[28]
Fast Healthcare Interoperability Resources (FHIR)	FHIR uses standardized “Resources” (i.e., predefined data formats and elements) to exchange EHR data.	[29]

II. Big Data For Precision Medicine

The invention of high-throughput -omic assays such as next-generation sequencing (NGS) and mass spectrometry (MS) has led to fast accumulation of -omic data. Likewise, the wide adoption of EHR for the entire population provides a foundation for studying healthcare efficiency and safety [19]. -Omic data analytics often aims at finding biomarkers by cleaning up raw data generated by NGS or MS, extracting molecular profiles, identifying statistically significant molecules, constructing models describing molecular interactions or temporal system behavior, and validating putative biomarkers. EHR data analytics typically aims at predicting future outcome based on population and individual longitudinal data. The analytics has a similar process such as data cleaning, clinical features identification, predictive model construction, and clinical validation.

A. Biomedical Big Data

A.1. Big -Omic Data

-Omic data contain a comprehensive catalog of molecular profiles (e.g. genomic, transcriptomic, epigenomic, proteomic, and metabolomic as explained in Table II) in biological samples that provide a basis for precision medicine [17]. The genome, transcriptome, and epigenome are upstream of the proteome and metabolome. A genome is unique and mostly invariant over time with its knowledge embedded in single nucleotide polymorphisms (SNPs), frameshift mutations (insertions or deletions; or indels), copy number variations (CNVs), and other structural variations (SVs) [30, 31]; transcriptomic knowledge is contained in gene expression, transcript expression, gene fusion, and alternative splicing [32, 33]; epigenomic knowledge is carried in protein-DNA binding sites, histone modification patterns, and DNA methylation patterns [34]; proteomic knowledge is reflected by protein expression, post-translational modification, and protein-protein interactions [35]; and metabolomic knowledge is shown in the abundance of metabolites [36]. Because epigenomic information impacts transcriptomic, proteomic, and metabolomic profiles [37], and the proteome and metabolome are directly responsible for the establishment of phenotypes, uncovering interactions among the proteome, metabolome, and upstream processes is a key towards precision medicine.

A.2. Big Electronic Health Record Data

EHR data can be unstructured (e.g., clinical notes) or structured (e.g., ICD-9 diagnosis codes, administrative data, chart, and medication) [38]. Written or dictated clinical notes describe the patient’s condition and are the most efficient and human-intuitive way for clinical documentation. However, they are the most challenging for computer analysis because of (1) unstructured and heterogeneous data formats, (2) abundant typing and spelling errors, (3) violation of natural language grammar, and (4) rich domain-specific abbreviations, acronyms, and idiosyncrasies [39]. Structured EHR data can be categorized into two classes [40]. Administrative data include those remain unchanged during the entire course of a clinical encounter (e.g., demographic data), and those keep updating over time (e.g., diagnoses and procedures) [41]. Ancillary clinical data are frequently recorded during a clinical encounter that can be discrete (e.g., physiological measures, medication, and lab tests), or continuous (e.g., respiration, blood pressure, pulse oximetry, and electrocardiography waveforms captured by sensors, either through bedside monitoring devices or ambulatory, implanted, or wearable devices) [40].

B. Challenges Associated with -Omic and EHR Data

-Omic and EHR big data analytics is challenging due to data frequency; quality; dimensionality; and heterogeneity.

B.1. Diverse Data Collection Frequency

First, different data modalities have different data collection frequency. For example, in -omic data, a genome is invariant over a long period of time, and often only needs a one-time data acquisition, while other types of -omic data vary with environment, tissue types, and time that require multi-time-point acquisition. In EHR, bedside monitoring data are captured at very high frequency, while lab tests may be taken a few times a day. In addition, data generation frequency may be influenced by cost (e.g. proteomic/metabolomic data generated by MS versus genomic/transcriptomic/epigenomic data generated by NGS). Second, data collection frequency can be irregular. For example, in EHR, most clinical variables have irregular sampling frequencies, depending on the criticality of a patient and the easiness of a measurement.

B.2. Inherent Data Quality Issues

In -omic data, quality issues are caused by a combination of biological, instrumental, and environmental factors such as sample contamination [42, 43], batch effects [44, 45], and low signal-to-noise ratios [46, 47]. In EHR data, quality issues include missing data because recorded clinical variables vary with each clinical encounter and depend on clinical team’s assessment of the patient’s condition [48], and erroneous data entries happening due to data entry mistakes or misinterpretation of original documents when entering [49]. For high-resolution waveform data, common quality issues include random noise, gaps in the waveform, and artifacts (e.g., patient’s motion) [50]. These data quality issues may lead to wrong conclusion, but correcting these remains challenging.

B.3. High Dimensionality

A big challenge in either -omic or EHR data mining is the “curse of dimensionality” associated with high-dimensional data [51]. -Omic data often have many dimensions or features (may be more than 10⁴) much larger than the number of samples available, while EHR data may contain a large sample size of high-dimensional data but with each individual sample only sparsely populated. Making sense of these data with statistical significance presents to be challenging.

B.4. Heterogeneous Data Types

In -omics, using underlying molecular fingerprints to characterize disease subtypes may require heterogeneous multi-omic data. For example, the integrative personal omics profile (iPOP) project has integrated multiple molecular expression profiles to uncover dynamic molecular changes between healthy and diseased states [52]. However, integrating multi-omic data is challenging because of variations in represented biological processes, technical and biological noise levels, identification accuracy, spatiotemporal resolution, and many other confounding factors [53]. In EHR, the data are inherently heterogeneous. To accomplish precision medicine, it is necessary and critical to make sense of heterogeneous data.

C. Big Data Analytics for Precision Medicine

C.1. General Analytics for Biomedical Big Data

Most -Omic and EHR are high-dimensional data that not only require longer computational time but also affect the accuracy of analysis. Thus, we try to reduce data dimensionality by identifying a subset of variables or latent factors that preserve as much of the characteristics of the original data as possible with two strategies (Table III): (i) feature selection that aims to select an optimal subset of existing features, and (ii) feature extraction that aims to transform existing features into a more compact set of dimensions [56].

TABLE III.

Selected Methods for Dimensionality Reduction

Method	Advantages	Limitations
Feature extraction: PCA, SVD, tensor-based approaches* [54]	Reduces dimensionality; relatively immune to noise	Performance usually inferior to supervised approaches; difficult to interpret results
Feature selection: filter- based (mRMR), wrapper- based (sequential feature selection)* [55]	Reduces dimensionality; easy to interpret	Sometimes affected by noisy data

^*Highly impactful method with more than 50,000 relevant papers.

Feature selection techniques consist of filtering, wrapper, or embedded methods. Filtering methods limit the number of features by calculating a score designed to estimate the usefulness of each feature. Thus, they are generally faster and do not require explicit class labeling. The minimum redundancy maximum relevance (mRMR) method is a filtering method that iteratively selects features sharing the most mutual information (relevance) with the least redundancy [57]. In contrast, wrapper methods select a subset of features (i.e. “wrap” the feature selection) for targeted learning models by using evaluation metrics such as cross-validation accuracy [58]. Embedded methods integrate machine learning algorithms (e.g. support vector machines) with recursive feature elimination [59].

Among feature extraction techniques, principal component analysis (PCA) is a basic method that identifies a small number of orthogonal linear vectors [51]. Its performance heavily depends on correctly identifying an optimal number of components, and requires careful testing and validation [60]. Other techniques include artificial neural networks such as autoencoders [61], and nonlinear kernels in PCA [62].

C.2. -Omic Data Pre-processing

NGS and MS high-throughput assays require different pre-processing methods that are summarized in Table IV. NGS is a popular assay for genomic, transcriptomic, and epigenomic studies. Its common pre-processing step is sequence mapping that identifies not only the origin but also the alignment of each read [78]. This step is computationally intensive and requires auxiliary data structures (e.g., the hash table [63] and the Burrows-Wheeler transform [64]), multithreading, or in-memory computing [65] for improved computational efficiency. Genomic studies typically aim to identify variants in a sequenced genome [79]. Small-scale variant (i.e., SNPs and indels) detection uses “per base differences” between reads and the reference genome as the evidence [30, 66]. Large-scale variant (i.e., SVs) detection uses read-pair-based, read-depth-based, split-read-based, and assembly-based methods [80, 81]. Transcriptomic studies mostly center on expression profiling, fusion gene detection, and alternative splicing detection [32, 33]. Expression profiling associates mapped reads with genes and isoforms. Different profiling methods handle multi-mapped reads differently, where some methods associate the reads with all loci [67, 68], while others probabilistically associate the reads with only a few model-inferred loci [69, 70]. Fusion gene detection relies on two factors, the spanning read pairs and the split read [24, 25]. Alternative splicing detection relies on either de novo transcriptome assembly [71, 72, 82–84], or inference from sequence mapping outputs [70, 73]. Epigenomic studies mainly focus on identifying patterns of protein-DNA binding sites, histone modification, and DNA methylation [34]. Epigenomic data pre-processing builds a profile representing the density of reads along the genome, models background noises, and determines statistically significant peaks [78].

TABLE IV.

Selected Tools for -Omic Data Pre-processing

Tool	Assay	-Omic Data	Key Functionality
GMAP* [63]	Next- generatio n sequenci ng	Genomic, transcriptomic , and epigenomic	Sequence mapping
BWA* [64]
STAR* [65]
GATK* [30]		Genomic	Genomic variant discovery
SAMtools* [66]
HTSeq* [67]		Transcriptomi c	Gene and transcript expression quantification
BEDTools* [68]
RSEM* [69]			Gene and transcript expression quantification
Cufflinks* [70]
defuse [25]			Gene fusion detection
TopHat-Fusion [24]
Trans-ABySS* [71]			Alternative splicing detection and quantification
Trinity* [72]
Cufflinks* [70]
Scripture* [73]
MACS* [74]		Epigenomic	ChIP-seq peak calling
SISSRs* [75]
OpenMS [76]	Mass spectrom etry	Proteomic and metabolomic	Peak detection and quantification
MZmine 2* [77]

GMAP stands for genomic mapping and alignment program; BWA, Burrows-Wheeler aligner; STAR, spliced transcripts alignment to a reference; GATK, genome analysis toolkit; RSEM, RNA-seq by expectation-maximization; Trans-ABySS, transcriptome assembly and analysis pipeline; MACS, model-based analysis of ChIP-seq; and SISSRs, site identification from short sequence reads.

^*Highly impactful tool with more than 50 citations per year.

MS is for proteomic and metabolomic studies, and its pre-processing steps include alignment, baseline correction, and peak detection [85]. In chromatography-coupled MS, chromatographic peak alignment can correct drift to ensure coherent retention time and accurate mass across samples, and mass-to-charge ratio (m/z) alignment can ensure mass spectra and component features are comparable among samples [86]. In MALDI (matrix-assisted laser desorption ionization) MS, baseline correction is particularly important. Low-mass measurement noise from the chemical matrix used in MALDI experiments affects the spectral baseline and needs to be removed prior to analysis [87]. Peak detection is then performed based on criteria such as signal-to-noise ratio, peak shape, and detection thresholds [88]. A common subsequent step is the identification, and potentially the filtering, of isotopic peaks from the spectrum [77].

C.3. Biomarker Identification Using -Omic Data

In practice, different groups of samples are collected for different biological conditions (e.g., disease vs. non-disease) or different time points (e.g., before vs. after a treatment). Thus, Table V summarizes selected tools that identify discriminatory biomarkers among different groups. Most -omic biomarkers are identified by investigating statistically significant differences among groups, such as differentially expressed genes or transcripts [94–96], differential alternative splicing [97, 98], differential protein-DNA binding [99], differential histone modification [100], and differential DNA methylation [101]. The basic idea is to quantify and then fit the abundance of each group to Poisson-based distributions (e.g., the Poisson distribution and the negative binomial distribution), followed by statistical tests (e.g., the Fisher’s exact test and the likelihood ratio test) that determine the statistical significance of each molecular feature. For genomic data, genome-wide association studies (GWAS) uses different approaches (e.g., the chi-squared test or logistic regression) to assess the degree of association between each variant and a targeted trait, and then select most significant variants as biomarkers [105]. Most GWAS focuses on SNP association [89–91], while only a few infer CNV or SV association [92, 93].

TABLE V.

Selected Tools for -Omic Biomarker Identification

Tool	-Omic Data	-Omic Biomarker	Approach
SNPassoc [89]	Genomic	Significant SNPs associated with traits	Genome- wide association studies
SNPTEST* [90]
VAT [91]		Significant SNPs and indels associated with traits
PLINK* [92]		Significant SNPs, indels, and CNVs associated with traits
CNVRuler [93]		Significant CNVs associated with traits
edgeR* [94]	Transcriptomic	Differentially expressed genes /transcripts	Differential analysis (model fitting and statistical tests)
DESeq2* [95]
omniBiomarker [96]
DiffSplice [97]		Differential alternative splicing
MATS [98]
DBChIP [99]	Epigenomic	Differential binding sites
ChIPDiff [100]		Differential histone modification sites
QDMR [101]		Differentially methylated regions
DetectTLC [102]	Proteomic and metabolomic	Molecular patterns in mass spectrometry images	Similarity scoring
Automics [103]		Differentially abundant metabolites	Supervised and unsupervised learning
MetaboAnalyst* [104]

SNPassoc stands for SNP-based whole genome association studies; VAT, variant association tools; PLINK, population-based linkage analyses; edgeR, empirical analysis of digital gene expression data in R; MATS, multivariate analysis of transcript splicing; DBChIP, differential binding with ChIP-seq data; and QDMR, quantitative differentially methylated regions.

^*Highly impactful tool with more than 50 citations per year.

C.4. Systems Biology Modeling Using -Omic Data

To gain insights about a complex molecular system, we can conduct systems biology modeling using either “static network analysis” or “dynamic temporal analysis” based on -omic features (Table VI).

TABLE VI.

Selected Tools for -Omic Data Modeling

Tool	Modeling Type	Approach	Key Functionality
WGCNA* [106]	Static Network analysis	Correlation between quantitative variables	Network construction, module detection, and gene selection
CODENSE [107]		Summary graphs and dense subgraphs for frequent edges	Mining frequent coherent dense subgraphs across large numbers of massive graphs
MEMo* [108]		Mutually exclusive genomic alterations	Network construction, module detection, and gene selection
CellDesigner [109]	Dynamic Temporal Analysis	Ordinary and partial differential equations	Graphical interface for ODE or PDE model implementation and simulation; systems biology markup language compatibility
NetLogo* [110]		Agent-based models	General-purpose modeling environment capable of simulating hundreds to thousands of interacting agents
BoolNet [111]		Boolean models	Simulating and analyzing Boolean and probabilistic Boolean models
Snoopy [112]		Petri nets	Network modeling using Petri nets; hierarchical structure and multiple class compatibility

WGCNA stands for weighted correlation network analysis; CODENSE, coherent dense subgraphs; and MEMo, mutual exclusivity modules in cancer.

^*Highly impactful tool with more than 50 citations per year.

Static network analysis studies the interactome (i.e., a complete set of molecular interactions) with three steps [113]: identifying a network scaffold that describes interactions among -omic features [106, 108]; decomposing the network scaffold into smaller network modules [106–108]; and mathematically representing each network module for downstream simulation and analysis [114]. Most interactome networks use a single -omic data such as metabolic networks and gene regulatory networks. Few incorporate multi-omic data but are limited to simpler organisms (e.g., S. cerevisiae and C. elegans) [115].

Dynamic temporal analysis (e.g., ordinary or partial differential equations, Boolean networks, agent-based models, and Petri nets [116]) makes use of temporal measurement of - omic data to develop and validate dynamic predictive models of complex systems. For example, a recent study on A. thaliana used a Granger causality model to integrate two types of metabolomic data acquired at multiple time points for studying the interaction of primary and secondary metabolism [117].

C.5. EHR Data Pre-processing

Information embedded in EHR is abundant but disorganized in nature. Thus, EHR data requires systematic pre-processing that are summarized in Table VII. On EHR missing data, conventional approaches either impute missing values by the mean or median in a population, or list-wise or pair-wise delete records with missing values. These approaches are simple and easy to implement, but they ignore the underlying data structure and tend to introduce additional biases [128]. Thus, more robust missing data imputation methods such as interpolation [129], multiple imputation [130], expectation maximization [131], and maximum likelihood [132] are needed.

TABLE VII.

Selected Methods for EHR Data Pre-processing

Method	Advantages	Limitations
Missing data: list-wise deletion, mean filling* [118, 119]	Simple to implement; complete case analysis	Loss of statistical power; introduces biases; underestimates variances
Missing data: hot deck, nearest neighbor* [120]	Simple to implement and interpret; immune to cross-user inconsistencies	Introduces biases; underestimates variances
Missing data: interpolation (linear, piece-wise linear, spline, cubic) [121]	Simple to implement and interpret; direct estimation on the basis of neighbors	Does not account for relationships among different features
Missing data: model-based filling (expectation maximization, maximum likelihood, multiple imputations)* [122]	Accounts for uncertainty in imputations	Does not account for missing data mechanisms (i.e., MCAR, MAR, and MNAR)
Waveforms: noise filtering (IIR, FIR, PCA, ICA, Kalman filter, wavelets) [50, 123]	Generally simple to implement	Falls short in situations where “true” waveform is obscured by artifact such as patient motion
Waveforms: signal quality indices [123–125]	Human-interpretable metrics of signal quality	Can be complex to implement and computationally intensive; may require ad-hoc calibration based on the features of the target waveform
Waveforms: sensor fusion [126, 127]	Improved SNR; reduces data dimensionality while increasing data quality	Computationally intensive; loss of detail from individual sensor waveforms

^*Highly impactful method with more than 50,000 relevant papers.

On high time-resolution waveform data quality issues [50], we can use (i) filtering strategies such as median filtering, Kalman filtering, and model-based filtering to handle noise [50, 123]; (ii) signal quality indices that detect the presence of expected physiological features, quantify the agreement between signals with mutual information, or infer other ad hoc definitions of signal quality to identify artifacts and gaps in the waveform [123–125]; and (iii) sensor fusion techniques (e.g. using redundant measurements of electrocardiography, blood pressure sensors, and photoplethysmography to derive a more reliable measure of heart rate than any single signal alone [126, 127]) to correct artifacts and fill in gaps in the waveform.

C.6. EHR Data Mining

To derive actionable knowledge from complex EHR big data, two strategies such as static endpoint prediction and temporal data mining are summarized in Table VIII.

TABLE VIII.

Selected Methods for EHR Data Mining

Method	Advantages	Limitations
Logistic regression, cox regression, local regression (LOESS)* [133]	Simple to implement and interpret; direct estimates of relevant hazards for Cox regression	Sensitive to outliers
Logistic regression with LASSO regularization [134]	Reduces feature space	Prone to overfitting
Hidden Markov models [135]	Simultaneous detection, segmentation, and classification in a waveform	Sensitive to the design of the Markov model being trained
Conditional random fields [136]	Supports temporal analysis; resistant to differences in class prevalence	Sensitive to regularization and feature space size
Relational subgroup discovery, episode rule mining, windowing* [137]	Valid sequential techniques for some clinical applications	Tradeoffs between simplicity, complexity, and temporal resolution
Rule mining, Allen’s interval algebra, directed acyclic graph* [138]	Temporal mining/modeling capabilities	Requires specific experimental design

^*Highly impactful method with more than 50,000 relevant papers.

C.6.1. Static Endpoint Prediction

After dimensionality reduction, we can model the relationship between selected clinical features (i.e. the patient’s condition) and targeted clinical endpoints (i.e. the clinical outcome) with three groups of techniques: Regression analysis is a statistical process that estimates the relationship between independent variables (i.e., features) and dependent variables (i.e., endpoints). If dependent variables follow distributions such as normal, Poisson, and binomial, we can use a generalized linear model for regression model fitting; Classification involves building statistical models that assign a new observation to a known class. Many classification techniques such as decision trees, k-nearest neighbors, and support vector machines (SVM) prove to be effective in clinical applications; Associate Rule Learning (ARL) discovers frequent and reliable associations among clinical variables, and these association rules describe that if all elements in the antecedent occurs, all elements in the consequent should occur with certain confidence [28]. In general, these machine learning techniques prefer a large sample size.

C.6.2. Temporal Data Mining

EHR captures diagnosis, treatment, and outcome chronologically throughout a medical encounter, and thus, it is important to model temporal relationship between events using temporal data mining techniques such as the hidden Markov model (HMM) and the conditional random field (CRF) [135, 136]. One constraint of HMM and CRF is that they require predefined clinical variables and outcome categories ofen difficult to generalize for a given treatment of a given patient. Thus, temporal association rule mining (TARM) is proposed to discover causality between the event and outcome. A temporal association rule, denoted by A →_T C, describes an antecedent A followed by a consequent C separated by a time difference T. Because the selection of the event and outcome is flexible, TARM model can be tailored for any event-outcome combination in various clinical settings [137, 139].

D. Enablers of Biomedical Big Data Analytics

The big data revolution has led to the development of enterprise tools and platforms for extracting, summarizing, and interpreting knowledge from rapidly generated data, for business intelligence, analytics, and predictive modeling as summarized in Table IX [11, 140, 141].

TABLE IX.

Selected Platforms for Big Data Analytics

Platform	Advantages	Limitations
Apache Hadoop (MapReduce)* [11, 142]	Horizontally scalable; fault- tolerant; designed to be deployed on commodity-grade hardware; free and open-source	Generally most effective for batch-mode processing; not always appropriate for real- time, online analytics
IBM InfoSphere Platform* [143]	Includes purpose-built tools to handle streaming information; integrates with open-source tools such as Hadoop	Commercial licensing
Apache Spark Streaming* [144]	Integrates with the Hadoop stack; allows one code base for both batch-mode and online analysis	Depends on more expensive hardware with large amounts of RAM to work efficiently
Tableau, QlikView, TIBCO Spotfire, and other visual analytics tools*	Visualization of large and complex data sets	Generally incomplete solutions, requiring other tools to effectively handle data storage

^*Highly impactful platform.

Distributed computing systems such as Apache Hadoop (based on MapReduce) provide the storage and processing backbone for dealing with very large datasets [142]. Specific tools also exist to solve more specialized problems. For example, IBM InfoSphere Streams and Apache Spark Streaming can handle real-time streaming data [143, 144]. Cloud computing providers such as Amazon Elastic Compute Cloud (EC2) can provide on-demand computing power to accommodate scalable growth from development to truly big data production [145]. Many cloud-based services in bioinformatics such as Illumina’s BaseSpace [146] and the Galaxy project [147] are deployed on Amazon EC2.

To deploy biomedical big data for precision medicine in health care, there is a critical need to address domain-specific challenges such as the requirements of HIPAA (Health Insurance Portability and Accountability Act), HITECH (Health Information Technology for Economic and Clinical Health), and other privacy regulations. Thus, security is an important enabling technology (e.g., encryption for protected health information) in biomedical big data [140, 145, 148].

III. Case Studies

In this section, we present two real-world applications to illustrate the utility of biomedical big data analytics for precision medicine: (1) integrative -omic data for the improved understanding of cancer mechanisms (see Fig. 2), and (2) the incorporation of genomic knowledge into the EHR system for improved patient diagnosis and care (see Fig. 3).

Fig. 2.

Integrative analysis of multi-omic data leads to the improved understanding of cancer mechanisms, which in turn enables more precise classification of cancer subtypes.

Fig. 3.

Integrating derived -omic knowledge into the existing EHR system is an approach to utilizing molecular information for clinical decision support, and it also help deliver precision medicine.

A. Integrative -Omics for Precise Cancer Understanding

One notable effort that integrates multi-omic data for the improved understanding of cancer mechanisms is The Cancer Genome Atlas (TCGA) [149]. TCGA hosts public datasets of 27 cancer types with more than 11,000 patient cases. Each patient is annotated with clinical data (i.e. demographic, diagnostic, and survival data) and multimodal -omic data (i.e., genomic, transcriptomic, epigenomic, and proteomic).

We use head and neck squamous cell carcinoma (HNSCC) as an example to illustrate the integrative multi-omic study for precision medicine [150]. In 2014, a pan-cancer study with twelve cancer types using multi-omic TCGA data was performed [151]. Among 3,527 samples in total, 305 were HNSCC. Six different data types (i.e. DNA copy number, methylation, mutation, and expression of mRNAs, miRNAs, and proteins) were analyzed both separately and integratively. By using clustering-based methods, pathway activities (inferred from gene expression and copy number data) have shown common copy number variations, mutation frequency patterns, and survival patterns between HNSCC and lung squamous cell carcinomas or some bladder cancers. Such integrative pan-cancer analysis provides more precise subtyping across multiple cancers sharing common molecular-level processes underlying cancer development. This new subtyping system reflects precision medicine because it finds precise classification of patients into disease subgroups.

TCGA Research Network has published more than 30 articles describing multi-omic investigation on numerous cancer types, and identified more precise, clinically relevant subtyping for multiple cancers [152–154].

B. Adoption of Genomics in EHR for Precision Medicine

In a clinical setting, healthcare providers use electronic medical record (EMR) for clinical decision support. Thus, it is important to incorporate -omic data and knowledge into EMR. The Electronic Medical Records and Genomics (eMERGE) Network consortium aims to identify causal genomic variants (mostly SNPs) for EMR-based phenotypes and to integrate identified genotype-phenotype associations into the EMR system [155]. One crucial challenge is on how to store variants present in an individual or even in family members in the EMR [156]. The consortium has proposed several recommendations on augmenting the current EMR structure: (1) it should store various genomic variants, such as SNPs, indels, and CNVs, in a discrete computable format; (2) it needs to satisfy interoperability to reduce the burden in data transfer and update within and between healthcare facilities; (3) it has to support rule-based decision support engines; and (4) it must contain abundant visualization elements for easier interpretation [157]. Another big challenge is that each individual typically has millions of variants. The consortium has proposed one potential solution that stores only known pathological variants in the EMR system. However, because the set of known pathological variants may change over time, this approach may lead to the inclusion of false positive and the exclusion of false negative variants. Thus, an alternative solution is to archive raw data in separate repositories easily accessible when necessary [158].

EMR is only for local clinic and hospital, while EHR contains and shares medical records among all participant clinics and hospitals [159]. Thus, interoperability is critical in using big data for precision medicine. Recently, the Health Level Seven International (HL7) proposed the Fast Healthcare Interoperability Resources (FHIR) standard that addresses this important issue. On clinical genomics, several new FHIR resources and extension definitions are designed for variant data [160]. With such the standardized data exchange protocol, clinicians can utilize genomic information with other existing EHR data to determine the most effective treatment for each patient, which is a paradigm shift towards precision medicine.

IV. Biomedical Big Data Initiatives

To apply big data analytics for precision medicine, Table X summarizes multiple consortium initiatives that collect and organize data from various projects and trials, and make them available to the research community for secondary data use.

TABLE X.

Selected Biomedical Big Data Initiatives

Consortium	Data Sources	Data Elements
TCGA	Multi-omic data for 27 cancer types, covering more than 11,000 cases	Clinical, genomic, transcriptomic, epigenomic, and proteomic data
Project Data Sphere	Patient-level data from comparator arms of Phase IIB and III clinical trials; currently contains 33 trials covering 12 cancer types	Common data include baseline, safety, efficacy, medication, dosing, lab test, medical history, and demographic data
TARGET	Multi-omic data for 7 types of childhood cancers	Clinical, genomic, transcriptomic, and epigenomic data
1,000 Genomes Project	Large-scale genome sequencing project for populations of African, European, and East Asian ancestry	Low-coverage whole genome sequencing for 179 individuals; high-coverage targeted exome sequencing for 697 individuals
100,000 Genomes Project	Large-scale genome sequencing project for studying cancers and rare diseases in the UK	Genome sequencing will be completed in 2017
ICGC	Genomic data for 18 cancer types; partially overlap the TCGA data	SNPs, CNVs, methylation, and gene and miRNA expression
RD- Connect	Infrastructure project funded by European Union for facilitating rare disease research	Currently links to 3 biobanks and more than 150 rare disease registries
ADNI	Multi-center, longitudinal study with elderly control subjects, early Alzheimer’s disease subjects, and mild cognitive impairment subjects	Clinical, genetic, magnetic resonance imaging, and positron emission tomography imaging data
iDASH	Data from 17 focused trials, each of which represents a specific objective and a patient population	Imaging, EHR, sensor, and genomic data from multiple clinical trials
GHO	Worldwide population and environmental data for infectious diseases, noncommunicable diseases, sexually transmitted diseases, and children’s health	Population-level statistics and modeling
BMIC	Large trials encompassing thousands of samples	EHR, imaging, genetic, and social research data
MIMIC II	ICU data for more than 30,000 patients with more than 40,000 ICU stays	Chart data, administrative data, alert data, lab results, electronic documentation, and bedside monitor trends and waveforms
HIW	Federal data for aggregated health indices by geography; covers data from claims, healthcare cost, to population statistics	Data element varies, depending on the trials

TCGA stands for The Cancer Genome Atlas; TARGET, Therapeutically Applicable Research to Generate Effective Treatments; ICGC, International Cancer Genome Consortium; ADNI, Alzheimer’s Diseases Neuroimaging Initiative; iDASH, Integrating Data for Analysis, Anonymization, and Sharing; GHO, WHO Global Health Observatory Data Repository; BMIC, Trans-NIH BioMedical Informatics Coordinating Committee; MIMIC II, Multiparameter Intelligent Monitoring in Intensive Care II; and HIW, Health Indicators Warehouse.

First, initiatives such as Project Data Sphere aim to improve research efficiency and to encourage collaboration by integrating information of clinical trials for different cancers. For example, the European Union-funded RD-Connect aggregates data of multiple rare diseases from around the world. The Cancer Genome Atlas (TCGA) of US, Therapeutically Applicable Research to Generate Effective Treatments (TARGET), and the International Cancer Genome Consortium (ICGC) aim to study multiple aspects of individual diseases by collecting multi-omic data of hundreds of patients for each cancer type. In contrast, the US 1,000 Genomes Project and the UK-based 100,000 Genomes Project aim to connect genotypes with phenotypes using single -omic data type.

Second, large data repositories have been created and maintained by organizations such as US National Institutes of Health (NIH) and the World Health Organization (WHO). The Trans-NIH BioMedical Informatics Coordinating Committee (BMIC) has established a data repository that archives the data from 61 large multi-center studies for promoting secondary use of biomedical data [161]. The Global Health Observatory Data Repository (GHO) is maintained by WHO for population-level health studies [162]. As another example, the Health Indicators Warehouse (HIW) within the US Department of Health and Human Services provides country-level and state-level aggregated clinical information [163].

V. Discussion

Among many data types included in the NIH Big Data to Knowledge (BD2K) Initiative, -omic data, EHR, and medical imaging data are the three most important biomedical big data. We conducted the review of -omic and EHR data because of their close relationship with precision medicine [3, 5, 17, 18].

Big data have had major societal impact in energy, environment, financial, and others. They motivate rapid advances in data storage, data mining and analytics, data retrieval, and data visualization [164, 165]. When applying to biomedicine and healthcare, big data will improve quality and outcome by (i) discovering new knowledge (e.g., automated identification of postoperative complications in EHR data [166]); (ii) disseminating new knowledge (e.g., data-driven clinical decision support systems such as IBM Watson); (iii) incorporating -omic data into EHR (e.g. eMERGE network [167]); and (iv) implementing patient-centered care (e.g., e-health [168]).

To accelerate the delivery of precision medicine, more research is needed in the following biomedical big data areas:

1. -Omic Data Integration: As illustrated by the TCGA case study, integrative multi-omic data analysis is of growing importance because it provides holistic view of molecular fingerprints for each patient’s condition. Recent research has shown positive impact of knowledge and insight obtained from integrative analysis of genomic and transcriptomic [169], transcriptomic and proteomic [170], and multiple -omic data types [53, 151] on disease diagnosis, prognosis, and treatment. The next important direction is the development of guidelines (or best practices) for -omic data integration and interpretation that will in turn enable better prediction of bio-system behavior, and safer and more effective therapeutics.

2. Waveform and Irregularly Spaced Time Series Analysis: Real-time streaming data analytics needs to be further developed due to the pervasive use of wearable sensors in either the critical care setting or in the continuous home monitoring setting for fitness and preventative medicine [171], and the need to reduce alarm fatigue [172]. However, the challenge for irregularly sampled temporal data remains and requires advanced imputation techniques and robust parameter extraction techniques [50, 173].

3. Patient Similarity: Precision medicine promotes precise subgroup classification of patients based on biological basis such as molecular profiles. Thus, EHR mining can assist in patient classification based on clinical measurements (e.g. drug responses, physiological signals, and disease susceptibility). However, because of high patient variability for any disease, the precise subtypes of many diseases remain unknown as of today, and it requires systematic big data analytics to model physician knowledge to validate the reliability of patient subgrouping based on EHR mining.

VI. Conclusion

In this review, we present -omic and EHR big data challenges and current progressed. We provide case studies to show how big data analytics can facilitate precision medicine. Because biomedical big data analytics is in its infancy, more biomedical data scientists and engineers are needed to gain necessary biomedical knowledge, to use large data provided by biomedical big data initiatives, and to put concerted effort in areas such as multi-omic data integration, waveform and time series data analysis, and patient similarity and so on to speed up big data research for precision medicine. By delivering the most suitable and effective treatment to each patient based on their precise subtyping information, the healthcare system can achieve better care efficiency and quality.