Clinical phenotyping in selected national networks: demonstrating the need for high-throughput, portable, and computational methods

Abstract

Objective

The combination of phenomic data from electronic health records (EHR) and clinical data repositories with dense biological data has enabled genomic and pharmacogenomic discovery, a first step toward precision medicine. Computational methods for the identification of clinical phenotypes from EHR data will advance our understanding of disease risk and drug response, and support the practice of precision medicine on a national scale.

Methods

Based on our experience within three national research networks, we summarize the broad approaches to clinical phenotyping and highlight the important role of these networks in the progression of high-throughput phenotyping and precision medicine. We provide supporting literature in the form of a non-systematic review.

Results

The practice of clinical phenotyping is evolving to meet the growing demand for scalable, portable, and data driven methods and tools. The resources required for traditional phenotyping algorithms from expert defined rules are significant. In contrast, machine learning approaches that rely on data patterns will require fewer clinical domain experts and resources.

Conclusions

Machine learning approaches that generate phenotype definitions from patient features and clinical profiles will result in truly computational phenotypes, derived from data rather than experts. Research networks and phenotype developers should cooperate to develop methods, collaboration platforms, and data standards that will enable computational phenotyping and truly modernize biomedical research and precision medicine.

1. Introduction

The national Precision Medicine Initiative aims to enroll one million members in a national cohort that will integrate data from biospecimens, sensor and mobile technologies, and health-care, largely from electronic health record (EHR) data, to advance biomedical discovery and improve health [1]. The realization of this vision will require efficient and effective methods to convert data from EHRs into specific and reliable phenotype characterizations that can be used to predict an individual’s risk of disease or response to drug therapy.

Phenotypes are the measurable biological, behavioral and clinical markers of a condition or disease. The process of deriving research-grade phenotypes from clinical data using computer-executable algorithms is called computational phenotyping (phenotyping for short) [2]. Phenotyping includes a range of approaches from finding a phenotype using expert-derived rules and those phenotypes emerging from novel computational methods that potentially represent new clinical entities. The widespread adoption of EHRs will increase the reliance on phenotyping for a number of activities, including genomic studies of disease and drug response, clinical predictive modeling, pragmatic clinical trials, and healthcare quality measurement. Current methods face bottlenecks for development, implementation, sharability, and the ability to derive novel, not-foreseen findings. We provide a survey of the approaches to computational phenotyping and challenges experienced by several national research networks with which we are affiliated, and provide supporting literature in the form of a non-systematic review. The aim of this paper is to provide a summary of the approaches and tools that clinical research networks are using to realize the scale of high-throughput computational phenotyping. Based on the common challenges faced by these networks, we suggest cultural change and resources that will be needed to support computational phenotyping on a grand scale and advance data-driven precision medicine research.

2. National networks and phenotyping activity

A number of national research and public health surveillance networks have leveraged data from EHRs for defining conditions and risk. The Electronic Medical Records & Genomics (eMERGE) Network, formed in 2007 and arguably the pioneer of computational phenotyping, has investigated more than 40 phenotypes for genomic studies using algorithms that combine billing codes, medication data, laboratory and test results, and natural language processing of clinical notes [3,4]. Sites from the Pharmacogenomics Research Network (PGRN) have used EHR data to identify genetic predictors of drug-response phenotypes across multiple sites [5–7]. The Mini-Sentinel surveillance initiative, funded by the U.S. Food and Drug Administration, uses phenotype algorithms to define conditions from administrative data from 18 national health plans to identify adverse drug outcomes [8–15]. In addition, provider networks use computational phenotyping to identify patients with particular conditions for health services or population-level research. These include the Health Care Systems Research Network, formerly known as the HMO Research Network, and the Observational Medical Outcomes Partnership (OMOP) [16], now part of the Observational Health Data Sciences and Informatics (OHDSI) collaborative [17].

A number of disease-specific research networks and multi-site registries have developed and validated EHR-based phenotype definitions for specific conditions [18,19]. In the National Institutes of Health’s Health Care Systems Research Collaboratory, a number of multi-site pragmatic clinical trial demonstration projects are using computable phenotypes for cohort identification, development of interventions, and study outcomes [20–22] More recently, the Patient Centered Outcomes Research Institute funded the National Patient-Centered Clinical Research Network (PCORnet) to conduct comparative effectiveness studies across 13 Clinical Data Research Networks and 21 Patient Powered Research Networks [23]. Partnering institutions are expected to support up to 200 queries in the next few years, signifying the imminent need for high-throughput and reproducible phenotyping methods.

Although the aforementioned research networks have unique objectives and constraints, they share common challenges related to the use of clinical data for research, including heterogeneous EHR systems, a lack of standardized data, concerns about data completeness and inherent biases, and variation in medical diagnosis, procedures, treatments, and data documentation across providers, organizations, and regions. In response, several networks have published methodological guides for data quality assurance [24–29].

3. Evolution of phenotyping methods

Research networks by their very nature require scalable approaches that can be implemented quickly with reproducible performance characteristics in multiple settings and information systems. There are several broad classes of methods to computational phenotyping that are continuously improving.

The use of expert-defined rules is most widely adopted method for phenotyping, and this approach was used for the early phenotypes developed from the eMERGE network, such as type 2 diabetes [30] and cataracts [31]. This approach begins with the manual development of an algorithm – often using Boolean logic, scoring thresholds, or a decision tree – based on domain expertise. The logic is then iteratively enhanced through validation and chart review on EHR data. Advantages of this approach are that it yields human-interpretable algorithms, which can be portable to other sites [32], and the number of charts needed to review to train/validate an algorithm can be lower. However, the effort and time for developing the algorithms can be significant, requiring clinical and informatics knowledge, and this approach cannot be used to identify phenotypes not first envisioned by a researcher.

Machine learning methods rely on data patterns to develop the phenotype definitions, and can reduce the effort required from clinical domain experts. Supervised learning aims to construct classifiers to differentiate cases (positive for the phenotype) and controls (negative for the phenotype). The high level steps involve (1) characterizing patients as feature vectors, (2) determining the class label (case vs. control) for each patient, (3) building and optimizing the classifier. Typically the number of charts reviewed is higher than required for rule-based algorithms, a time-consuming task requiring domain experts. Chen et al. explored active learning as a more efficient labeling process, demonstrating reduction in the number of cases needed [33]. However, machine learning classification models can be difficult to interpret, require significant training data, and may not transfer well to other sites, as a model may learn features that are unique to an institution (e.g., physician name, local note type, or clinical unit). Yu et al. extracted clinical features from publicly-available knowledge sources to develop more “interpretable” machine learning algorithms that performed as well as or better than expert-derived algorithms [34].

Unsupervised learning provides approaches to cluster EHR data into patient groups corresponding to phenotypes or subtypes. Unsupervised learning does not require expert labels, which tremendously reduces the time needed for manual chart review. However, the validation of the resulting phenotypic groups is challenging, as no clear ground truth on those groups are given. While these methods require very large volumes of training data, they do not carry costs of manually labeling individuals as cases or controls. Various tensor factorization methods have been developed for unsupervised phenotyping [35–37]. Deep learning is another approach which has successfully identified patterns in clinical data representing distinct phenotypes [38].

Because important relevant clinical data is included in narrative clinical notes rather than structured data elements or standardized coding systems, natural language processing methods can be used to extract phenotypes from clinical notes [39,40] and to process data for more advanced machine learning techniques. Phenotype definitions including general purpose natural language processing (NLP) tools [41–43] have accelerated the widespread use of NLP, which is an important component of some complex phenotypes [44].

4. Toward a future of higher throughput phenotyping

The planned Precision Medicine Initiative study will require higher-throughput, more easily shared computational approaches than have been demonstrated to date. Scalable precision medicine will require clinical phenotypes that can be rapidly developed, executed in high volume, and easily adapted to new sites with high algorithm reliability (Fig. 1).

Fig. 1.

The evolution of computational phenotyping methods and the key biomedical applications that needs high-throughput phenotyping.

The vision of rapid, portable phenotyping implies that multiple providers and applications can reuse computational methods and definitional logic, enhanced by accessible repositories for phenotyping logic and methods. While individual research networks undoubtedly have infrastructure for sharing, national repositories for definitions and methods will enable cooperation across networks. The Phenotype Knowledge Base (PheKB) is one such knowledge resource for phenotyping methods, which hosts algorithms from eMERGE, PCORnet, PGRN, and other sites and networks [45]. Many algorithms on PheKB have been validated by multiple sites [28].

Several tools have been developed to simplify “EHR-wide” analyses. These include groupings of billing codes into meaningful phenotypes, such as done by the Agency for Healthcare Research and Quality’s Clinical Classifications Software [46,47] or the phenome-wide association study (PheWAS) tools [3,48–50]. PheWAS approaches have been used for a number of genomics and clinical studies, including recent studying leveraging text. Data driven approaches leveraging PheWAS have also elucidated sub-classifications of phenotypes, such as for rheumatoid arthritis and autism [51]. Other resources, such as mapping medications to their indications or adverse effects [52,53], offer aids to building algorithms. Mo et al. offer desiderata for computable representations of EHR-driven phenotype algorithms [54].

The implementation of phenotypes (in multisite clinical or research networks) can be accelerated by using harmonized clinical data. One approach is the use of a common data model (CDM) to create a common set of definitions, formats, and allowable values for data elements from heterogeneous EHRs. A number of different CDMs are used by different research networks for different purposes, including the OMOP CDM [55] adopted by OHDSI [17], the Mini-sentinel data model [56], the PCORnet CDM [57]. Other clinical data models include the Health Level Seven (HL7) virtual medical record (vMR) model for clinical decision support [58,59], OpenEHR Archetypes [60,61], and the HL7 Detailed Clinical Models [62,63]. An important consideration is that most CDMs are limited to the types of data they include, and unstructured or ad hoc structured data types are often excluded. Thus, many of the phenotype algorithms executed within eMERGE and potentially as a part of precision medicine are not fully addressable via existing CDMs. Novel combinations that facilitate the use of a CDM in tandem with custom applications of NLP or other cohort selection tools will be needed.

For rapid authoring, the Phenotype Execution Modeling Architecture (PhEMA) project is a phenotyping algorithm authoring and execution platform [64] designed to streamline the process for developing computer-readable, rule-based phenotyping algorithms that can be shared as executable representations between different sites. By creating a common representation model for phenotypes (such as the Quality Data Model from the National Quality Forum), PhEMA could enable a common language to be used against a variety of CDMs and EHR systems [65]. Other tests of computable phenotyping approaches have leveraged the Konstanz Information Miner (KNIME) [66,67] or Drools [68]. The recent desiderata recommend use of a common data model, and the use standardized terminologies and ontologies, and facilitate reuse of value sets [54].

5. Discussion

This survey represents the opinions of authors based upon our experience with computational phenotyping within several national research networks, and does not represent a systematic review or national consensus. However, the research networks we represent are leveraging deep multidisciplinary expertise and collective resources to develop new methods and tools for computational phenotyping across heterogeneous organizations, data systems, and populations. These methods promise to significantly advance the identification of disease cohorts and the quantification of an individual’s risk for disease or drug toxicity, and provide path for rapid translational of research findings in healthcare delivery.

The number of national initiatives and activities focused on the use of clinical data is promising, but effective realization of precision medicine on a grand scale will require higher-throughput computational phenotyping. Machine learning approaches have great potential to transform our understanding of disease by allowing phenotypes to be defined by what patients present with, rather than by what research experts know or believe. This will result in truly computational phenotypes, derived from data rather than experts. We might never get to full confidence in the output of a purely machine-generated phenotype, but a goal instead could be to get to a well characterized phenotype, complete with a confidence level or interval, based on the data availability and quality, and the specific use case where it can be applied. New tools will facilitate faster development and implementation of computational phenotyping and lead to more nuanced understanding of “phenotypes.” [69] The latter is a primary goal of precision medicine, to achieve a deeper understanding of diseases and drug response that allow tailoring of therapy toward better health on the basis of clinical and molecular data. In the future, this goal will be enhanced with synergy and harmonization with phenotypes and phenotypic traits defined the biology community [70].

In parallel, national incentives for quality measurement and reporting are driving the development of standardized approaches for deriving research-grade phenotypes from EHR data to advance precision medicine. The future alignment between research phenotyping and quality improvement efforts will enable further efficiencies for both domains. This should include libraries of phenotype definitions, annotated by use case and data required, that can easily be discovered by implementers working in genomic research, precision medicine, or healthcare quality measurement. The next step – and the greater challenge – is to build a culture and supporting infrastructure to share the knowledge and tools that can advance all these efforts. Clinical research networks and phenotype developers should cooperate to develop methods, collaboration platforms, and data standards that will enable high-throughput phenotyping to be implemented across millions of individuals, for a spectrum of use cases from personalized medicine to drug safety surveillance and population health. Successful synergy between clinical research networks, providers, and national initiatives will truly modernize computational phenotyping, biomedical research, precision medicine, and health outcomes.