Integrating laboratory tests for deep phenotyping and biomarker discovery

Clinical laboratory tests are one of the key components of electronic health records and contain rich patient phenotypic information. However, there are often multiple similar laboratory tests for the same medical issue, creating a data integration problem when being used for translational research. In our paper, we presented a novel approach that allows one to semantically integrate laboratory tests by mapping their results into Human Phenotype Ontology terms. The transformation generated detailed patient phenotypic profiles that can be used for biomarker screen.

Electronic health records (EHRs) have been widely adopted in US hospitals since the last decade, generating a vast amount of data waiting to be explored. Clinical laboratory tests are one of the key components of EHR. Because of the objectivity of laboratory tests and the sheer volume that each person can accumulate in routine health care, they are a gold mine for extracting detailed phenotypic profiles of patient, a practice that we referred to “deep phenotyping”. Laboratory tests are uniquely identified with Laboratory Observation Identifier Names and Codes (LOINC), which is a universal code system that defines various kinds of clinical laboratory tests and other measurements. The entire LOINC database have more than 86,000 codes, however, multiple codes are used to represent different ways of measuring the same metabolite. For example, there are five different types of laboratory tests that measure nitrite in urine, which differ in the kind of assays utilized, output formats, or the output units. Although such technical details are useful for uniquely identifying the laboratory tests, they pose a challenge if one hopes to combine them for statistical analysis as they require domain expertise and are hard to scale.

Our group has contributed to patient phenotyping by developing the Human Phenotype Ontology (HPO), which is a vocabulary for describing patient abnormal phenotypes. HPO currently (April 2019) contains ~14,000 terms, many of them represent the medical implications of clinical laboratory tests, such as Hypokalemia and Nitrituria. The terms form a hierarchical tree structure, where the root terms are more generic and the leaf terms are more specific. According to the hierarchical structure of the HPO, if a patient is annotated to a certain phenotype such as Sinus venosus atrial septal defect, then that patient is inferred to have all of the more general ancestor terms such as Atrial septal defect and so on. In addition, HPO also contains disease annotations for all rare and many common diseases. HPO has been adopted by organizations such as the 100,000 Genomes Project, the NIH Undiagnosed Diseases Network and Project, and many others and become the de facto standard for computational genomic diagnosis. In the current paper, we represent laboratory test results with HPO terms, thereby integrating similar laboratory tests together and extracting detailed patient profiles at the same time.  

(Figure credit: adapted from Nicole Vasilevsky)

The essential component of our approach is an accurate map between laboratory test results to their medical implications in HPO terms. For instance, a test for the concentration of potassium in the blood (LOINC:6298-4). If the result is high, our procedure infers the corresponding HPO term for Hyperkalemia (HP:0002153). Analogously, a low result is mapped to Hypokalemia (HP:0002900). The HPO is an ontology of abnormal phenotypes, and thus there is no term that specifically represents a normal test result. However, computational analysis can record negated HPO terms, and the normal test result is represented as NOT Abnormal blood potassium concentration  (HP:0011042). Other types of laboratory tests, such as ordinal type that have negative or positive findings, or nominal type that have a list of potential findings, are also mapped to their corresponding HPO terms. We have a team of multiple MDs and biomedical PhDs to ensure that we achieve the highest possible quality of mappings.

With the mapping in hand, our next step was to devise an algorithm that checks the value of a laboratory test with the reference ranges and then find the correct HPO term for the corresponding result. Because laboratory tests are reported in different formats depending on vendors, we targeted toward laboratory tests encoded as observations in the FHIR format, which is vendor-independent standards for transmitting EHR data. However, it is quite feasible to adapt our software to other formats, such as i2b2 or OMOP.


In our study, we also demonstrated our approach with an EHR dataset on 15,681 de-identified patients with asthma or asthma-like symptoms. We were able to transform 88.6% into HPO terms. On average, each patient were assigned to 633 HPO term-encoded phenotypes, which collapsed to 57 unique HPO terms because patients were tested multiple times with the same laboratory test over the tracking period. Using a logistic regression approach, we identified numerous HPO terms significantly associated with frequent prednisone prescriptions and acute asthma diagnosis, including many terms related to well known prednisone side effects as well as many published asthma biomarkers. Because our approach is very generic, it can be applied to many readily available dataset sitting in data warehouses and EHR systems to study other diseases of interest. Our library can also be coupled with more sophisticated statistical and machine-learning approaches to perform phenotype-driven analysis.

(Aaron Zhang & Peter Robinson)