EHR-based phenome wide association study in pancreatic cancer.

BACKGROUND: Pancreatic cancer is one of the most common causes of cancer-related deaths in the United States, it is difficult to detect early and typically has a very poor prognosis. We present a novel method of large-scale clinical hypothesis generation based on phenome wide association study performed using Electronic Health Records (EHR) in a pancreatic cancer cohort.
METHODS: The study population consisted of 1,154 patients diagnosed with malignant neoplasm of pancreas seen at The Froedtert & The Medical College of Wisconsin academic medical center between the years 2004 and 2013. We evaluated death of a patient as the primary clinical outcome and tested its association with the phenome, which consisted of over 2.5 million structured clinical observations extracted out of the EHR including labs, medications, phenotypes, diseases and procedures. The individual observations were encoded in the EHR using 6,617 unique ICD-9, CPT-4, LOINC, and RxNorm codes. We remapped this initial code set into UMLS concepts and then hierarchically expanded to support generalization into the final set of 10,164 clinical concepts, which formed the final phenome. We then tested all possible pairwise associations between any of the original 10,164 concepts and death as the primary outcome.
RESULTS: After correcting for multiple testing and folding back (generalizing) child concepts were appropriate, we found 231 concepts to be significantly associated with death in the study population.
CONCLUSIONS: With the abundance of structured EHR data, phenome wide association studies combined with knowledge engineering can be a viable method of rapid hypothesis generation.

Introduction

The Health Information Technology for Economic and Clinical Health (HITECH) Act introduced the concept of Meaningful Use of information technology in health care. As part of this process, the legislation mandated the use of standard terminologies for electronic exchange of health information. Patient clinical records represent a largely untapped treasure trove of research information, which only recently has become more accessible thanks to the increasing adoption of Electronic Health Records and healthcare data standards. The need to integrate and exchange clinical data has long been recognized1, but it was the HITECH Act that provided the final piece of the puzzle in terms of financial incentives.

A number of terminology standards are currently in use. LOINC (Logical Observation Identifiers Names and Codes) is a universal standard for identifying laboratory observations2. RxNorm is a standardized nomenclature for generic and branded drugs, as well as drug delivery devices. RxNorm provides normalized names for clinical drugs and links its names to many of the drug vocabularies commonly used in pharmacy management and drug interaction software3. The Healthcare Common Procedure Coding System (HCPCS) maintained by the Centers for Medicare & Medicaid Services (CMS) is a standardized coding system for describing items and services provided in the delivery of healthcare4. It incorporates Current Procedural Terminology (CPT), a coding system maintained by the American Medical Association (AMA) to identify medical services and procedures furnished by physicians and other health care professionals5. International Classification of Diseases, Clinical Modification (ICD-9-CM) is an adaption created by the U.S. National Center for Health Statistics (NCHS) and used in assigning diagnostic and procedure codes associated with inpatient, outpatient, and physician office utilization in the United States6. All these terminologies are integrated within the UMLS (Unified Medical Language System) maintained by the National Library of Medicine (NLM)7.

Current state of the art in extracting actionable information from EHR relies on large scale text-mining and NLP of clinical notes8,9 or either focuses on a specific terminology within the EHR, e.g., ICD-9-CM10,11 or looks into a handcrafted, small subset of EHR variables12. Our approach is novel in the sense that we analyzed the complete corpus of structured data within the EHR across all available terminology standards, as well as used an existing knowledge base (UMLS) to expand and generalize the findings.

Methods

Extract, Load and Transform (ELT)

A Limited Data Set, as defined under the Health Insurance Portability and Accountability Act (HIPAA), was obtained from the Medical College of Wisconsin Clinical Research Data Warehouse for this analysis. The data extract was in the form of standard Epic Clarity tables for a subset of patients that had an encounter or a problem list code in the Malignant neoplasm of pancreas (ICD9:157) code subset:

Malignant neoplasm of pancreas

Malignant neoplasm of head of pancreas

Malignant neoplasm of body of pancreas

Malignant neoplasm of tail of pancreas

Malignant neoplasm of pancreatic duct

Malignant neoplasm of islets of langerhans

Malignant neoplasm of other specified sites of pancreas

Malignant neoplasm of pancreas, part unspecified

Data was loaded into our in-house clinical analytics portal (ClinMiner), which was used to dynamically translate between any of the underlying clinical terminologies, and provided a consolidated view of the underlying patient data in a single UMLS perspective13. Drug information in EHR was encoded using MediSpan terminology, one of the RxNorm sources, which facilitated its automatic translation into UMLS. Labs were encoded as orders using CPT-4 codes or using a fixed category from the CLARITY_COMPONENT lookup table. We have manually mapped 130 tests from CLARITY_COMPONENT to LOINC, which provided coverage for over 97% of all lab observations (1 493 101 observations in total). Remaining ~3% lab observations were left unmapped and excluded from further analysis.

The source annotation space covered 6 617 unique ICD-9, CPT-4, LOINC, and RxNorm codes. This code set was then remapped into UMLS to facilitate further analysis, which resulted in 6 741 distinct UMLS CUIs (Concept Unique Identifiers). This code set was then expanded across a limited set of is_a and selected other relationships (e.g., has_ingredient for RxNorm drugs) as an extension of the method previously proposed in a method similar to that of parent child analysis described by Grossmann et al.14.

But not beyond the original set of UMLS Metathesaurus semantic types of the expanded concepts to exclude functional concepts from the analysis and to keep the general meaning of the originating concept in the expansion. Additionally, the UMLS traversal was limited to either the UMLS Metathesaurus itself, or any of the following terminologies specific to Meaningful Use: RxNorm, NDF-RT, LOINC, SNOMED CT, HCPCS, and ICD-9-CM. This resulted in 18 038 concepts. Finally, we discarded 7 874 concepts that did not increase information content (i.e., were redundant in terms of partitioning of the underlying data) to reach the final ‘phenome’ of 10 162 concepts.

Statistical analysis

A chi-squared test was used to (χ2) to test the significance of the associations. To correct for multiple testing we used a Bonferroni correction and tested at a level of p < 4.9 × 10−6(0.05/10162). Odds Ratio (OR) and Relative Risk (RR) were used to assess the effect size of associations found to be significant.

Results

713 concepts were found to be significantly associated with death in the study population. Where both parent and its child concepts were found to be significant, child concepts were removed to further generalize the results and final result set was thus reduced to 231 terms. A breakdown of all concepts by category and number of observations is shown in Figure!1.

Figure 1:

Breakdown of all 10 164 concepts by semantic type category and proportion of observations annotated with a particular concept. Categories are defined as follows. Lab is any UMLS concept in the semantic type tree: A2.3.1. Clinical Attribute, A2.2.1 Laboratory or Test Result, or B1.3.1.1 Laboratory Procedure. Procedure is a UMLS concept that is in the B1.3.1 Health Activity branch, but is not a 1.3.1.1 Laboratory Procedure. Problem is a concept with a semantic type under B2.2.1.2 Pathologic function or A2.2.2 Sign or symptom. Medication groups all concepts classified by the UMLS Semantic Network either under semantic type A1.4 Substance or under A1.3.3 Clinical Drug. Finally, OTHER groups all other semantic types.

Most of the terms were positively correlated with death and only the following 9 concepts were found to be associated with lower relative risk of death in the study population:

Immunoassay for tumor antigen, quantitative; CA 125

Vitamin D; 25 hydroxy, includes fraction(s), if performed

Prealbumin measurement

Racial group

Benzoic acid or derivative

Iodine AND/OR iodine compound

Ionic iodinated contrast media

Triiodobenzoic Acids

sevoflurane Inhalant Solution

For practical reasons, only the top ten (five from each side) significant associations are shown in Table 1. The complete result set encompassing all 231 significant associations is available as supplementary materials at http://dx.doi.org/10.6084/m9.figshare.816958.

Table 1:

Top ten events by effect size significantly associated with death in the study cohort.

Label	CUI	Semantic Type	Exposed Deceased	Exposed Alive	Not Exposed Deceased	Not Exposed Alive	p	OR	RR
Increased Risk (RR > 1)
Cytopathology, fluids, washings or brushings, except cervical or vaginal; smears with interpretation	C0374051	Laboratory Procedure	28	8	808	310	8.31 × 10−11	9.12	2.80
Cimetidine	C0008783	Pharmacologic Substance	17	6	810	321	2.03 × 10−6	7.14	2.60
Hyposmolality and/or hyponatremia	C0020645	Finding	22	10	806	316	6.54 × 10−7	5.61	2.44
Osmolality; blood	C0373690	Laboratory Procedure	43	25	791	295	2.29 × 10−10	4.61	2.32
Haptoglobin; quantitative	C0373631	Laboratory Procedure	25	14	802	313	1.17 × 10−6	4.57	2.28
Decreased Risk (RR < 1)
sevoflurane Inhalant Solution	C1253873	Clinical Drug	7	85	731	331	1.90 × 10−6	0.18	0.24
Triiodobenzoic Acids	C0041013	Organic Chemical	56	332	484	282	3.00 × 10−15	0.28	0.39
Ionic iodinated contrast media	C0361904	Indicator, Reagent, or Diagnostic Aid	57	332	484	281	6.66 × 10−15	0.29	0.39
Iodine AND/OR iodine compound	C0303013	Inorganic Chemical Pharmacologic Substance	60	338	478	278	1.38 × 10−14	0.30	0.40
Benzoic acid or derivative	C0578497	Organic Chemical	58	328	488	280	4.41 × 10−14	0.30	0.41

Abbreviations: CUI – Concept Unique Identifier; OR – Odds Ratio; RR – Relative Risk.

Discussion

As with any retrospective observation the primary limitation is a lack of a prospective control group, which means the results can be biased due to an imbalanced design. It is also worth noting, that correlation does not imply causation. For example, while cytopathology was found to be associated with an increased risk of death, it is more likely due to selection bias. Patients with more advanced disease more frequently underwent the procedure as part of their diagnostic process. There are also limitation due to data incompleteness. For example, here we looked at known deaths from the EHR only and did not include data from outside sources such the National Death Index. On the other hand, retrospective designs have the advantage of observing real clinical practice.

We have observed that the use of contrast media and medical gases used to induce anesthesia lowered the risk of death in the study population. This confirms an already known association between hospital resource utilization and patient mortality15,16.

Cimetidine, an H2 receptor antagonist, has a known off-label use as an anticancer drug17–19. Paradoxically, its use was associated with an increased risk of death in our study population. However, this subpopulation was also older than the rest of the cohort, and likely increased mortality was due to a more advanced disease process. Without access to clinical notes, we can only speculate that perhaps this was a part of an experimental treatment.

We see the potential to use this approach to automatically generate groupers or value sets of closely related concepts. This could be used either in the EHR to alert the physician to other possibly relevant features of patient presentation as well as on the research side to make more informative patient cohort selections.

A major critique would be that they we only looked at association of the concept and not the value. Presence of an observation on a patient-level also discards the temporal and frequency information. On the other hand, this would also increase dimensionality of the analysis (cf. curse of dimensionality) and would require not only a more sophisticated statistical approach but could also suffer from lower statistical power. These are some of the challenges that we hope to address in future work.

Conclusions

Information contained in EHR combined with knowledge engineering could be used a a viable method of rapid hypothesis generation, but requires comprehensive validation.