A knowledge-based, automated method for phenotyping in the EHR using only clinical pathology reports

Abstract

The secondary use of electronic health records (EHR) represents unprecedented opportunities for biomedical discovery. Central to this goal is, EHR-phenotyping, also known as cohort identification, which remains a significant challenge. Complex phenotypes often require multivariate and multi-scale analyses, ultimately leading to manually created phenotype definitions. We present Ontology-driven Reports-based Phenotyping from Unique Signatures (ORPheUS), an automated approach to EHR-phenotyping. To do this we identify unique signatures of abnormal clinical pathology reports that correspond to pre-defined medical terms from biomedical ontologies. By using only the clinical pathology, or “lab”, reports we are able to mitigate clinical biases enabling researchers to explore other dimensions of the EHR. We used ORPheUS to generate signatures for 858 diseases and validated against reference cohorts for Type 2 Diabetes Mellitus (T2DM) and Atrial Fibrillation (AF). Our results suggest that our approach, using solely clinical pathology reports, is an effective as a primary screening tool for automated clinical phenotyping.

Introduction & Background

Electronic health records (EHR) capture an increasing variety and amount of clinical data leading to initiatives that are leveraging this potential for knowledge discovery. From adverse event and medical error detection for patient safety1,2 to case-control studies3, those new tools often rely on the researchers’ ability to isolate accurate cohorts of patients with a given phenotype. In this context, the term phenotyping has been used to describe automated and manual methods for identifying these patient cohorts in the EHR4. Advancement of automated phenotyping algorithms is a major roadblock in the field4. Several nationwide efforts, such as eMERGE5 and SHARPn6, have developed selection algorithms for high-throughput phenotype extractions. Those algorithms often comprise of a series of arithmetic and logical operations that are applied to the clinical data. The data types used in these algorithms are heterogeneous and may vary between institutions necessitating continual re-evaluation7. There is an opportunity in phenotyping to apply statistical learning methods, like Association Rule Mining (ARM), for modeling selection algorithms8 or the use of tensor factorization of medications and diagnoses to identify patients9. Other approaches have focused on certain types of clinical data like the diagnoses codes, which often are ICD-9-CM codes. Machine learning techniques trained on these data have been able to classify patients even when data are missing by using inductive logical programming10. The exclusive use of a particular clinical data type (e.g., medications or clinical pathology reports) is advantageous because it allows the exploration other the other data types in the selected cohort while minimizing bias to the extent possible. In particular, ICD-9-CM codes have been widely used for phenotyping and, in some cases, enhanced by additional information, such patient-reported data11. However, ICD-9-CM are primarily used for billing purposes and not for differential diagnosis, introducing complicated biases12. Clinical pathology is the medical subfield that deals with the analysis of bodily fluids for diagnosis and prognosis and clinical pathology reports, commonly called “lab reports,” may be more reliable than ICD-9 codes for EHR phenotyping, while maintaining the same level of standardization.

We present Ontology-driven Reports-based Phenotyping with Unique Signatures (ORPheUS), a knowledge-based phenotyping method that generates a unique clinical pathology signature for each term of a given ontology (i.e. each disease phenotype). Each “phenotype signature” is comprised of a set of abnormal laboratory tests (ATs). Our approach relies on only one type of clinical data – the clinical pathology reports – to minimize biases and increase interoperability. In total we generated clinical pathology signatures for 858 distinct diseases. We validated three of these signatures against reference patient cohorts using definitions from PheKB.org. We evaluated for precision and recall as well as the recovery of known co-morbidities. In each case we found that ORPheUS significantly outperforms the null model, with the T2DM signature recovering 17.2% of diabetics at 81.4% precision (F1 score=0.28).

Methods

Clinical Data Sources

The New York Presbyterian/Columbia University Medical Center (NYP/CUMC) clinical data warehouse contains about 470 million laboratory values from clinical pathology reports from more than 1.3 million patients over the last decade. We selected 177 of the most commonly ordered tests performed from blood, urine, plasma, and cerebrospinal fluid. We restricted our cohort of study to patients over 18 years old at order time with specified sex and at least one of these 177 laboratory tests. It narrowed our study to 767,389 patients with 172,518,869 values total. We preprocessed these data to assert if those reports were normal, abnormal, high, or low accounting for the patients’ age and sex, and according to our normal ranges database (Yahi, et al, in preparation).

Annotating abnormal laboratory tests with ontology terms

ORPheUS uses abnormal laboratory tests (ATs). We associated each AT to the medical terms from a given ontology through statistical enrichment analysis. We created the initial set of annotations by defining a search term by concatenating the name of the laboratory test with its non-normal status (i.e., “blood glucose low”, “blood glucose high”, etc.). Then we searched for each of these terms in the medical search engine UpToDate (www.uptodate.com) and gathered the titles of the first three pages of results. Once regrouped in a text file, these titles were annotated with the Annotator API by the NCBO (www.bioontology.org)13 and counted the number of times an ontology term would appear. We attributed 10 points for an exact match and 8 points for a synonym match. This is a one-time process associate ATs to clinical ontology terms and it is not repeated for the following steps of the phenotyping. We looped through all the terms of the ontology to associate each medical term with the ATs associated with its semantic descendants. We performed a Fisher’s exact test and a permutation analysis on these annotations sets to identify the ATs significantly associated to each ontology term, assessing significance using a FDR <= 0.05. Therefore, each ontology term (e.g., “Diabetes mellitus”), we have a set of significant ATs. We call this set of ATs the phenotype signature.

Selecting cohorts of patients for reference standard

We applied phenotype selection algorithms available on PheKB (www.phekb.org) to construct a reference standard. We therefore identified case cohorts for Atrial Fibrillation (AF)14 and Type 2 Diabetes Mellitus (T2DM)15,16. The data required by these algorithms consists of ICD-9 codes, CPT-4 codes, drug prescriptions, and clinical notes. We tested the performance of ORPheUS on these reference groups of patients.

Phenotyping with ORPheUS

We identified the presence of the phenotype signatures, complete (i.e., all the ATs of the signature are found in the patient’s clinical history) or partial (i.e. a subset of the ATs in the signature), in a patient’s clinical pathology records. For each patient, we look for the presence of any of the ATs belonging to the signature in his medical record to consider this patient as a potential candidate. We referenced laboratory tests with a universal code system named Logical Observation Identifiers Names and Codes (LOINC)17 and we used these codes to match ATs. We sorted those candidates by the number of distinct ATs of the target signature they had without any constraint in time. We designated by true positive (TP) the patients at the intersection of each of these prediction sets and its reference cohort of patients. To assess statistical significance, we compared the precision of the predictions from the signatures to a randomly selected cohort of the same size. For each group of candidates with N distinct ATs, we compared the precision of the prediction against the precision of a randomly selected cohort of the same size relative to all the patients with at least N distinct clinical pathology reports. We performed this random selection 20 times for each category. To compute the recall, we proceeded the same way except that the predictions were evaluated against the complete cohort of reference patients.

Results

Signatures

We annotated 351 abnormal laboratory test (ATs) with terms from the Human Disease Ontology (DOID)18. We then identified those ATs that were specific to each term to generate 858 signatures. The average signature contained 10.8 ± 14 ATs. The minimum number of ATs in a signature was 1 (for 95 signatures), and the maximum 50 (DOID:1579 Respiratory system disease). We did not construct a signature for parent term, “Disease,” in the ontology. Diabetes Mellitus with 14 distinct ATs is a little above the average of signatures (Table 1 – Signature for Diabetes Mellitus). Congenital heart disease presents 16 ATs and Mycoardial infarction 14 (Table 2 and 3).

Table 1.

– Signature of Diabetes Mellitus (DOID:9351)

Diabetes Mellitus (DOID:9351)
Clinical Pathology Report	Status
Glucose in Serum or Plasma	High/Low
Fasting glucose in Serum or Plasma	High/Low
Glucose in Blood	High/Low
Glucose in Serum or Plasma post challenge	High/Low
Hemoglobin A1c/Hemoglobin.total in Blood by HPLC	High/Low
Glucose in Blood (Meter)	High/Low
Hemoglobin A1c/Hemoglobin.total in Blood	Low
Hemoglobin in Blood	High

Table 2.

– Signature Congenital heart disease (DOID:1682)

congenital heart disease (DOID:1682)
Clinical Pathology Report	Status
Carbon dioxide, total in Arterial blood	High/Low
Carbon dioxide, total in Serum or Plasma	High
Estradiol (E2) in Serum or Plasma	High
Thyroxine (T4) free in Serum or Plasma	High
Calcium.ionized in Arterial blood	High
Erythrocyte mean corpuscular volume by Automated count	Low
Oxygen saturation in Arterial blood	High/Low
Oxygen saturation Calculated from oxygen partial pressure in Blood	High
Oxygen saturation in Venous blood	High/Low
Oxygen [Partial pressure] in Arterial blood	High/Low
Oxygen [Partial pressure] in Venous blood	Low
Thyroxine (T4) in Serum or Plasma	High

Table 3.

– Signature of myocardial infarction (DOID:5844)

myocardial infarction (DOID:5844)
Clinical Pathology Report	Status	Clinical Pathology Report	Status
Basophils [#/volume] in Blood	High	Platelet mean volume in Blood	High
Eosinophil [#/volume] in Blood	High	INR in Platelet poor plasma by Coagulation assay	High
Eosinophils [#/volume] in Blood by Manual count	High	Carbon dioxide [Partial pressure] in Arterial blood	High
Fibrinogen in Platelet poor plasma by Coagulation assay	High	Platelets in Blood	High
Hematocrit of Blood by Automated count	High	Potassium in Arterial blood	High
Hematocrit of Blood	Low	Sirolimus in Blood	High
International Normalized Ratio POC	High	Thrombin time in Platelet poor plasma by Coagulation assay	High

Phenotyping performances

We computed the precision and recall curves for the Diabetes Mellitus in 83,246 patients with T2DM as determined by the reference standard. We observed that of the 14 T2DM specific ATs in the signature, we only found up to 10 simultaneously in a single patient’s record. The precision is significantly better than by chance and increases above 80% with when at least 4 ATs are matched. At 6 or more distinct ATs the recall falls to below 5% (Figure 1).

Figure 1.

(left) Precision and Recall curves for Diabetes Mellitus signatures tested on T2DM patients

We also explored the cohorts of 80,163 patients with Atrial Fibrillation and evaluated the signatures of two of AF’s known comorbidities: myocardial infarction19, and congenital heart disease20. We observed an interesting precision for Congenital Heart Disease (Figure 2.a.) reaching a plateau around 80% from 10 distinct ATs. Myocardial Infarction (Figure 2.b.) presented a better precision, needing only 6 distinct ATs to reach 80%. However, despite a better initial recall, we witnessed a faster drop in sensitivity for the myocardial infarction signature than the congenital heart disease one. Finally, we observed that for 10 distinct ATs the predicted set of patients was so small that the precision fell to zero.

Figure 2.

(a) Congenital heart disease and (b) Myocardial infarction signatures in Atrial Fibrillation patients

Discussions

In this paper we present a novel automated EHR phenotyping algorithm by defining signatures of abnormal laboratory tests and scanning for matches in a patient’s longitudinal medical record. These signatures are knowledge-driven and rely on only one type of clinical data helping to minimize biases and improve interoperability. Since the signatures are knowledge-based they are not directly exposed to any clinical data before they are used for phenotyping. In total we generated 858 disease signatures. We validated two (atrial fibrillation and type 2 diabetes mellitus) of these signatures against a reference cohort of patients identified using eMERGE algorithms available at PheKB.org. We did not revalidate the PheKB algorithms in the CUMC database, however, previous implementations showed a 98% Positive Predictive Value for AF, and between 98 and 100% for T2DM.

In future studies, we would like to consider co-occurrences of those signatures across time. We might consider restricting the time windows from 1 to 12 months in patients’ records and look for the phenotype signatures, keeping only the maximum number of distinct simultaneous ATs in these windows. It might improve the precision of our predictions since some patients present sparse clinical pathology reports. Dynamical phenotyping using those reports has shown promising opportunities21. We would also like to investigate the potential of combining different phenotypes signatures. We also envision a possible approach for robustness assessment, which would consist of mapping ontological terms, in this example a DOID term, to ICD-9-CM diagnoses codes. This would allow us to evaluate performance of our all or most of our generated phenotype signatures systematically.

The EHR systems are in constant evolution, and many efforts are focused on designed new models learning from data and mitigate complex, inaccurate and frequently missing clinical values4. Indeed, the need for normalization in the information models that are use and the use of standardized vocabularies would ensure a better end-to-end connectivity over platforms allowing more reliable high-throughput phenotyping6. Meanwhile, as clinical notes still remain a critical source of information for phenotypic characteristics, phenotyping techniques using natural language processing (NLP) has been widely used and are gaining popularity22. The term of “Verotype” as a matching of genotype, phenotype and disease subtype has also been described23 to make a step forward to personalized medicine. The systematic inclusion of genotype and phenotype data in future EHR would be critical for this purpose24.

Conclusion

We presented Ontology-driven Reports-based Phenotyping with Unique Signatures (ORPheUS), a knowledge-based automated method for EHR-phenotyping, using only clinical pathology reports. We evaluated the performances of our phenotype signatures for T2DM and AF and demonstrated the potential use of this method for phenotyping. Our ontology-driven approach could allow us in future work to use other medical semantic fields and study for example adverse events signatures.