Building an OMOP common data model-compliant annotated corpus for COVID-19 clinical trials.

Clinical trials are essential for generating reliable medical evidence, but often suffer from expensive and delayed patient recruitment because the unstructured eligibility criteria description prevents automatic query generation for eligibility screening. In response to the COVID-19 pandemic, many trials have been created but their information is not computable. We included 700 COVID-19 trials available at the point of study and developed a semi-automatic approach to generate an annotated corpus for COVID-19 clinical trial eligibility criteria called COVIC. A hierarchical annotation schema based on the OMOP Common Data Model was developed to accommodate four levels of annotation granularity: i.e., study cohort, eligibility criteria, named entity and standard concept. In COVIC, 39 trials with more than one study cohorts were identified and labelled with an identifier for each cohort. 1,943 criteria for non-clinical characteristics such as "informed consent", "exclusivity of participation" were annotated. 9767 criteria were represented by 18,161 entities in 8 domains, 7,743 attributes of 7 attribute types and 16,443 relationships of 11 relationship types. 17,171 entities were mapped to standard medical concepts and 1,009 attributes were normalized into computable representations. COVIC can serve as a corpus indexed by semantic tags for COVID-19 trial search and analytics, and a benchmark for machine learning based criteria extraction.

Background

Since the first reported cases in December 2019, Coronavirus Disease 2019 (COVID-19) has spread rapidly from country to country and become a global pandemic [1]. As of February 2021, over 111 million confirmed positive cases have been reported worldwide, leading to 2.46 million deaths [2]. To deal with one of the worst pandemics in the world’s history [3], numerous research studies assessing the efficacy and safety of COVID-19 treatments are being conducted at an unprecedented rate. As of September 31, 2020, over 3,500 clinical trials targeting COVID-19 were registered in ClinicalTrials.gov, the largest clinical trial registry in the world. However, the free-text eligibility criteria are not amenable for computational analyses [4] or direct cohort queries using from electronic health records (EHRs) [5].

Formal representations for eligibility criteria have been pursued by the biomedical informatics research community to facilitate cohort identification, trial search [6] and clinical analytics [7], [8]. Several groups [4], [9], [10] have published datasets of annotated eligibility criteria, the largest of which containing criteria of 1,000 randomly selected clinical trials including 15 entity types and 12 relationships [11]. Manual annotation of clinical trials contributes to high-quality datasets but entails considerable costs of time and human labor. Automatic information extraction methods that leverage natural language processing (NLP) technologies may reduce the required curation time and effort but the performance is limited due the semantic complexity of eligibility criteria.

To address the challenge of lacking annotated COVID-19 trials, here we present a semi-automatic approach to annotating COVID-19 trial eligibility criteria by enhancing NLP-assisted manual concept recognition with machine-based concept normalization. The named entities in eligibility criteria are first annotated and classified by an automated Named Entity Recognition (NER) module, and then reviewed and updated by domain experts. The verified entities are converted to standard concepts by a concept mapping tool called Usagi [12]. Standard concepts (“concept” refers to “standard concept” in this paper) are medical terminologies that can be used as normative expressions of a clinical entity within standardized analytics. Translating source specific codes into standard concepts will promote the usage of eligibility criteria in a “common language” and leverage the power of rich clinical corpus and EHR database. Incorrectly mapped concepts are updated by experts in an iterative fashion. Associated relationships are also manually annotated. By integrating manual and NLP efforts, this approach could offer a cost-effective solution to generate trustworthy structured eligibility criteria. In this paper, we describe this semi-automatic eligibility criteria annotation framework and present our generated dataset COVIC with 700 clinical trials acquired from the ClinicalTrials.gov by querying all the trials indexed with “COVID-19” and including 11,710 annotated criteria formatted with the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) [13]. Compared to other data models such as Integrating Biology and the Bedside (i2b2) data model [16], Sentinel Common Data Model (SCDM) [17] or Patient Centered Outcomes Research Network (PCORnet) [18], OMOP CDM has an international orientation and participation with terminology sets from multiple countries [19], which is widely used in the medical research community for health data standardization.

To the best of our knowledge, COVIC is the first structured eligibility criteria dataset for

COVID-19 clinical trials. COVIC provides standardized human- and computer-interpretable annotation and representation of COVID-19 trials, which can greatly facilitate patient recruitment and exhibits great analytical flexibility. The proposed scheme defines a hierarchical trial annotation model with four layers that accommodate different levels of annotation granularity, such as study cohort groups, named entities or standard concepts, providing a more comprehensive annotation specification. To address the imperative for sharing machine readable datasets to battle the current pandemic, we released version 1 of the dataset with 700 trials in September of 2020 at (https://github.com/WengLab-InformaticsResearch/COVID19-Structured_Trials). We also provide related source code and documentation along with the dataset, which allows for maximum understanding of the data and promotion of its use among non-experts. This dataset will be periodically updated to ensure accuracy of the data being shared.

Methods

Clinical trials are research studies performed on human subjects often to evaluate a new medical treatment, drug or device. The most complete repository of clinical trials is ClinicalTrials.gov, a web resource created and maintained by the U.S. National Library of Medicine (NLM) [14], which includes a collection of clinical trial registrations in key-value pairs. Most values within these records are very short and limited to a few discrete value types which can be considered structured data, such as target condition, start and end dates, phase, recruitment status, etc. In this paper, they are referred to as metadata. Eligibility criteria, although playing a central role in clinical research for specifying rules for screening clinical trial participants, exist as free text. We proposed a semi-automatic annotation method to generate computable representations for eligibility criteria and present them together with metadata as a machine-readable dataset. Fig. 1 is an overview of the methodology framework.

Fig. 1

Overview of the structured dataset generation framework.

First, 700 trials were exported from the Aggregate Analysis of ClinicaTrials.gov (AACT) database by querying all the trials indexed with “COVID-19” as its target condition. AACT is a publicly available relational database which contains information about studies registered in ClinicalTrials.gov and is provided for research purposes by the Clinical Trials Transformation Initiative (CTTI) [15]. Next, relevant metadata and eligibility criteria were extracted separately from each trial. A hierarchical trial annotation schema following OMOP CDM was developed to guide the iterative annotation process, with the results reviewed and validated by domain experts. Next, six commonly used criteria across all COVID-19 trials about current age, high-risk status (e.g., hospital worker), COVID-19 status (e.g., yes, cleared, no), days since diagnosis (if relevant), current hospitalization or intensive care unit (ICU) admission, and pregnancy status are specified as “key criteria” and recorded separately from the annotation data table. Finally, all data are packaged as a COVID-19 structured trial dataset COVIC. We list a few application scenarios with COVIC, such as semantic retrieval of COVID-19 trial, eligibility query generation for patient cohort identification, trial similarity and collaborative analytics, and machine learning model training for information extraction from eligibility criteria.

The annotation schema

To accommodate annotation at different granularity levels, we built a hierarchical trial annotation model with four layers: study cohort layer, eligibility criteria layer, named entity layer and standard concept layer. Fig. 2 shows the hierarchy and consisted elements of the four layers.

Fig. 2

Hierarchical trial annotation model with four layers: study cohort layer, eligibility criteria layer, named entity layer and standard concept layer.

Study cohort layer

To make sure the structured eligibility criteria can be used effectively for patient recruitment, the number of sub-study or cohort groups utilized in each clinical trial was identified first during the annotation process. If a clinical study includes multiple sub-studies or cohort groups, it will be divided into a group of trials and each of them will be treated as a new trial for annotation with a suffix added to their previous ID. Table 1 lists an example of a trial with three cohort groups and the names of three ‘new’ trials generated from the old one.

Table 1

Example of a trial with three cohort groups.

#	New Trial ID	Cohort Group Topic
1	NCT04494893_exposed	Cohort 1. Exposed to coronavirus disease
2	NCT04494893_active	Cohort 2. Infected with coronavirus disease
3	NCT04494893_recovered	Cohort 3. Recovered from coronavirus disease

Eligibility criteria layer

In this layer, the “key criteria” and their types need annotation. First, criterion specifying any of the six “key criteria” discussed above will be identified and labelled. Next, “regular” criteria and “particular” ones are classified. Most of the criteria are “regular” ones annotated by terms, but there are four types of criteria labelled as a whole object because of their particular purposes, although they may contain entities or attributes. The four types are “non-query-able”, “informed consent”, “competing trial” and “post-eligibility”. “non-query-able” means a given criterion cannot be used to query relevant patients from a relational database. Such criteria usually specify implicit or flexible eligibility instead of determinate requirements. “informed consent” criteria typically specify the signing of informed consents for participation in a clinical trial which is outside of the scope of database queries. “competing trial” typically excludes simultaneous participation in other clinical trials and reinforces exclusivity of participation. “post-eligibility” typically involves requirements that are met after the patient enrolls on the clinical trial. Table 2 shows examples of these four types of criteria.

Table 2

Examples of four particular types of criteria.

Type	Example
Non-query-able	Any condition unsuitable for the study as determined by investigators;Any other reason that the Investigator considers makes the patient unsuitable to participate
Informed consent	Patient or responsible family member or surrogate signs informed consent
Competing trial	Not participate in any other clinical trial for an investigational therapy through day 30
Post-eligibility	Able to attend all scheduled visits and to comply with all study procedures;Agrees to required laboratory data collected which will include the baseline organ function and regular ongoing assessments done as part of routine care.

Named entity layer

For “regular” criteria, entities and their associated attributes and relationships will be extracted and annotated. We adapted a simplified version of the Chia annotation model created by Kury et al[11] by not using its “Scope” mechanism considering its complexity. Table 3 lists the categories and examples of eight domains of entities and seven categories of attributes in this layer.

Table 3

Entity and attribute with different categories and examples.

	Category	Example
Entity	Condition	Patients with bleeding disorders
	Observation	History of stem cell transplant
	Drug	Has received treatment with systemic anticancer treatments
	Measurement	SpO2 < 92%
	Procedure	requires supplemental oxygen 2 LPM
	Person	If female, subject must meet one of the following conditions
	Visit	Patients admitted to hospital
	Device	Has a Ventricular Assist Device
Attribute	Value	SpO2 < 92%
	Temporal	Has had plasmapheresis within the previous 24 h
	Qualifier	Significant cardiovascular disease
	Reference_point	within two weeks of a blood transfusion
	Mood	Patients who require renal replacement therapy
	Negation	no alternative explanation for current clinical condition
	Multiplier	at least two of the following high-risk criteria

Most entity types and attributes are defined by the literal meaning of their names in Chia model. One attribute which requires specific explanation is “Reference_point”. It comes downstream (usually directly) from a parent “Temporal”, and specifies a concept whose timestamp is pivoting that “Temporal”. For example, in “within two weeks of a blood transfusion” this entire text string is one “Temporal”, and it contains (overlaps) the “Reference_point” “blood transfusion.” Eleven types of relationships were used to specify the association between pairs of entities (AND, OR) or entity and its corresponded attribute: SUBSUMES, HAS_NEGATION, HAS_MULTIPLIER, HAS_QUALIFIER, HAS_VALUE, HAS_TEMPORAL, HAS_INDEX (target argument is reference_point), HAS_MOOD and HAS_CONTEXT (target argument is observation and not included in above relationships). More details can be found in the Supplementary File of the Chia Annotation Model [11].

Standard concept layer

In this layer, extracted named entities are converted to the standard concepts in the OMOP CDM. Each standard concept has a unique ID and belongs to one domain, which defines the location where the concept would be expected to occur within data tables of the OMOP CDM. Table 4 lists various examples of extracted entities mapped to standard concepts. In the first two lines of examples, the name of entity is the same with its mapped concept, meaning the entity itself is already a standard concept. In the third example, the names of entity and concept are similar but not exactly the same. In example 4, the entity name is an abbreviation of the standard concept. The entity and mapped concept have totally different names in example 5 and 6, but they are semantically equivalent. In COVIC, all types of entities are mapped to standard concepts.

Table 4

Examples of entity and standard concept.

#	Entity Name	Domain	Standard Concept	Concept ID
1	acute hepatic failure	Condition	acute hepatic failure	4026032
2	discharged from hospital	Observation	discharged from hospital	4084843
3	drug addiction	Condition	Dependent drug abuse	4275756
4	CPAP	Device	Continuous positive airway pressure (cpap) device	2616666
5	shortness of breath	Condition	Dyspnea	312437
6	solid tumor	Condition	Neoplasm	4030314

Semi-automatic annotation and normalization

To maximize annotation quality while minimizing manual annotation effort, manual and machine-assisted annotations are combined in a semi-automatic approach. First, trials with multiple cohorts are manually identified and labeled. Next, an information extraction tool is used to automatically locate and classify entities in eligibility criteria, and the recognized entities results are verified and corrected manually by medical domain experts to overcome the limitations in the tool. Next, Attributes and relationships are manually annotated by domain experts because existing criteria extraction tools could not achieve satisfactory performance due to the large number of attribute and relationship types defined in our annotation model. Finally, extracted entities are mapped to their corresponding concepts and normalized automatically by the concept mapping and normalization modules. Fig. 3 shows the semi-automatic annotation process.

Fig. 3

Semi-automatic eligibility criteria annotation and normalization.

An information extraction tool Critera2Query [5] is used in this study to automatically identify entities in eligibility criteria. Criteria2Query is a natural language interface that transforms eligibility criteria into executable query for cohort definition, and can be used for entity recognition. Compared to other NER tools like cTAKES [23] or MetaMap [24], Criteria2Query follows OMOP CDM with considerable precision (~90%) and recall (~71%) rates. Eligibility Criteria with identified entities are imported to an annotation tool Brat for verification, and to be annotated for attributes and relationships. Brat is a web-based, interactive annotation environment with a visualization frontend interface [20]. Entity or attribute annotation can be defined using a contiguous span, beginning at the start of a phrase and ending at the completion of the phrase to capture instances rather than individual word tokens. An entity can be linked by multiple entities or attributes to build various types of relationships. The annotation makes a directed acyclic graph, which can be easily transformed into Boolean logic to form a database query. Fig. 4 shows an example of annotated criterion with Brat tool.

Fig. 4

Example of the eligibility criteria annotations on Brat tool.

The annotated eligibility criteria include extracted name entities and their domain tags, which can feed into the concept mapping module to obtain each entity’s mapped concept name and ID. Usagi is used in this study to assist with concept mapping. It is a software tool created by the OHDSI team and used in the process of mapping codes from a source system into the standard terminologies stored in the OMOP vocabulary [12]. With the free-text string as input, medical concept with the highest ‘mapping accuracy score’ from the returned results will be selected as the mapped concept. If a mapped concept suggested by Usagi is not accepted by the expert after reviewing it, the expert will suggest a new mapping, and a “entity-concept” dictionary will be automatically updated. The dictionary is built and maintained to complement Usagi mapping in cases where there is no appropriately mapped concept. A segment of the dictionary is shown in Table 5 .

Table 5

A short segment of the dictionary.

Entity Source Text	Domain	Concept ID	Concept Name
GI perforation	Condition	4202064	Gastrointestinal perforation
encephalitis	Condition	378143	Encephalitis
oxygen saturation (pulse oximetry)	Measurement	4310328	Blood oxygen saturation
individual NEWS parameters	Measurement	44808684	National early warning score
hematopoietic stem cell transplant	Procedure	4120445	Hemopoietic stem cell transplant
transthoracic echocardiogram	Procedure	4335825	Transthoracic echocardiography

The dictionary is a lookup table that saves all the manually corrected mappings and will be used for concept mapping before applying Usagi. It ensures the mappings updated by domain experts will be applied first in order to preserve the quality of annotations. The update of the dictionary follows an iteratively incremental learning process and all inappropriate mappings in previous annotations are avoided in mapping annotations of new trials.

The “value” and “multiplier” attributes are normalized automatically as numerical values with upper and lower boundaries. For example, the value attribute ‘>20%’ in the phrase ‘>20% of the body surface’ will be coded into a minimum boundary as ‘20%’ and maximum boundary as ‘infinity’ for ‘body surface’. Supplementary Table 1 lists the mathematical operators in word (string type) expression and its corresponded numerical symbol. Regular expressions are used to identify and covert different type of operators. For the normalization of “temporal” attribute, all temporal expressions are unified to the same unit (days) by SUTime from Standard NLP group[21]. For example, “for at least 1 year before the screening visit” will be coded into 365 days before the ‘the screening visit’. After normalizations, these attributes are converted from strings to numerical data types and are comparable in a quantitative manner. After labelling the five key criteria for COVID-19 trials, all annotated data are integrated together and reviewed and updated by domain experts. The structured dataset will be generated after that.

Annotation effort assessment

As discussed above, manual review was included in each step of the annotation schema to overcome the limitations of automated tools and ensure the corpus quality, we empirically assessed the annotation effort. Given the nature of the pandemic and the focus on rapid dissemination of annotated data at the outset of this project, specific tracking of the time needed to annotate each trial was not performed. However, in speaking with the annotators, early annotation work required around 45 min to one hour per trial which was reduced to around 15–20 min per trial at the end of the research effort as the annotators were more comfortable with the brat software and the iteratively generated mapping dictionary improved first-pass concept mapping results. The greatest amount of manual time for annotation review occurred at the eligibility criteria layer in adjusting entity recognition. Such an emphasis on manual review was due to our desire to ensure completely accurate data in our annotated corpus. To be specific, in the study cohort layer, the annotation is straightforward because most eligibility criteria contain explicit indicators such as “study group x” or “study cohort x”; in the edibility criteria layer and named entity layer, the annotation time depends on the length of criteria and number of included instances, and it usually takes 12–15 min; in the standard concept layer, new appeared concepts will be automatically added to the “entity-concept” dictionary and used for future annotations, so the annotation cost of this layer is supposed to continue to decrease as more and more concepts are included in the dictionary. It takes about 3–5 min for each trial in the current stage.

Validation

The creation of COVIC was performed by medical professionals (AB and LS). An annotation manual was developed to guide the data processing and format conversion. A medical terminology search engine Athena (http://athena.ohdsi.org) that provides searching of concepts in the OMOP CDM is used to assist experts in annotations and reviews in case of doubt for some concepts. For the first 100 trials, both annotators regularly discussed adaptations to the annotation model based on their experience and reviewed each other’s work to make sure their annotations were consistent. Next, each annotator received a separate set of criteria loaded in brat and hand-created the entities and relationships as expressed above. All annotated trials will be reviewed by a third annotator (YS) to check whether there are missing annotations.

When the annotations for all 700 trials are finished, we randomly selected 70 trials (10%) from them to evaluate the inter-annotator agreement. Each text-bound annotation in Brat is an annotation instance. Two identical instances are considered as instance-level agreement. If two annotators agree on the annotation of an entity, they agree on both the entity span and the category. For example, the italic phrase in the criteria “subjects with a history of systemic autoimmune diseases” was annotated as “Condition” type entity. An agreement is counted if and only if annotators consistently recognized the sequence “systemic autoimmune diseases” and classified it as a “Condition” type entity. Each document can be split up into multiple tokens. If annotations on the same token are identical, annotators achieve token-level agreement. For example, entity “tamoxifen citrate” annotated as “Drug” type is an instance. It contains two tokens: “tamoxifen” and “citrate”, and they are both “Drug” type entities. Given a clinical trial annotated separately by two annotators, the “instance-level” agreement is calculated by 2|A ∩ B|/(|A|+|B|) [25], where A and B are sets of instances created by the two annotators. The “token-level” agreement is also be calculated by 2|A ∩ B|/(|A|+|B|) while A and B are sets of tokens. For entities and attributes, both instance level and token level agreements are automatically measured by a Python tool “Bratiaa”[22]. For relationships, instance level agreements are manually calculated.

Results

In this section, we discuss the detailed description of the COVIC dataset. COVIC contains 44, 622 annotations for 11,710 inclusion and exclusion eligibility criteria from 700 COVID-19 trials conducted in the United States. 39 trials have multiple study cohorts and they were divided into separate trials each of which includes eligibility criteria for one study cohort. The descriptive statistics are computed from the trial metadata and listed in Table 6 .

Table 6

Statistical information of 700 trials: 6.A Study Type, 6.B Study Allocation, 6.C Study Phase, 6.D Trial Locations.

Table 6.A
Study Type	Count	%
Interventional	503	71.86%
Observational	153	21.86%
Observational [Patient Registry]	31	4.43%
Expanded Access	13	1.86%

Most of the trials are interventional (71.86%) and randomly allocated (54.86%) studies. 27.71% of the studies focus on Phase 2. Places like New York, California and Massachusetts states have the most trial recruitment sites and trial seekers living around these states may find eligible trials more easily than others in US. We also notice that 40.00% of the trials have no information on the type of allocation and 43.14% of the trials do not specify any phase stage.

For all the trials in COVIC, 18,161 entities were extracted and annotated by various domain names. Table 7 lists the counts and percentage of annotated entities in eight domains. “Condition”, “observation”, “drug” and “measurement” have relatively higher annotations than the other four domains, which makes sense in terms of the purpose of clinical trials. Of the 7, 743 annotations for attributes, most of them fall into “value”, “temporal” and “qualifier” categories, as show in Table 8 . We have 16,443 annotations for 11 types of relationships among entities and their attributes, and the statistics is provided in Table 9 . Categories of “OR”, “has_value” and “has_temporal” were annotated more than other relationships, which is consistent with the statistical results of annotations in domains and attributes.

Table 7

Total count and percentage (%) of annotated entities in 8 domains.

Domain	Count	%
Condition	6,614	36.42%
Observation	4,243	23.36%
Drug	2,179	12.00%
Measurement	2,112	11.63%
Procedure	1,364	7.51%
Person	1,008	5.55%
Visit	470	2.59%
Device	171	0.94%
In total	18,161	100.00%

Table 8

Total count and percentage of annotated entities in 7 attributes.

Attribute	Count	%
Value	2,832	36.57%
Temporal	1,803	23.29%
Qualifier	1,613	20.83%
Reference_point	537	6.94%
Mood	520	6.72%
Negation	380	4.91%
Multiplier	58	0.75%
In total	7,743	100.00%

Table 9

Total count and percentage of annotated entities in 11 relationships.

Relationship	Count	%
OR	3,155	19.19%
has_value	2,910	17.70%
has_temporal	2,651	16.12%
has_qualifier	1,599	9.72%
AND	1,549	9.42%
has_context	1,529	9.30%
subsumes	1,382	8.40%
has_mood	572	3.48%
has_index	529	3.22%
has_negation	509	3.1%
has_multiplier	58	0.35%
In total	16,443	100.00%

Besides the above annotations based on terms, we also have 2,210 annotations for criteria from four categories. They usually have particular purposes different than most of the criteria as discussed in section 2.1. The counts and percentage of them are listed in Table 10 .

Table 10

Total count and percentage of the four particular types of criteria.

Type	Count	%
Non-query-able	1,112	50.32%
Informed consent	478	21.63%
Competing trial	331	13.08%
Post eligibility	289	13.08%
In total	2210	100.00%

In the criteria normalization, we have 17,171 entities were mapped to 4,140 different medical concepts, and the top 10 of the most common concepts mapped in each domain are listed in Table 11 . Each concept is listed by its ID and name retrieved from OMOP CDM. As the table shows, concepts like “Disease caused by severe acute respiratory syndrome coronavirus 2”, “Detection of 2019 novel coronavirus using polymerase chain reaction technique”, “Artificial respiration” and “Oxygen therapy” are mapped more frequently which corresponds with the targeted disease COVID-19 of our dataset.

Table 11

The count (#) of most common (top 10) concepts mapped in each domain.

Condition			Observation			Drug			Measurement
ID	concept	Freq.	ID	concept	Freq.	ID	concept	Freq.	ID	concept	Freq.
37311061	Disease caused by severe acute respiratory syndrome coronavirus 2	660	4188893	History of	523	21605200	Corticosteroids	78	37310255	Detection of 2019 novel coronavirus using polymerase chain reaction technique	192
4299535	Patient currently pregnant	372	40218805	CDC laboratory	148	21602457	Prolactine inhibitors	66	4146380	Alanine aminotransferase measurement	118
437663	Fever	133	4185135	Breastfeeding	124	1777087	Hydroxychloroquine	53	4263457	Aspartate aminotransferase measurement	114
312437	Dyspnea	118	36685445	on room air at rest	91	21603891	IMMUNOSUPPRESSANTS	52	4310328	Blood oxygen saturation	96
254761	Cough	104	37310260	Close exposure to 2019 novel coronavirus infection	52	21601386	ANTINEOPLASTIC AND IMMUNOMODULATING AGENTS	43	4233883	Ratio of arterial oxygen tension to inspired oxygen fraction	86
439727	Human immunodeficiency virus infection	83	4120014	Polymerase chain reaction	43	2718732	Immunosuppressive drug not otherwise classified	42	4267147	Platelet count	74
316866	Hypertensive disorder	68	45766517	Confirmatory technique	40	1507835	Vasopressin (USP)	41	4313591	Respiratory rate	72
255573	Chronic obstructive lung disease	66	4244251	Confirmed by	40	40171288	tocilizumab	36	44806420	Estimation of glomerular filtration rate	69
443392	Malignant neoplastic disease	61	4289014	Normal breast feeding	35	1792515	Chloroquine	30	3027315	Oxygen [Partial pressure] in Blood	66
4212484	Multiple organ failure	60	4142947	Symptomatic	33	1309944	Amiodarone	23	44789311	Pregnancy test	60
Procedure			Person			Visit			Device
ID	concept	Freq.	ID	concept	Freq.	ID	concept	Freq.	ID	concept	Freq.
4230167	Artificial respiration	151	4265453	age	567	38004515	Hospital	298	4139525	High flow oxygen nasal cannula	28
4239130	Oxygen therapy	123	442986	female	200	38004311	Inpatient Hospice	29	4138614	BiPAP oxygen nasal cannula	16
4052536	Extracorporeal membrane oxygenation	71	442985	Male	100	38004519	Home Health Agency	21	2614925	Cannula nasal	15
4032243	Dialysis procedure	67	4323831	old	34	32037	Intensive Care	19	2616666	Continuous positive airway pressure (cpap) device	14
4202832	Intubation	60	4046779	Adult	23	9201	Inpatient Visit	13	4145528	Nonrebreather oxygen mask	12
4273629	Chemotherapy	59	4119673	year	7	38004522	Department Store	10	4030875	Cardiac pacemaker	6
44790095	Invasive ventilation	51	1332764	children	9	38004284	Psychiatric Hospital	8	4232657	Vascular stent	5
4208341	Solid organ transplant	45	4305451	Infant	3	8717	Inpatient Hospital	8	4234106	Metal periosteal implant	4
37018292	Continuous renal replacement therapy	37	42073776	Newborn	5	8676	Nursing Facility	8	4148006	Epidural catheter	3
40486624	Noninvasive positive pressure ventilation	34	2090691	Mothers	3	9202	Outpatient Visit	8	45760696	Spinal catheter	3

The same 70 trials were provided to the two annotators (AB and LS) to annotate independently using the Brat annotation tool to evaluate the inter-rater agreement. In total, 331 out of 1,732 entities, 204 out of 726 attributes, and 288 out of 1,538 relationships were annotated with different tags, making the inter-rater agreement of the annotations 80.9% (1,401/1,732) for entities, 72% (522/726) for attributes, and 81.3% (1,250/1,538) for relationships. Table 12 and Table 13 list the instance-level and token-level agreement rates of entities and attributes in different type and their average. Inter-rater agreement for each trial is listed in Supplementary Table 2.

Table 12

Instance-level and token-level agreement rates of entities in different types and the average (arithmetic mean for all trials).

Type	Instance-level	Token-level
Measurement	0.874	0.887
Condition	0.832	0.845
Person	0.79	0.830
Drug	0.781	0.825
Visit	0.765	0.792
Procedure	0.728	0.746
Observation	0.601	0.718
Device	0.565	0.714
Average	0.809	0.825

Table 13

Instance-level and token-level agreement rates of attributes in different types and the average (arithmetic mean for all trials).

Type	Instance-level	Token-level
Temporal	0.713	0.752
Value	0.927	0.954
Negation	0.653	0.693
Qualifier	0.66	0.741
Multiplier	0.678	0.764
Reference_point	0.344	0.372
Mood	0.624	0.643
Average	0.72	0.798

A few terms with multiple interpretations bring the disagreements for the two annotators. For example, “COVID-19 PCR” can be treated as an entity from “measurement” domain, but it also makes sense for “COVID-19” as a ‘condition’ entity and “PCR” as a measurement entity. Further, when assessing agreement between specific annotation types, the lowest level of agreement is observed in “reference_point” annotation types. The reason is that a “reference_point” attribute always overlaps a long “Temporal” type attribute, and easily to be ignored during annotation. The disagreements in annotations will not have effect to the availability of COVIC.

Usage discussion

As the first large structured dataset for COVID-19 clinical trials, COVIC can support the research community and facilitate the development of various applications on automatic cohort identification, fine-grained trial generalizability analysis or similarity comparison. It can also serve as a shared benchmark for machine learning based information extraction from free-text clinical trial eligibility criteria. We describe below four examples of data usages for different applications.

COVID-19 cohort identification

Researchers worldwide have worked diligently to understand the mechanisms of transmission and action for COVID-19 and to discover effective treatments and interventions. One important data source for COVID-19 research is patients’ clinical data stored in electronic health records (EHRs). EHRs use structured data elements to document patient information with controlled vocabulary, but clinical trial eligibility criteria are usually written in free-text descriptions and the majority of which include semantically complex language hard for computational processing. It requires domain experts to understand the eligibility criteria and construct the query manually, which is time-consuming and error-prone.

With COVIC, the EHR data queries can be automatically generated with a few lines of extra codes. For example, in the criterion “Blood pressure<90 mm Hg systolic or 60 mm Hg diastolic recorded on at least two readings 30 min apart” (ID NCT04335123), “blood pressure systolic” and “blood pressure diastolic” are extracted and annotated as “measurement” entities and mapped to “systolic blood pressure (ID 4152194)” and ”diastolic blood pressure (ID 4154790) “ respectively. “<90 mm Hg”, “<60 mm Hg” and “at least two” are normalized as numerical values. “Readings” and “30 min apart” are annotated as “observation” entities and normalized as “reading (ID 3243724)” and “Every thirty minutes as required (ID 45757503)”. The original criterion can then be coded as “systolic blood pressure < 90 mm Hg” or “blood pressure diastolic < 60 mm Hg” with “ 2 reading” and “every thirty minutes as required”. A few applications have been developed to convert such structured statements to queries for cohort definitions, such as Circle-be (https://github.com/OHDSI/circe-be) and Atlas (http://www.ohdsi.org/web/atlas/).

COVID-19 trial search

While we are fortunate to have a growing number of trials available for coronavirus therapies, it is often difficult to determine which specific studies are appropriate for individual patients. Keyword-based search engines allow users to search for trials using a combination of terms, though this method is often relatively nonspecific and can result in information overload when too many trials are returned. With the help of COVIC, semantic trial search engines can be developed for patients to find trials for which they are eligible and appropriate. Once the data are structured and coded, algorithms can be applied to match a patient to clinical trials based on their answers to eligibility questions. Rather than obtaining a long list of results, patients can receive a personalized list of a few trials for which they may be eligible. COVIC will be an enormous help to busy clinicians, allowing them to efficiently link patients with the trials best suited to them. A web application “COVID-19 Trial Finder” (https://covidtrialx.dbmi.columbia.edu) has already been developed by our team to provide patient-centered search of COVID-19 trials with the support of the COVIC dataset. It allows potential candidates to pre-screen their eligibility through a set of short medical questions with minimum effort. Its initial version was released in May 2020, and>2,300 page visits from 30 countries were recorded by Feb. 15th, 2021, including 1,348 visits from 40 states in the United States, and the average session duration was 2 min and 24 s, according to the report of Google Analytics.

COVID-19 trial collaboration analytics

Researchers across the world are racing to conduct COVID-19-related clinical trials. However, many of these trials have significant redundancy as well as limited or biased target patient populations. Clinical trial collaboration analytics tries to answer the question: given two clinical trials investigating the same scientific question, can we tell if they are studying comparable cohorts? The manual labor required from domain experts for such appraisal is prohibitive. Metadata such as study type, location, age, or sex information are usually used for coarse-grained trial comparison. COVIC provides more features within eligibility criteria to identify opportunities for collaborative recruitment such as the category or mapped concepts of annotated entities. They can be used for building recruitment optimization systems to minimize the unnecessary local competition among trial investigators in COVID-19 clinical trials. We have developed a prototype application “Collaboration Opportunities” (http://apex.dbmi.columbia.edu/collaboration/) based on COVIC dataset to support coordinated meta-analyses.

Machine learning based information extraction for clinical trials

Non-interoperable data cannot be interpreted by computers and this inhibits machine learning based NLP. Large machine-readable datasets are needed to assist the development and evaluation of the biomedical NLP systems. Most of previous datasets only include a few trials which are not sufficient enough to train robust machine learning models. Chia has a large size of annotations for 1,000 trials but only for the domains of entities, attributes, and their relationships. COVIC includes not only the three types of annotations that Chia provides, but also the mapped concepts for each identified entity. This allows COVIC to serve as a benchmark for both training and evaluative purposes for NLP related tasks like Named-entity Recognition or Named-entity Linking. As a manually annotated and validated large corpus, COVIC can be a knowledge base to catalyze clinical trial analytics and criteria extraction. A web application “Criteria2Query” (http://www.ohdsi.org/web/criteria2query/) was developed by our group for automatic clinical trial eligibility criteria extraction and query generation based on the structured datasets.

To further improve the usage of the proposed schema and corpus, a couple of limitations might need more work. First, each criterion was annotated only one kind of label in this work, but sometimes a criterion with multiple interpretations might have various types of annotations. For example, “women who have been pregnant” can be considered as an “Observation” entity as a whole, but the token “women” and “pregnant” can also be labelled by “Person” and “Condition” types separately: “women [Person] who have been pregnant [Condition]”. “COVID-19 PCR” can be treated as an entity from “measurement” domain, but it also makes sense for “COVID-19” as a ‘condition’ entity and “PCR” as a measurement entity: “COVID-19 [Condition] PCR [Measurement]”. Eligibility criteria with different types of annotations can be further explored. Second, although the annotation process is semi-automatic, it still took lots of human efforts to check and update the annotations due to the limitations of NER and concept normalization tools. Also, only new appeared criteria need to be focused by annotators and repetitive annotations should be automatically filtered or labelled. Future directions aim to improve the performance and efficiency of the automated tools. Third, COVIC contains a large number of annotations for 700 COVID-19 trials, but considering the nearly 4,800 COVID-19 trials registered in ClinicalTrials.gov as of Feb. 2021, it is worthy to expanding the corpus with more annotated trials.

Conclusions

In the context of the COVID-19 health crisis, a multitude of clinical trials have been designed to seek effective ways to manage this pandemic. In this research, we propose a semi-automatic clinical trial annotation method and contribute a multi-purpose large annotated dataset COVIC with structured and semantically annotated criteria for 700 COVID-19 clinical trials. We provide a detailed description of the method and the dataset, present examples of how it can assist in patient recruitment and provide insights into trial collaboration opportunities. COVIC is currently being used in a number of active systems and exhibits great analytical flexibility. It promises to support more applications.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.