Automated clinical trial matching: take the first leap toward digital trials


While the pandemic has necessitated an acceleration for non-Covid clinical trials (reduced by almost 57% during 2020-21), it has also meant that Sponsors/ Institutions must find an innovative, yet compliant, way of running them. Moreover, the number of complex trials is set to increase, so the time taken to identify the rightful patient cohort is bound to increase. Already, states that it may take almost 2 hours to manually screen a patient for cancer trials. No wonder then that Global data reports that with digital trials set to increase by 28% during 2022, it will require digital ways to screen and engage with the patients. Tools like Circlebase’s Automated Clinical Trial Matching (ACTM) can digitize the recruitment process by analyzing complex inclusion and exclusion criteria and creating a patient cohort based on all the relevant information in the medical record.

How automation works

The automated matching process essentially involves 3 components:

1. Annotation of the “Eligibility criteria” stated in the clinical trials
2. Annotation of the medical record of the available patient cohort
3. Match or closeness of fit between the annotations

To begin with, the criteria are automatically fetched from or are input through a user interface. The criteria are then annotated for the clinical terms (entities) and the relationships they exhibit. Through a similar approach, the medical record of the patient is also annotated, and standard clinical entities are extracted per the Observational Medical Outcomes Partnership (OMOP) Common Data Model. Next, entity-to-entity matching algorithms help establish the match score of a patient’s medical record versus the trial’s eligibility criteria.
If we look closely, the success of an automated clinical trial-matching project depends on 3 metrics:
1. How efficiently annotated are the data used for training the deep learning model?
2. How accurately can the deep learning models recall the pattern for extracting entities and their underlying relationships?
3. How effectively can the entities fuzzy match
At Circlebase, our state-of-the-art algorithms help us get the best result for these metrics. In fact, for curating the training data, Circlebase uses its proprietary annotation tool. It helps us identify eight different entity types, including “Condition”, “Measurement”, “Drug”, “Observation”, “Procedure” and “Person.” These entities can have multiple attributes such as “Value”, “Temporal”, “Qualifier” and ”Negation” associated with them. Additionally, the entities could be related by “has value”, “has temporal”, “OR”, “AND” etc.

The Named Entity Extraction & Relationship Model

Named entities are chunks of text with some relevance. Named-entity recognition (NER) is a subtask of information extraction that seeks to locate and classify named entities mentionedin unstructured text into pre-defined categories (source: Wikipedia) While matching patients to the trials, one needs to extract the clinicalentities and their relationships. For example:
Deep learning models are trained to extract these entities and their relationships. Most of the traditional models start with identifying the entities first and then building “triplet models” for predicting the relationships between the entities. These triplet models are based on “subject”, “predicate” and “argument” (here, entity relation). But these are prone to error propagation, which might get induced during the extraction of the entities itself, causing recognition to fail if the entities are overlapping. Circlebase uses a blended approach, including a “joint model” for extracting the entities and their relationship along with specific interjection points for consuming the outcome of “individual models”. Such an approach has enabled us to detect a larger number of clinical entities with accurate attributes and relations, thus resulting in a better match and hence, a more reliable patient cohort for the trials.

Entity Matching Algorithms

Entities need to be standardized for them to be compared. Hence, they are linked to clinical terminologies like SNOMED CT, ICD 10, RxNorm, and LOINC and then compared. Still, there could be several clinical entities that might fail to match, even if clinically equivalent. An example could be “ductal carcinoma (Concept ID: C0007124)” versus “malignant neoplasm of breast (Concept ID: C0006826)” which may not match directly. This calls for an innovative approach to establishing the match. At Circlebase, our patent-ready algorithm, besides learning from the patterns, utilizes a smarter way to match the entities. This reduces our “miss-rate” and generates high-quality, reliable patient cohorts. Additionally, the algorithm also enables Study teams to see the gaps in their information collection process, thereby helping to streamline the process.

Getting ready for the automation

As with any AI project, ACTM needs access to anonymized data of the patients from which a study cohort must be built. Additionally, since the algorithms need to read through the medical records, an organization should assess the volume of medical records available in different formats like CCDA, FHIR, or PDF. Data sharing protocols should be agreed upon with the vendor, after the review for HIPAA compliance. Because PHI data is handled during the process, it is imperative that the solution be secured through multi-factor authentication, has encrypted transmission, and not store PHI information anywhere. Additionally, a full audit by the IT team may be useful.
It is also imperative for the organizations to determine the distinct financial and operational advantages of ACTM before adopting it. Indicators like revenue/cost per patient, time to enrollment, mismatch rate (when matched manually), etc. are good internal metrics to investigate. Additionally, the organizations should internally agree upon the distinct use cases of ACTM. As an example, it may very well serve beyond “patient matching.” In targeted recruitment calls, ACTM could help determine a patient’s eligibility based only on specific criteria. Similarly, beyond identifying the right patients from the patient registry, it may also help with study recommendations to the patients and thereby increase patient satisfaction rates.

We believe that an ACTM can eventually lead to improved clinical research and patient care and is a logical first step in streamlining digital trials. Circlebase, through its proprietary models and innovative algorithms, provides its clients with a thoroughly tested tool to embark on this automation journey. How have you been screening patients for trials? And how has automation helped you recently? We will be keen to hear from you and share our experiences. Please do write to us at: