Phenotyping to Facilitate Accrual for a Cardiovascular Intervention.

BACKGROUND: The conventional approach for clinical studies is to identify a cohort of potentially eligible patients and then screen for enrollment. In an effort to reduce the cost and manual effort involved in the screening process, several studies have leveraged electronic health records (EHR) to refine cohorts to better match the eligibility criteria, which is referred to as phenotyping. We extend this approach to dynamically identify a cohort by repeating phenotyping in alternation with manual screening.
METHODS: Our approach consists of multiple screen cycles. At the start of each cycle, the phenotyping algorithm is used to identify eligible patients from the EHR, creating an ordered list such that patients that are most likely eligible are listed first. This list is then manually screened, and the results are analyzed to improve the phenotyping for the next cycle. We describe the preliminary results and challenges in the implementation of this approach for an intervention study on heart failure.
RESULTS: A total of 1,022 patients were screened, with 223 (23%) of patients being found eligible for enrollment into the intervention study. The iterative approach improved the phenotyping in each screening cycle. Without an iterative approach, the positive screening rate (PSR) was expected to dip below the 20% measured in the first cycle; however, the cyclical approach increased the PSR to 23%.
CONCLUSIONS: Our study demonstrates that dynamic phenotyping can facilitate recruitment for prospective clinical study. Future directions include improved informatics infrastructure and governance policies to enable real-time updates to research repositories, tooling for EHR annotation, and methodologies to reduce human annotation.

Introduction

The adoption of electronic health records (EHRs) in the USA has generated large datasets that enable observational research. An essential step in utilizing these datasets for research is to identify patients with certain characteristics that match a set of eligibility criteria, a process known as phenotyping or patient cohort identification. Phenotyping requires significant manual effort to create labeled/validation datasets, and a population database is required to execute the algorithm [1].

Phenotyping has been reported to be useful for retrospective studies [2-4]; however, it has been relatively under-utilized for prospective studies. Although there has been considerable thought on phenotyping strategies for pragmatic clinical trials (PCTs) [5-7], much of the phenotyping for clinical studies has focused on developing disease registries or disease cohorts that serve as an initial focused pool for studies related to a particular disease [1, 8, 9].

Disease registries often cannot be readily used for prospective studies, as each study has specific eligibility criteria, and additional algorithms are thus necessary to implement the eligibility criteria to refine the disease registry, which costs significant effort. Due to the high cost and effort involved, phenotyping has largely been used for pre-screening before starting recruitment, and no previous reports of actively phenotyping for intra-study recruitment have been published to date. To address this gap, we describe an approach for phenotyping in tandem with manual screening, referred to as dynamic phenotyping (DP). We implement this approach to facilitate recruitment for a prospective study on heart failure.

The study team had utilized a phenotyping algorithm to identify an initial cohort of heart failure patients prior to commencing recruitment. However, as the study progressed, the study team recognized the need to refine the disease-cohort to match the eligibility criteria, which evolved into an approach of dynamically performing phenotyping in tandem with manual screening. We report on the challenges encountered when using this approach, and we test the hypothesis that DP can significantly improve the screening rate.

Materials and Methods

This study was carried out at Partners Healthcare, Boston and was approved by the Institutional Review Board. For the case study, we performed phenotyping using data for 193,808 patients with a coded diagnosis of heart failure. This study was conducted in compliance with the ethical standards of the responsible institution on human subjects as well as with the Helsinki Declaration.

DP

Our approach consists of multiple screen cycles. At the start of each cycle, a phenotyping algorithm is used to identify eligible patients from the EHR, creating an ordered list such in which those patients most likely to be eligible are listed first. This list is manually screened, and the results are analyzed to improve phenotyping for the next cycle. We applied this approach to EHR data at Brigham and Women’s Hospital (BWH) to accrue heart failure patients with reduced ejection fraction (HFrEF) for an intervention study, as described below in the methods section.

Heart failure study

Although national guidelines exist for the pharmacological therapy of HFrEF, many eligible patients do not receive optimal guideline-directed medical therapy. To address this gap, we created a program to remotely optimize HFrEF patient care using supervised non-licensed navigators. DP was used to identify patients eligible for the study. Treating providers were approached for consent to adjust medical therapy in accordance with ACC/AHA guidelines [10, 11]. Once consent was obtained, patients were contacted by phone by a navigator who completes a medication reconciliation and provides education over the telephone [28-32].

We obtained data for 193,808 patients with a coded diagnosis of heart failure from the institutional research patient data repository [12-15]. The eligibility criteria for the study included adult patients that had a diagnosis of HFrEF, and whose cardiac care was managed primarily by a cardiologist at BWH. Patients were excluded for the following reasons: end-stage renal disease, active chemotherapy, low life expectancy, risk of frailty, transplant, intravenous inotrope use, prior need for mechanical ventricular support, unstable disease, and/or a history of medication non-adherence.

The primary challenge was to identify eligible patients from the pool of patients that had a coded diagnosis of heart failure. As the distributions for exclusion criteria were expected to be small, we focused on implementing the inclusion criteria through in-silico phenotyping.

We utilized a variety of approaches to develop algorithms for extracting variables (Table 1), including rules, regular expressions for text processing and logistic regression [16]. In addition, we measured the performance of phenotyping in each cycle by measuring the positive screen rate (PSR) on the annotations resulting from manual screening in the next cycle.

Table 1

List of Algorithms for Pre-Screening

Inclusion criteria	Model types	Data modalities	Cycle
			I	II	III	IV
Heart failure	Logistic regression	Coded data and notes	√
	Simple inference	EF algorithm		√	√	√
EF ≤ 40	Simple regular expression	All notes	√
	Elaborate text-processing	Notes from cardiology service		√	√	√
Primary cardiologist at institution	Rules	Notes	√
	Rules	Coded data and notes		√	√
	Logistic regression	Coded data and notes				√

EF: ejection fraction.

Results

We applied the DP approach to facilitate recruitment for a real-world study on in HFrEF. Table 2 presents the results of the first four cycles, while Figure 1 summarizes the results for each cycle.

Table 2

PSR in Each Screening Cycle

Cycle	I	II	III	IV	Total
Screened in (A)	165	24	14	30	233
Screened out (B)	670	26	34	59	789
Total screened (C = A + B)	835	50	48	89	1,022
PSR (A/C)%	20	48	29	34	23

PSR: positive screening rate.

Figure 1

DP consists of multiple screen cycles. At the start of each cycle, phenotyping is used to identify eligible patients from the EHR, creating an ordered list in which the most eligible patients are listed first. This list is manually screened, and the results are analyzed to improve phenotyping for the next cycle. DP: dynamic phenotyping; EHR: electronic health record.

Cycle I

The initial cohort was developed using: 1) a logistic regression model trained on a dataset of 250 patient records that were manually annotated for heart failure, 2) a regular expression for extracting ejection fraction (EF), and 3) a rule that the care provider with the highest count of cardiology notes was identified as the primary cardiologist.

Major reasons for ineligibility included high EF, incorrect identification of primary cardiologist and no heart failure. Half of the false positives were due to a combination of incorrect extraction of EF, heart failure and incorrect identification of primary cardiologist, which motivated the study team to refine the phenotyping algorithm.

Cycle II

We improved the algorithm extraction of EF by developing patterns to identify echocardiograms and by using regular expressions to extract numbers, ranges, and prose mentions of EF. This approach was found to eliminate false positives due to incorrect EF extraction. Moreover, accurate EF extraction obviated the need for the heart failure algorithm, since, by definition, EF ≤ 40 is logically sufficient to enter the intervention program.

To improve identification of the primary cardiologist, we added the restriction of having two visits to the institutional cardiologist within a 2-year period, a finding that was suggested by the navigation team. However, as observed in Figure 2, this restriction did not decrease the number of false positives.

Figure 2

Composition of manual screening performed in each cycle. 1) No HF and EF accounted for 25% of false positives. HF was detected by machine learning and EF was extracted using a simple regular expression from clinical notes. 2) Optimization of EF parser and inferring HF from low EF eliminated many false positives. But failure to exclude patients not primarily managed by the outpatient cardiology clinic at BWH emerged as a challenge. This was because the cardiologist was inferred using total number of EHRs entries authored for the patient. 3) Cardiologist was inferred from number of EHR entries limited to the outpatient setting. But this did not significantly reduce false positives. 4) Use of machine learning to infer the primary cardiologist significantly excluded patients that are not managed at BWH. HF: heart failure; EF: ejection fraction; BWH: Brigham and Women’s Hospital; EHR: electronic health record.

Cycle III

We further attempted to improve detection of the cardiologist by using the total number of EHRs authored by the cardiologists. False positives due to EF extraction errors reappeared in this cycle due to a lag in updates to the research repository; the patients had undergone echocardiography after the repository update.

Cycle IV

The limited performance of rule-based approaches for identifying the primary cardiologist led the study team to develop a logistic regression-based model, which was found to reduce false positives (Fig. 2).

PSR

The PSR improved in cycles II-IV over the first cycle, and the iterative approach resulted in an improvement in PSR; the P-value of the Chi-square statistic upon comparing average PSR for cycles II-IV (36%) and to PSR for the first cycle (20%) is < 0.01.

Discussion

A total of 1,022 patients were screened, with 223 (23%) of patients being found eligible for enrollment into the intervention program. When a single run of phenotyping is used, the pool of eligible patients is expected to decline as manual screen progresses, resulting in more false positives and a subsequently downward trend in the PSR. We used iteratively improved phenotyping in each screening cycle. This process offset the dip in the PSR by providing an enriched cohort in every cycle. Without an iterative approach, the PSR was expected to dip below the 20% measured in the first cycle; however, the cyclical approach increased the PSR to 23%.

Recruitment for prospective clinical studies has been reported to be resource intensive requiring a large number of patients to be screened in order to achieve enrollment goals [1, 17-27]. Hence, DP may be useful to improve the screening rate for clinical studies.

One advantage of the dynamic approach is that manual screening performed by the navigators in each cycle provides a better perspective on the distribution of false positives. The latter helps direct the focus of the phenotyping team for refining the algorithms for the next cycle. However, one caveat is the additional resources required for phenotyping to keep up with the rapid operational pace of the clinical study.

A major challenge for DP is the lack of research informatics infrastructure necessary to maintain the performance and real-time data needs for such analyses. Specifically, the gaps are the latency in updates to research repositories from the EHR and lack of tooling for navigators to record annotations while performing manual screening. The latency in updates to research repositories lags behind EHR updates by days to weeks. Hence, patients selected by phenotyping often have events and visits during the lag period, which leads to false positives. For instance, the false positives in cycles III and IV due to incorrect EF were a result of patients having new echocardiograms after an update to the research repository. While it is true that most research or care innovation studies do not require the most recent patient data, DP for prospective studies presents a new use-case that requires a near real-time repository [28].

The second infrastructure barrier was a lack of tooling that allows navigators to store their annotations for chart reviews. Instead, navigators used spreadsheet software to record their annotations, which required substantial curation for re-integration into the subsequent DP cycle.

Data governance also poses a challenge for DP, as researchers require the same data fields to develop algorithms that the navigators have access to in the EHR system. For instance, our research repository lacked the metadata of “clinical service type” for notes, and there was no encounter metadata for notes, both of which were available for navigators that helped researchers identify the primary cardiologist for the patient. The absence of these fields in our research repository resulted in additional effort being necessary for the development of the phenotyping algorithm.

Furthermore, we terminated the manual screening phase (in the DP cycles) when the PSR was below a subjective threshold. The manual screening was often prolonged due to logistical reasons such as delays in the updating of the research repository for executing the phenotyping. In an ideal scenario, DP cycles are executed in real-time and the phenotyping is executed after each manual screening/annotation.

Limitations

One limitation in our study was that we did not extensively investigate optimization of the phenotyping algorithms. Secondly, we did not implement a model for exclusion criteria, which forms a significant proportion of false positives (Fig. 2). Both of these limitations were due to the required fast pace of the study; in retrospect, it would have further improved the PSR. Also, the phenotyping algorithms may be converged more rapidly by utilizing active learning and semi-supervised approaches [29-31].

Conclusions

Our study demonstrates that DP can facilitate recruitment for prospective clinical study. Future directions include improved informatics infrastructure and governance policies to enable real-time updates to research repositories, tooling for EHR annotation and methodologies to reduce human annotation.