PEPRVision: A Visualization Framework for Pharmacoepidemiology Research

Abstract

Medical chart review is commonly used in epidemiologic research to evaluate health outcomes, refine measurement algorithms, and identify potential study biases. In pharmacoepidemiologic research, administrative claims and other structured healthcare databases provide longitudinal detail suitable for similar purposes but are typically presented in tabular formats that are cumbersome to interpret. This study introduces PEPRVision, a visualization framework that transforms structured longitudinal healthcare data into intuitive graphics to support patient profile review in pharmacoepidemiology research. PEPRVision was developed based on principles of preattentive visual processing, leveraging visual attributes such as color, spatial position, shape, and length to facilitate rapid pattern recognition. We demonstrated PEPRVision using US administrative claims data through four case studies: (1) evaluating a claims-based algorithm for measurement of congenital hearing loss, (2) enhancing study design decisions by identifying reverse causation in a drug safety study, (3) assessing the face validity of signals of potential teratogenicity detected through data mining for triage for advancement to causal inference studies, and (4) identifying mechanisms of inappropriate prenatal exposure to teratogenic medications to inform improvement of risk mitigation strategies. As a complement to population-based causal inference studies, PEPRVision facilitates review of patient-level variation, enhancing the validity and clinical relevance of pharmacoepidemiologic research.

Introduction

Medical chart reviews are used in clinical care and research for a variety of purposes, leveraging the longitudinal, granular detail on healthcare services to support inferences about disease etiology and the effects of medical interventions. For example, chart reviews are used to evaluate computable phenotypes identified from electronic health records (EHRs) or administrative claims data using structured information such as diagnosis or procedure codes. They are also used to triage signals from active surveillance efforts of medications and other medical technology to identify obvious biases and support decisions on whether formal causal inference studies to address a safety signal are needed.^1,2 Finally, in clinical settings, they are used to support root-cause analyses of medical errors, such as inappropriate medication use, to highlight healthcare delivery breakdowns and identify quality improvement opportunities.³ However, traditional medical chart reviews are time-consuming and resource-intensive, and vary in methodology.^4,5

In an increasing number of countries, administrative claims data are used for pharmacoepidemiologic studies that aim to make causal inferences about drug effects. Although EHRs provide more clinical detail for specific encounters, they are often limited to individual healthcare institutions and therefore cannot longitudinally follow patients across different providers and settings.^6–8 Claims data also offer more reliable information on medication exposure from outpatient pharmacy dispensing claims, as opposed to EHRs, which provide information on medication prescribing. Beyond causal inference, both claims data and structured EHR data have been used for purposes similar to those of medical chart review.

Several structured approaches have been developed to facilitate patient-level review of coded healthcare data. In the United States, the Food and Drug Administration (FDA) has used the Sentinel System’s Patient Episode Profile Retrieval (PEPR), which consolidates administrative claims into a chronological, tabular list of healthcare encounters, to evaluate the face validity of safety signals detected in active surveillance of medications and vaccines,^9–11 to improve claims-based algorithms for specific adverse outcomes,^12,13 and to evaluate medication errors.¹⁴ Similar structured profile review approaches have been developed for EHR data, such as the Knowledge-Enhanced Electronic Profile Review (KEEPER) system. These approaches present structured clinical information in tabular format; however, researchers and clinicians may at times prefer visual representations that facilitate identification of patterns, potential confounders, or data anomalies.

To address this gap, we developed PEPRVision, a visualization framework inspired by the structured profile review approach of PEPR and applicable to any longitudinal healthcare data to facilitate pharmacoepidemiology research. In this paper, we describe the development of PEPRVision and demonstrate its utility through four case studies: (1) evaluation of claims-based computable phenotypes, (2) assessment of potential biases in real-world studies, (3) triage of drug safety surveillance signals from data mining, and (4) identification of root causes for medication errors.

Method

The design of PEPRVision is grounded in the principles of preattentive processing, leveraging visual cues that are rapidly processed by the human visual system before engaging attentive processing mechanisms.¹⁵ The specific preattentive attributes¹⁶ incorporated into the design are summarized in Table 1.

Table 1.

Preattentive Visual Attributes in the Design of PEPRVision

Attribute	Implementation in PEPRVision
Orientation	Each code for diagnosis, procedure, or medication is displayed in its own row, ordered from top to bottom by date of first occurrence relative to an index date (e.g., conception, delivery, or outcome date). Within each row, repeated occurrences of the same code are displayed horizontally from left to right.
Enclosure / Grouping	Reference lines and shaded regions are utilized to emphasize relevant dates and periods, e.g., the time from conception to the outcome. Additional reference lines denote the beginning and end of the exposure of interest.
Color Hue	Distinct color coding is applied to differentiate among diagnoses, procedures, medications, and the exposure of interest.
Shape / Length	Medication exposures are depicted using a Gantt chart format to represent the duration of exposure. Diagnoses and procedures are represented by dots. In some cases, we use open circles to represent health services in outpatient settings and filled dots to represent hospitalizations.

PEPRVision requires healthcare data to be transformed into a structured format containing patient identifiers, service codes with their corresponding code category (e.g., diagnosis, procedure, or medication), service setting (inpatient or outpatient), and service start and end dates calculated as relative dates anchored to an analytic index date (e.g., date of conception). Within the observation window, records with the same code occurring on the same date are deduplicated, and all codes are mapped to their corresponding descriptions. Project-specific clinically relevant dates and time windows (e.g., the start and end dates of the exposure of interest) can be specified as additional input variables. The processed data are then used to generate PEPRVision, with the visual design implemented as described in Table 1.

In this study, we present four cases to illustrate how we have used PEPRVision in pregnancy and pediatric research to (1) evaluate a novel claims-based algorithm to identify congenital hearing loss, (2) evaluate study design decisions to mitigate design-related biases in a causal inference study evaluating the safety of doxycycline on pregnancy loss, (3) triage signals of medication teratogenicity detected through data mining for formal study, and (4) examine mechanisms of medication errors such as prenatal exposure to teratogenic medications. Data were de-identified and modified as necessary to protect patient privacy. The visualizations presented were produced using Tableau Desktop, version 2024.2 (Tableau Software, LLC, Seattle, WA). Interactive examples of PEPRVision cases are available at https://bit.ly/3LPqHrO. To facilitate broader implementation, the required data structure and a web application are provided in the Appendix S1.

Results

Case Study 1: Evaluate claims-based computable phenotypes

Claims-based pharmacoepidemiologic studies typically employ validated algorithms to measure clinical conditions. Validation usually relies on linkage to medical records and expert review of identified cases, but this process is time- and labor-intensive and sometimes infeasible due to data privacy requirements.

Figure 1 illustrates a case to demonstrate how PEPRVision was used to evaluate and refine a claims-based algorithm for congenital hearing loss (cHL).

Figure 1.

Example case of transient hearing loss misidentified as congenital hearing loss by a claims-based algorithm

To enhance validity and align with the clinical workflow for diagnosing cHL, we included infants with at least two hearing loss (HL) diagnosis codes and at least one audiometric test code. We required the following three events to occur in a specific temporal order in our algorithm. The first and second diagnosis codes needed to be at least one day apart. The date of the audiometric test should be after the date of the first HL diagnosis, and the second HL diagnosis could occur on the same date or any time after the audiometric test date. This temporal sequence was designed to mirror clinical practice, in which infants who undergo infant hearing screening receive more comprehensive audiometric testing before they are diagnosed with hearing loss. Although universal newborn hearing screening is implemented across the United States and most initial diagnoses (failed screenings) are expected to occur within the first month of life, we were also interested in evaluating cases in which follow-up after a failed screening test was delayed. This interest was motivated by national Early Hearing Detection and Intervention (EHDI) data published annually by the Centers for Disease Control and Prevention, showing that fewer than half of infants who fail screening complete a documented diagnostic evaluation in the same reporting year.¹⁷ Hence, our PEPRVision review aimed to assess whether an algorithm that relaxed the time window for cHL diagnoses to occur within the first year of life could distinguish cHL from hearing loss due to other causes acquired during infancy rather than during pregnancy.

Figure 1 displays the claims profile of an infant who met the algorithmic definition of cHL by having two separate HL diagnosis codes and at least one audiometric test. The first unspecified HL diagnosis occurred on day 141, followed by an audiology test and tympanostomy with a conductive HL diagnosis on day 143, and another conductive HL diagnosis on day 170, satisfying the sequential logic of our algorithm.

However, the claims visualized in PEPRVision revealed that the child had multiple diagnoses of otitis media (OM). After an initial OM diagnosis on day 108, additional OM diagnoses were recorded on days 136–141, just before the first HL diagnosis. Because acute or recurrent OM can cause transient conductive hearing loss in infancy, these claims likely represent a temporary hearing impairment rather than a congenital anomaly. The sequence – HL diagnoses at days 141, 143, and 170 occurring after repeated OM – is more consistent with infection-related conductive loss than with cHL.

In our final definition of cHL, we excluded conductive HL and unspecified HL cases if OM was diagnosed within one year prior to the first HL diagnosis.

Case Study 2: Evaluate study design decisions to mitigate design-related biases

This case study demonstrates how PEPRVision can prevent study design features that create, rather than eliminate bias, introduced by reverse causation or protopathic bias. Doxycycline is used to treat a variety of infections, including respiratory tract and sexually transmitted infections, pelvic inflammatory disease, and acne. Despite its therapeutic significance, concerns have been raised about doxycycline’s potentially harmful effects on fetal bone development because of its calcium-chelating properties and known effects on dental enamel.¹⁸ Well-designed observational studies using large population-based claims data can provide real-world evidence to support decision-making regarding the risk-benefit of using doxycycline during pregnancy. We recently completed a study evaluating whether doxycycline increases the risk of spontaneous abortion in pregnant women. In this cohort study, we followed patients with acne who filled prescriptions for doxycycline or for topical clindamycin, an active comparator treatment that is considered safe during pregnancy. Although used for the treatment of a broad range of infections, we intentionally evaluated doxycycline for acne because, beyond infections, acne has not been linked to pregnancy loss; hence, this design reduced the chance of confounding by indication.

While it is intuitive that, combined with an acne diagnosis, topical clindamycin was likely intended to treat acne vulgaris, the use of doxycycline, while indicated for acne, could also be for a concurrent acute infection; hence, the observed co-occurrence with an acne diagnosis could be coincidental. We therefore excluded all cases with any healthcare encounter with a diagnosis of infection within 7 days before or after the doxycycline prescription fill. PEPRVision was then used to ensure that we had created “clean” scenarios, in which, by eliminating other indications, doxycycline was used for acne and nothing else.

Figure 2 illustrates a case of a 27-year-old pregnant woman with a diagnosis of acne. The patient had been treated for acne with medications like benzoyl peroxide and topical erythromycin prior to her pregnancy. Her pregnancy ended in a missed spontaneous abortion at approximately 10 weeks of gestation. A month before her pregnancy, she underwent a routine gynecological exam, followed by her first prenatal care visit with routine testing and a diagnosis of normal pregnancy at 6 weeks of gestation. A subsequent ultrasound at 9 weeks of gestation indicated an abnormal product of conception, followed by a diagnosis of threatened abortion one day later. The following day, the patient underwent surgical treatment of the missed abortion, where retained pregnancy tissue was removed.

Figure 2.

Example case of first-trimester doxycycline exposure with pregnancy potentially ending in spontaneous abortion

The patient filled a prescription for doxycycline the day before the diagnosis of a missed spontaneous abortion and the surgery. Given the timing and context, it is unlikely that doxycycline was prescribed for acne treatment; rather, it was likely intended as surgical prophylaxis to prevent post-abortion infection, which is an established indication for doxycycline.^19,20 Including this case in the study would introduce protopathic bias, or reverse causation, where doxycycline was used in response to a missed abortion, i.e., embryo dies in utero but has not been expelled, rather than being a causal factor. Although signs of a failed pregnancy had already occurred prior to doxycycline dispensing, these codes have not been validated to confirm a spontaneous abortion and were therefore not included in our algorithm to measure pregnancy outcomes. Review of other cases also suggested that these codes are not consistently used during similar office visits that precede surgical treatment of a missed abortion. After reviewing several cases in PEPRVision, we therefore decided to delay assessment of pregnancy loss to at least seven days after doxycycline initiation to minimize the risk of reverse causation. We further required dispensing of at least 30 days’ supply, because acne is typically treated over several weeks or months, whereas abortion-related infection prophylaxis requires only short treatment courses.

Case Study 3: Triage signals of medication teratogenicity detected via data mining for formal study

In our ongoing study, Big Data Approaches for Safe Therapeutics in Pregnancy (BOOST-HP), we use tree-based scan statistics to simultaneously evaluate multiple congenital malformations after in utero medication exposure. Tree-based scan statistics allow simultaneous assessments of specific outcomes (e.g., ventricular septal defect) and in broader hierarchically organized clusters (e.g., congenital malformations of cardiac septa, congenital malformations of the circulatory system). Examination of larger clusters might enhance statistical power by increasing the number of cases, while examination of specific types of outcomes might better isolate the drug effect, increasing specificity and the magnitude of the observed association. However, a challenge to the data-mining approach is how to triage statistical signals to assess the plausibility of pharmacologic pathways. To identify potential major congenital malformations resulting from target medication exposure in BOOST-HP, we scan about 1,000 individual malformation diagnoses and groupings, organized in a hierarchical structure, to detect unexpected adverse events. Since the tree-based method searches for patterns and clusters of unusual activity but has no insight as to why the patterns or clusters occur, signals require validation and triage. Further, because a multitude of outcomes are evaluated, the use of validated algorithms for specific phenotypes is difficult.

To prioritize signals for expert review, we applied statistical criteria based on the strength of the association (e.g., relative risk) and statistical significance (e.g., p-value). We then used PEPRVision to evaluate the “face validity” of sample cases for each signal. This evaluation was focused on the validity of the outcome and obvious biases that would negate a causal association. For this case study (Figure 3), we present a case demonstrating how we evaluated a signal suggesting a possible association between azelastine, an intranasal antihistamine frequently used for allergic rhinitis, and neural tube defects (e.g., anencephaly and spina bifida).

Figure 3.

Example case of signal triage in a data-mining study on congenital malformations

The patient, a 33-year-old woman, delivered a liveborn infant at 38 weeks of gestation by cesarean section. Before and during pregnancy, she was treated for allergic rhinitis. From the 90-day pre-conception period through late pregnancy, she filled multiple prescriptions for azelastine hydrochloride as well as montelukast and mometasone, consistent with long-standing allergy management.

Routine prenatal laboratory tests and ultrasound were performed at 11 weeks and repeatedly throughout gestation. During the second trimester (weeks 14–20), maternal claims included diagnosis codes for “maternal care for suspected fetal abnormality” and “spina bifida, unspecified,” often accompanied by ultrasound and follow-up visit codes. Additional spina bifida diagnoses, including “sacral spina bifida without hydrocephalus,” appeared again during her delivery hospitalization at week 38. However, these diagnoses were limited to maternal claims; no procedure codes or specialist visits indicated confirmatory evaluation or intervention for a fetal neural tube defect.

Review of the infant’s claims from delivery through 180 days showed only routine newborn and pediatric care without evidence of spina bifida or other neural tube defects. At birth, the infant was coded as a single liveborn delivered by cesarean and received standard evaluations, including bilirubin testing, auditory screening, and immunizations. Subsequent visits documented common neonatal issues such as jaundice and feeding difficulties, along with routine preventive care including well-child visits, developmental screening, and vaccinations at 2, 4, 9, and 18 weeks. Throughout the six-month follow-up period, no diagnoses, procedures, or specialist evaluations suggested the presence of spina bifida.

The maternal and infant history of care indicate that the spina bifida diagnoses billed on maternal records most likely reflect prenatal suspicion during ultrasound assessment, rather than a confirmed diagnosis in the infant.

Because infant claims are sometimes billed under parental insurance at birth, we conducted the primary tree-based scan using both maternal and infant records. However, this case illustrates that including maternal records may introduce misclassification of outcomes. To assess the impact of this potential misclassification, we conducted a sensitivity analysis restricted to infant-only records.

Case Study 4: Identifying Mechanisms of Prenatal Exposure to Isotretinoin

This case study focuses on utilizing PEPRVision to identify the mechanisms of how prenatal exposure to teratogenic medications occurs within a strict Risk Evaluation and Mitigation Strategy (REMS) program. REMS programs encourage or enforce safe use behaviors that prevent certain medication side effects to ensure the medication’s benefits outweigh its risks.²¹ To prevent prenatal exposure to medications that are contraindicated in pregnancy because of their teratogenicity, REMS programs can set certain requirements, such as mandatory use of contraceptives or pregnancy testing before drug initiation. While some REMS programs have been shown to be effective in reducing prenatal exposure,^22–24 pregnancies exposed to teratogenic medications that have a REMS continue to occur.^25,26 Case views similar to root cause analyses conducted for sentinel events in clinical practice can elucidate our understanding of the circumstances of prenatal exposure to teratogenic medications and help improve the design of the REMS program to ensure patient safety.

Isotretinoin is one of the most potent teratogenic medications, associated with a 20% risk of spontaneous abortion and an 18–28% risk of embryopathy in live births.²⁷ Indicated for treatment of severe recalcitrant acne unresponsive to conventional therapy,²⁸ it is commonly used by women of childbearing age, making it one of the most critical targets for prevention of prenatal exposure.²⁹ Isotretinoin has a REMS program that requires a 30-day wait period while using contraception before the medication can be prescribed. Two negative pregnancy tests at least 19 days apart, with the second test taken during the first 5 days of the menstrual period immediately before starting the medication, are required during this wait period. While the requirements for pregnancy testing were relaxed to allow home testing during the COVID-19 pandemic, during the study period, both initial tests had to use specimens collected during an office visit and the second test had to be conducted by a certified laboratory. The prescribed days’ supply is limited to 30 days, and a repeat pregnancy test is required each time a new 30-day supply of isotretinoin is prescribed. Pharmacies must check the pregnancy test result entered by the physician into an online registry before they may dispense the medication. In addition, patients must agree to use two forms of birth control unless they commit to sexual abstinence or are sterile.³⁰

In our efforts to understand causes of isotretinoin REMS failure, we identified pregnancies with prenatal exposure in administrative claims data using previously described methods.²⁵ Records of pregnancies with exposure to isotretinoin were then used to construct PEPRVision graphs and reviewed. In Figure 4, we show a real-world case representing a scenario where conception occurred during isotretinoin treatment. A 24-year-old female patient has several pharmacy claims for oral hormonal contraceptives in the 150 days prior to conception. Estimation of exposure duration based on dispensed days’ supply suggests that contraceptives were discontinued about 70 days before conception. Because prescriptions and therefore claims do not capture the indication for prescribed medications in the US, it is unclear whether the hormonal contraceptive was used for contraception or acne. Acne appears as a diagnosis on office visits repeatedly and is, without doubt, the indication for isotretinoin, likely prescribed during an office visit 54 days after discontinuation of the hormonal contraceptive and dispensed six days later. Two pregnancy tests (hCG tests) occur prior to the isotretinoin prescription fill. Conception occurs two weeks after isotretinoin initiation. The pregnancy ends in a missed abortion at about 2 months of gestation, which is a common outcome following in utero exposure to isotretinoin.

Figure 4.

Example forensic case study of iPLEDGE program failure resulting in prenatal isotretinoin exposure

Prenatal exposure to isotretinoin can occur either when physicians initiate the medication during pregnancy, attributable to failure to test for pregnancy, or when patients conceive during use due to the ineffectiveness of contraception. These two failure pathways would require different improvements in the REMS program, and quantifying their relative contributions is an important next step. In this case, the patient was not using contraception, likely because she had pledged sexual abstinence, but she conceived shortly after isotretinoin initiation.

Discussion

We developed an intuitive and easily interpretable visualization framework, PEPRVision, to visualize patients’ healthcare services, including sequences of diagnoses, medical procedures, and medication use, providing a complete longitudinal picture of information captured in structured healthcare data. We demonstrate the use of PEPRVision in four areas: (1) evaluating claims-based computable phenotypes, (2) investigating potential biases in study design, (3) assessing the validity of signals detected in safety surveillance data mining, and (4) identifying the etiology of medication errors. PEPRVision emphasizes the chronological alignment of health care encounters with diagnoses, procedures, and medication use, using distinct visual elements to differentiate types of healthcare services. The patient profile display can span several pages, and reference lines can be added to highlight key periods such as pregnancy or drug exposures of interest, helping investigators interpret events in context. Unlike traditional tables where repeated diagnoses, procedures, or medications appear as multiple rows, PEPRVision places them on the same row, making recurrent conditions, procedures, or prescriptions immediately recognizable as ongoing or recurring patterns. This design makes it easy to assess temporal and proximate relationships between exposures and outcomes, and supports the assessment of clinical plausibility and potential biases in observational studies using routinely collected health data.

Several limitations should be acknowledged. First, when applied to claims data, PEPRVision is limited to the information available in these data. Claims data typically do not capture test results, radiology interpretations, or other details of clinical assessments (e.g., progress notes). On the other hand, unlike electronic health records, they are not constrained to specific healthcare systems but instead reflect the entirety of healthcare services that were covered by health insurance. Second, because claims are created for billing rather than research purposes, coding follows administrative rules that may not reflect clinical priorities and will not capture services paid for out of pocket. Third, the dates we used to build PEPRVision are derived from claims and do not always reflect the true timing or consumption of a medical service or medication. For example, drug exposure is assumed to begin on the dispensing date and end when the days’ supply is exhausted, although patients may delay initiation or discontinue treatment early. As a result, the Gantt bars representing medication exposure may imply continuous uninterrupted use when, in reality, patients may use medications intermittently or discontinue early. Similarly, conception is typically estimated based on gestational age from the delivery claim, or when unavailable, using proxies such as recommended prenatal screening schedules. Hence, the chronological layout anchored to an index date may create an impression of temporal precision that the underlying data do not fully support. This uncertainty is illustrated in Figure 4: while oral contraceptives can fail, it is also possible that the patient did not adhere to treatment. Alternatively, conception may have been estimated too early, though this is unlikely since the first prenatal visit with a pregnancy diagnosis occurred during the estimated exposure period. A thorough understanding of how claims data are generated, as well as the delivery of medical care, is essential when using claims data for research. Furthermore, PEPRVision is designed for structured coded data and cannot incorporate unstructured clinical documentation. As such, PEPRVision supports face validity assessment and contextual interpretation but does not replace a traditional medical chart review for outcome validation. The latter, however, is also limited by the completeness and accuracy of information documented in medical records.

Despite these limitations, both epidemiologists and clinicians on our team value the clinical detail available for individual cases, especially in vetting design and measurement decisions in causal inference studies. When permissible, all cases in a causal inference study can be made available on a PEPRVision platform to allow peer reviewers and readers to directly examine the data underlying medication safety or effectiveness studies that reported significant risk estimates. Importantly, in contrast to manual chart review based on unstructured notes, the ability to magnify a specific exposure/outcome period and highlight prespecified variables increases the efficiency of the review process, allows processing of larger numbers of cases in a limited amount of time, and hence, enhances the ability to identify important patterns that can help refine measurement algorithms and study design. The PEPRVision visualization framework could be extended to structured EHR-based patient-profile systems such as KEEPER and adapted to non-pregnant clinical populations by redefining the anchor event and tailoring clinical domains. Future extensions could incorporate key clinical information extracted from unstructured text through natural language processing. Another important area for future research is the formal evaluation of the framework’s impact on review efficiency and decision-making.

Conclusion

PEPRVision effectively visualizes patients’ longitudinal healthcare services, providing a framework for disaggregated, qualitative review of patient-level data that complements the aggregated results of population-based studies. By integrating PEPRVision with analytical approaches, researchers can better understand root causes of medication errors and other quality deficits in healthcare, assess the validity of signals identified through data mining techniques, detect potential biases in study design, and refine measurement algorithms.

Data Availability

All data are incorporated into the article and its online supplementary material.