Harnessing Electronic Health Records to Study Emerging Environmental Disasters: A Proof of Concept with Perfluoralkyl Substances (PFAS)

1.1 Exposure to Environmental Disasters Impacts Human Health

Environmental disasters are catastrophic events that occur and are the result of human activity (i.e., anthropogenic) and often inadvertent. These differ from natural disasters, which are presumed to be non-human caused (e.g., hurricane, volcano eruption), and intentional acts such as nuclear bombings. Environmental disasters often affect human health directly. Famous environmental disasters include the Chernobyl nuclear meltdown in Ukraine on April 26, 1986. The resulting effects on human health included an increased rate of spontaneous abortions, including an excess loss of 400 male fetuses following first-trimester exposure, and sex-ratio disturbances observed throughout Northern Europe followed by increased rates of thyroid cancer ^1-5.

Another devastating environmental disaster was the Seveso Herbicide plant explosion that occurred on July 10, 1972 in Seveso, Italy. This resulted in the highest known exposure to dioxin or 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD) in residential populations ^6-9. Spontaneous abortions (i.e., miscarriages) increased 67.7% following the Seveso plant explosion, in addition fathers exposed tended to have more female offspring then expected^6-9. Chloracne (a dermatological condition) also resulted but not due to prenatal exposure. More recently, the Fukushima-Daiichi nuclear disaster occurred in Japan on March 16, 2011 resulting in increased exposure to radiation and other chemicals. Review of medical and health records revealed no observed change in the fetal loss or anomaly rate following the Fukushima-Daiichi nuclear disaster ¹⁰.

The typical epidemiology framework for studying environmental disasters includes conducting interviews of the exposed and affected populations, surveys to collect health outcomes, and in many cases review of health records. High-quality epidemiology studies use biomarkers of exposure, including serum concentrations of PFAS. However, these are expensive, typically costing over $100 per test and thereby limiting the sample size for study populations. In addition, there are many different species of PFAS (described further below).

1.2 PFAS Exposure: a Critical Environmental Disaster

A recent environmental disaster in the United States of America (US) involves perfluoroalkyl substances (PFAS). PFAS are anthropogenic persistent organic pollutants resistant to degradation and high heat and consist of different species of Fluorinated Organic Compounds (FOCs). PFAS were used in a variety of products (and in some cases are still used although production in the US has ceased). Products include firefighting foams, non-stick cookware, dental floss, fire and chemical resistant tubing, and more (Table 1)¹¹. Due to their chemical structure, PFAS have surfactant properties, and are resistant to degradation and high heat. The two most common PFAS species found in serum in the US are perfluorooctanesulfonate (PFOS) and perfluorooctanoate (PFOA) ^12,13. The effect of PFAS on human health is only recently gaining awareness and therefore could be termed an emerging environmental disaster.

PFAS were detected in the Philadelphia area, specifically in Horsham-Warminster-Warrington following a series of events occurring in 2014 through May 19, 2016. In 2014, the established Provisional Health Advisory Levels (PHAL) were 200 ppt for PFOS and 400 ppt for PFOA (both types of PFAS). The Warrington Township Water and Sewer Department (WTWSD) promptly shut down public wells 1, 2 and 6 on October 29, 2014 (the same day they received the sampling results). Horsham Water and Sewer Authority (HWSA) took Well No. 26 and Well No. 40 off-line by August 2014, and three other wells had PFAS below the PHAL (HWSA Wells 10, 17 and 21). Warminster Municipal Authority (WMA) removed three wells

On May 19, 2016, US Environmental Protection Agency (EPA) released a new combined Health Advisory Level (HAL) for PFOS and PFOA of 70 ppt. WTWSD promptly shut down public wells 3 and 9 that very same day. HWSA shut down 3 wells in compliance. WMA removed another three wells, and started purchasing supplemental water supply from the Forest Park Water Treatment facility through the North Wales Water Authority (NWWA). Overall, three of the top 10 water utilities with the highest nationally recorded PFAS levels were located close to Philadelphia: Horsham-Warminster-Warrington ¹⁴.

PFAS exposure is known to increase the risk for various health outcomes, including high cholesterol¹⁵, ulcerative colitis^16,17, thyroid disease^17,18, kidney cancer¹⁹, pregnancy-induced hypertension, and preeclampsia^17,20. In highly exposed US populations, indirect estimates of PFAS exposure were associated with preeclampsia ^21,22, which puts these women at increased risk of cardiovascular disease (CVD) later in life²³. Several observational studies have found PFAS exposure associated with high serum uric acid levels ^24-26; epidemiologic evidence has linked elevated serum uric acid concentrations to hypertension ²⁷.

1.3 Informatics Methods are needed to Harness Electronic Health Records for the Study of Health Impact

As mentioned previously, environmental epidemiologic studies typically rely heavily on interviews of the exposed and affected populations, surveys to collect health outcomes, recruitment of affected populations and in many cases manual review of health records. These studies require a painstaking amount of work and manual curation of specially designed datasets.

Data science has exploded these past few years with increasing numbers of studies seeking to incorporate ‘big data’ towards studying human health and disease. Electronic Health Records (EHRs) contain millions of patient records across billions of visits, including details such as laboratory values, medications, procedures, diagnostic imaging, clinical notes, and demographic data. Therefore, EHRs contain a treasure-trove of data that could be useful for studying the health impacts of environmental disasters, such as PFAS exposure in a given population, if the appropriate informatics methods are applied.

Previously, several disease-environment correlation studies have been performed using EHRs. The first was Patel et al. who coined the term EWAS: Environment-Wide Association Study ²⁸. Others have studied environmental challenge in disease progression ²⁹, thyroid cancer hotspots in a rural community in Vermont ³⁰, and air pollution data in Italy ³¹. We have also conducted several studies investigating the relationship between birth season (as a proxy for seasonal variance in environmental exposures at birth) and disease risk ³². We conducted our initial study using New York City (NYC) data with novel findings that were validated in a separate EHR system ³³. We expanded our study to include six sites including South Korea and Taiwan ³⁴. This enabled us to develop a method that correlated birth season and trimester information with climate and pollution variables using location information ³⁴. We have also used our methods to study the effects of social exposures that can manifest as birth month relationships for Attention Deficit Hyperactivity Disorder (ADHD) ³⁵ and increased the number of datasets included in our assessment of female fertility and birth season to assess deeper and more complex multi-factorial environmental exposures ³⁶. We have found that prenatal exposure to fine air particulates increased the risk of atrial fibrillation later in life in humans ³⁴; this finding was supported by a study in a canine population ³⁷.

These methods demonstrate that informatics methods can be developed to overcome EHR biases ^38,39 and can be used to probe disease-environment interactions. However, they fall short in that they assess environmental exposures that are always present (e.g., air pollution) rather than the effect of an environmental disaster on human health and disease. An environmental disaster is more complex in that there is an exposure period, a remediation period and a post-exposure period. This differs from less-severe exposures that do not have a Government mandated remediation, clean up and post-remediation period.

This purpose of this study is to present a proof-of-concept detailing how EHRs can be used to study the impact of environmental disasters, such as PFAS exposure. We will compare the results generated from our EHR-only study to those that resulted from carefully curated epidemiology and government-based studies to demonstrate that we can effectively capture the health impacts of environmental disasters using EHR data only.

2.1 Integrating Diverse Data Sources to Define PFAS ‘Contaminated Regions’

The PFAS exposure in Horsham-Warminster-Warrington, PA was via the water supply. However, these three towns are located outside of Philadelphia city proper and therefore, some homes are on city/town water line while others are on private water wells. Exposure via the public water utilities (city/town water) is shown in Figure 1.

• Download figure
• Open in new tab

Figure 1. PFAS Exposure Levels as Reported in Public Water Utilities.

Results over the Minimum Reporting Level (MRL) defined by UCMR-3 are recorded in µg/L and null values indicate results less than the MRL.

2.2 Linking Individual-Level Exposure to PFAS ‘Contaminated Regions’ with EHRs

For the purposes of this proof-of-concept study, we focused on women living in the three towns that were exposed to extremely high levels of PFAS. These towns are Horsham, Warminster and Warrington ¹⁴ referred to as Horsham-Warminster-Warrington throughout this study. These towns receive water supply from three of the top 10 water utilities with the highest nationally recorded PFAS levels in the entire US ¹⁴. Using address information contained in the EHRs, we linked patients to their precise geolocation within the Horsham-Warminster-Warrington area for visualization purposes. We used zip code information for the three towns, Horsham-Warminster-Warrington, to extract patients in exposed regions for our association analyses. We then extracted all diagnosis information contained in International Classification of Diseases (ICD) coding schema versions 9 and 10 for data collected between 2010-2017. These data were obtained during routine clinical care and stored in EHRs. The first PFAS exposure was detailed on May 19, 2016 and remediation occurred subsequently. We have data through 2017 and therefore all data is either pre-remediation or during the beginning of the remediation period. The biological half-life for PFAS species (shown in Table 1) ranges from 2 to 10 years depending on the compound⁴⁴: 2-4 years for PFOA, 4-6 year for PFOS, and 8-10 years for PFHxS ⁴⁴. Therefore, all patients with EHR data between 2010-2017 found in this heavily PFAS-exposed region would be expected to have elevated PFAS levels.

2.3 Statistical Analysis: Determining if Known PFAS Exposure-Disease Associations are Captured Using EHR clinical data alone

First we split our data into inpatient and outpatient records. We report the results obtained from the analysis of the inpatient records as these are less subjected to data missingness. We then performed Fisher’s exact test to determine the association between disease/condition/symptoms codes having at least 100 patients diagnosed at PennMed. WE compared the disease/condition/symptoms frequencies between those heavily exposed to PFAS (i.e., living in the Horsham-Warminster-Warrington towns) versus the overall cohort of patients treated at PennMed. Our approach was not ‘hypothesis-free’ in that we focused on diseases that were known to be related to PFAS exposure. This allowed us to confirm if EHRs are able to capture the known PFAS-disease relationships established in the epidemiology literature. These known PFAS-disease associations were established using carefully curated epidemiology studies and therefore they serve as a ‘gold-standard’. Associations have been established between high PFAS exposure and high cholesterol¹⁵, ulcerative colitis¹⁶, thyroid disease¹⁸ and kidney cancer¹⁹. Importantly, this research is mainly with PFOA and PFOS rather then all PFAS species found in the Horsham-Warminster-Warrington area. We will determine if we can effectively replicate all known PFAS-disease associations reported in the literature. We are excluding associations that are related to pregnancy (preeclampsia) as this requires capture of the pregnancy timeline. In addition, we are excluding testicular and prostate cancer, as our population is restricted to women.

The University of Pennsylvania’s Institutional Review Board approved this study.

3.1 EHRs can Capture Patients Exposed to Environmental Disasters such as PFAS

Using the zip code for individuals who were treated at Penn Medicine, we identified those living in Horsham-Warminster-Warrington. All of these women were included in our association analysis (results shown in section 3.2). For mapping purposes, we linked address information recorded in the EHR to latitude and longitude. We identified the two major contamination sites in the Horsham-Warminster-Warrington area (depicted as gray shapes in Figure 2), along with the private well data denoting wells where PFAS contamination was reported for those exposed to the heavily polluted areas of Horsham-Warminster-Warrington (Figure 2). Contaminated wells are denoted with red circles in Figure 2.

• Download figure
• Open in new tab

Figure 2. Heatmap detailing a subset of PennMedicine patients living in Horsham-Warminster-Warrington, PA along with PFAS contamination sites and water wells

3.2 Validation of known disease-PFAS associations using clinical data from EHRs only

Women living the Horsham-Warminster-Warrington are exposed to PFAS at high levels. Therefore, we investigated whether or not we could replicate diseases known to be associated with high PFAS exposure among women visiting PennMed that lived in Horsham-Warminster-Warrington between 2010 and 2017 (during the highly exposed period). We investigated all diseases known to be elevated among women exposed to PFAS and compared the disease incidence to the background population. We found that high cholesterol, colitis and thyroid diseases (including multiple thyroid conditions) were elevated among women living in Horsham-Warminster-Warrington (Table 2). We also found the proteinuria was elevated, which is related to the elevated uric acid reported in the literature. We did not find kidney cancer to be elevated among these women, but found that complications related to kidney transplants were elevated, which could be indicative of prior kidney cancer (given that if kidney cancer was present previously then the diseased kidney may have been replaced with a transplanted kidney). We used Steenland et al.’s comprehensive review⁴⁵ to determine ranges of odds ratios (ORs) reported in literature for diseases linked with PFAS exposure. If multiple studies were found, we took the study with the lowest reported OR, which for high cholesterol was 1.38 ^45,46. There are reports of cardiovascular disease and cerebrovascular disease increases among PFAS exposed but these studies were too small for general conclusions according to Steenland ⁴⁵. This review found no relation between kidney function and PFAS exposure; however, another study ¹⁹ reported a hazard ratio of 1.10 for kidney cancer for every 1 unit increase in PFOA in serum. Therefore, we looked for any kidney-related diagnoses that were nominally associated with PFAS exposure and found complications of transplanted kidney, which could indicate some impairment of kidney function among heavily exposed.

Overall, these results demonstrate that clinical outcomes known to be due to high PFAS exposure in the literature are extractable from EHRs. In addition, it demonstrates that EHRs can be used to study the impact of environmental disasters on human health and disease without the need for specialized study recruitment and surveys.

4.1 Power of EHRs for Studying Impact of Environmental Disasters: Reduced Cost Enables More Exposed Populations to be studied

Epidemiology studies focused on environmental disasters – such as PFAS exposure in Horsham-Warminster-Warrington are expensive and often limited to a particular disease area, such as cancer. Government agencies, such as the EPA and the Agency for Toxic Substances and Disease Registry (ATSDR), often need to perform public health assessments on hazardous sites facing designation as a Superfund or Brownfield site. However, these public health agencies face many resource and financial constraints when deciding what outcomes to study, the wash out period following exposure, and in which locations ⁴⁷. This leads to intense frustration in communities if their locations are not selected for further study, or if the outcomes they are most interested in are not explored – especially if they are already noticing alarming signs of deteriorating health among their loved ones. Even when a community is selected for study, it can be incredibly difficult for community members to wait for the results of the assessment, let alone the long process during which EPA and ATSDR decide on removal and remediation efforts. We should point out that the CDC did not select the Horsham-Warminster-Warrington site for future study likely due to budgetary constraints. In this current funding environment, studies using less expensive methods by harnessing EHR data become imperative.

The power of using EHRs to study environmental disasters such as PFAS, is that a wide range of health outcomes can be relatively quickly explored in an EWAS-style manner ²⁸. The purpose of this current study is to serve as a proof-of-concept to demonstrate that we can replicate the known diseases associated with PFAS exposure in gold standard epidemiologic studies. This is a first step towards further detailed high-throughput analyses on diseases that individuals are at increased risk for following PFAS exposure. Furthermore, we demonstrate that EHRs are a powerful tool that can be used by public health investigators who are interested in studying disparate health outcomes.

4.2 Power of EHRs for Studying Impact of Environmental Disasters: Increased Inclusion and Diversity

Socioeconomic analyses in the study of PFAS exposure and its subsequent effects on human health and disease are also important ⁴⁸. In this proof-of-concept, we ignore the socioeconomic factors because our desire was to replicate prior research on the health risks associated with PFAS exposure and the prior research often ignores socioeconomic co-contributors. A meta-analysis found that with a doubling of income the exposure to PFAS increased on average 10-14% ⁴⁹. This is contrary to the typical environmental justice hypothesis that higher income individuals have lower exposure to pollutants in general ⁴⁹. Many theories exist to explain this counter-intuitive result; some conclude that it could be due to routes of exposures (e.g., dental floss) or perhaps a result of the way many prior studies were conducted.

However, traditional epidemiology studies often require patient recruitment and consent and therefore are likely to result in inequities in part because recruiting representative populations is challenging ⁵⁰. Recruiting minority populations is especially challenging due to historical inequities and distrust ⁵⁰. Using EHRs to study environmental disasters, such as PFAS exposure, would address many of these issues. With a ready availability of clinical data from the EHR, a representative sample of the exposed population (in terms of race/ethnicity/sex/gender/various income levels) can be derived. This can mitigate some of the data collection and case ascertainment challenges experienced by traditional epidemiology studies.

4.3 Statistical Methods with EHR data can enable better Assessment of Biases in Environmental Studies

Analyzing EHR data can enable elucidation of issues with specific datasets, for example if a clinical dataset is derived from one type of clinic (e.g., ‘oncology’) the quality of the health assessment for other diseases that are non-oncology related will be affected ⁵¹. These dataset stratification issues could result in a biased assessment of certain exposure-disease related findings. Much work has been done to assess the bias of EHR methods and their portability from one clinic to the next ^34,52-55. However, many environmental studies consist of meta-analyses of results across different sites without exploring site-specific biases that may affect the results. These types of biases can be addressed using informatics methods applied to EHRs that contain rich phenotypic data that is not often available from epidemiology studies. This is another strength of using EHRs and informatics methods to study environmental disasters. In addition, we open a new area of research for informatics through the exploration of health effects following environmental disasters.

4.4 Our Findings Versus the Literature: PFAS and Health Risks

4.5 Limitations and Future Work

There are several limitations of our work. We investigated findings reported in the literature to determine if we could replicate these well-known findings. We successfully validated known PFAS-disease exposures. However, it is important to note that other environmental exposures can also cause many of these diseases. For example, Endocrine Disrupting Chemicals can disrupt the thyroid system (e.g., PCBs and Dioxins). Our primary purpose was to determine if EHRs were useful in studying environmental disasters - given our ability to replicate the known PFAS-disease relationships, we conclude that they are a useful and cost-effective method of studying the health impacts of environmental disasters. However, additional studies would be required to rule out other exposures and concomitant comorbidities that may play a role.

From an informatics perspective, the alignment between diagnosis, symptom and condition codes found in the EHR and the diseases reported in these prior studies may not be exact. Future work, includes our plan to link disease codes from ICD-9 and ICD-10 to higher-level disease categories using ontologies such as the Disease Ontology or the Human Phenotype Ontology to assist in linking these codes to exact disease entities (although many of these methods have other limitations). However, it is still likely that those entities may differ from what was reported in the literature by other researchers. Therefore, we also plan to expand our analysis to include laboratory values; however, this is challenging due to the many different laboratory-testing services that exist with different reference standards and the porting of information in the EHR usually occurs in PDF format. Therefore this is the subject of future work. We also plan on using medication histories to determine treatment status (especially important for diseases/conditions like hypercholesterolemia and thyroid disease).

In addition there is the limitation of address accuracy, we used patient zip code information - if this was incorrect then we may be misclassifying patients as cases. Also, we do not have information on whether or not patients were using bottled water or not - therefore, it is likely we are underestimating the true effect of the PFAS exposure on human health. In addition, we classified everyone living between 2010-2017 in the Horsham-Warminster-Warrington as being exposed - this might not be the case if someone moved to the area in 2017 after the remediation period. This limitation also would lead us to underestimate the true effect of this exposure on human health.