Using electronic health record data to link families: an illustrative example using intergenerational patterns of obesity

Abstract

Objective

Electronic health record (EHR) data are a valuable resource for population health research but lack critical information such as relationships between individuals. Emergency contacts in EHRs can be used to link family members, creating a population that is more representative of a community than traditional family cohorts.

Materials and Methods

We revised a published algorithm: relationship inference from the electronic health record (RIFTEHR). Our version, Pythonic RIFTEHR (P-RIFTEHR), identifies a patient’s emergency contacts, matches them to existing patients (when available) using network graphs, checks for conflicts, and infers new relationships. P-RIFTEHR was run on December 15, 2021 in the Northwestern Medicine Electronic Data Warehouse (NMEDW) on approximately 2.95 million individuals and was validated using the existing link between children born at NM hospitals and their mothers. As proof-of-concept, we modeled the association between parent and child obesity using logistic regression.

Results

The P-RIFTEHR algorithm matched 1 157 454 individuals in 448 278 families. The median family size was 2, the largest was 32 persons, and 247 families spanned 4 generations or more. Validation of the mother–child pairs resulted in 95.1% sensitivity. Children were 2 times more likely to be obese if a parent is obese (OR: 2.30; 95% CI, 2.23–2.37).

Conclusion

P-RIFTEHR can identify familiar relationships in a large, diverse population in an integrated health system. Estimates of parent–child inheritability of obesity using family structures identified by the algorithm were consistent with previously published estimates from traditional cohort studies.

BACKGROUND

With the widespread adoptions of electronic health records (EHRs), population health researchers increasingly use EHR data for epidemiological studies.¹ EHRs are a rich source of information and longitudinal data including lab measurements, insurance information, procedures, prescriptions, and emergency contact (EC) information. They can be used to strengthen both public health surveillance and the link between clinical care and public health practice.² However, EHRs are currently limited in several important ways, one of which being the lack of information about how individuals are related to each other. As family members are often part of the same healthcare system, connecting them would provide opportunities to answer important research questions around intergenerational health. Conventional research on families can be costly and may be nonrepresentative particularly when using twin or family cohort studies. EHR-based studies, which have been shown to reduce cost and time involved, may generate estimates that are more generalizable to certain patient populations.¹

One area that has been widely researched using both EHR and traditional cohort studies is obesity. Obesity and cardiovascular health have known associations with social determinants of health.^3–5 Factors that influence obesity include income, structural and community factors, cultural beliefs, social networks, chronic stress and education, among others.⁶ These factors tend to cluster within families—typically among people you would identify as an EC. Additionally, weight and height (and therefore BMI) in the EHR have low error rates; less than 1% in children.⁷

To define how individuals within the EHR are related to each other, a group at Columbia University developed an algorithm titled “Relationship Inference from the Electronic Health Record” or RIFTEHR.⁸ RIFTEHR demonstrated it is possible to create family trees using EC information in the EHR. However, the algorithm had yet to be tested in health systems outside New York City, and executing the algorithm required using multiple programming languages.

OBJECTIVE

We built on their previous script to create a Pythonic RIFTEHR (P-RIFTEHR) algorithm, simplified to use one programming language.⁹ Our aim was to improve its utility across different health systems with different population characteristics and optimize the code for scalability. We then modeled intergenerational patterns of obesity—a topic well-published in existing literature—to provide proof-of-concept for the P-RIFTEHR algorithm.^6,10–16 We hypothesized that families identified using P-RIFTEHR would share similar associations to those found in existing literature using family cohorts.

MATERIALS AND METHODS

Data source

The Northwestern Medicine Enterprise Data Warehouse (NMEDW) is a repository of all clinical and research data sources on the Northwestern Medicine campus to facilitate research, clinical quality, healthcare operations, and medical education. The NMEDW contains EHRs from approximately 10 million unique individuals. Of those, 3.6 million had EC information, and of those, 2.95 million have a contact date between January 1, 2000 and June 1, 2019. After cross-matching ECs with patients in the EDW, 1.2 million had ECs within the EDW (Figure 1). Ethical approval for the study and resulting analysis were provided by institutional review board at Northwestern University. Additional ethical concerns have been detailed in the Supplementary Material S1. Each part of our data release was approved by the appropriate compliance governance bodies: IRB approval and Northwestern Medicine Healthcare Corporation (NMHC) Data Steward approval. The IRB deemed this to be minimum risk to the subjects and thus granted a waiver of HIPAA authorization and waiver of patient consent for this research. The data underlying this article cannot be shared publicly due to our agreement to not reuse or disclose PHI inherent in those waivers.

Figure 1.

Validation flow chart. EC: emergency contact; EDW: Enterprise Data Warehouse; PRIFTEHR: Pythonic-Relationship inference from the electronic health record.

Pythonic RIFTEHR

The P-RIFTEHR algorithm is based on the RIFTEHR algorithm by Polubriaginof et al. The basic structure of the algorithm and its validation is described in detail elsewhere.⁸ Our team translated the existing SQL and JULIA code into Python and rewrote certain sections of Python code. The original RIFTEHR program consists of 5 steps that are not specific to the original dataset: matching provided patient EC data with other patients in the dataset, inferring new matches from those provided, assigning family IDs, checking for conflicts, and identifying twins. Three major changes from the original RIFTEHR code were made to increase computational speed and to increase confidence in the accuracy of the final results; we will detail the changes made in the P-RIFTEHR version here.

In the first step of identifying EC matches, both programs use 4 data points to search for a match: EC first name, EC last name, EC phone number, and patient zip code. In the P-RIFTEHR version, individuals are first matched on 4 of 4 data points, if all 4 elements do not match, then individuals are paired if any 3 out of 4 data points match. If 3 out of the 4 do not match, individuals are matched only if they match on first name and phone number, last name and phone number, or ZIP code and phone number. Finally, if a pair has not yet been found, they are paired if there is only a phone number match. Unlike the original program, the P-RIFTEHR program does not match based on first and last name, first name and ZIP code, last name and ZIP code, first name alone, last name alone, or ZIP code alone. While this does reduce the total number of matches initially returned, it also significantly reduces the number of conflicting relationships identified that would require later reconciliation, including the number of spousal relationships with one or both patients under 17 years old and the number of parent–child relationships with a less than 10-year age gap. This suggests that the new algorithm is reducing the total number of incorrect matches, including those that would not be filtered out through our age filters, and increasing the accuracy of the results in our dataset.

The second major change in the P-RIFTEHR program was also made to increase the accuracy of the results. While the original program inferred new relationships, assigned family IDs, and checked for relationship conflicts in 3 separate steps, the P-RIFTEHR program compresses these into one step. First, families are created using network graphs,¹⁷ sorted by size, and assigned a family ID. Then, within each family, we check each relationship for age conflicts (children who are older than their parents or grandparents, parents and children who are less than 10 years apart in age, spousal relationships where one or both are less than 17 years old) and for outright conflicts (person A says that person B is their parent, person B says that person A is their spouse). Opposite relationships are then created from remaining relationships (if A is the father of B, then B is the child of A). Finally, within each family ID, new relationships are inferred. For example, if A is the mother of B and B is the mother of C, the algorithm will infer that A is the grandmother of C. This new inferred relationship is then checked for age conflicts and for conflicting existing relationships as above before getting added to the dataset so that no new relationships can be inferred using an incorrectly inferred relationship, improving the accuracy of the results from the original algorithm. The inference step is then repeated until no new relationships are inferred. The inference logic of this step is outlined in the Supplementary Material S2.

The third major change was made to greatly increase computing speed. By assigning a family ID using network graphs before inferring new relationships between family members instead of after, the inference step is run only within families instead of across the entire dataset. This change increased the speed of the inference step by a factor of 60 on a subsample of the data (100 000 patients). This speed gain is somewhat reduced when running across the entire dataset because of the computation required to process the largest families, but still represents a major gain. Unfortunately, we do not have an A/B testing speed for this across the entire dataset because it was run on different systems, and the original algorithm’s computation time is too prohibitively large to run just for testing purposes. Figure 2 illustrates the main functionality of both algorithms and the differences between the 2.

Figure 2.

Illustration and comparison of RIFTEHR and P-RIFTEHR algorithms. PRIFTEHR: Pythonic-Relationship inference from the electronic health record; RIFTEHR: relationship inference from the electronic health record.

Validation

NMEDW includes a linkage between mothers and their children who are born within the Northwestern Medicine system. Using this linkage as the gold-standard reference, we calculated the sensitivity by determining the proportion of true mother–child relationships that were correctly identified by P-RIFTEHR.

Modeling intergenerational patterns

After running the P-RIFTEHR algorithm in the EDW, we pulled additional data for all the parent–child dyads. Dyads with at least one body mass index (BMI) (or weight and height) measurement in the EDW were included in analysis as long as the parent’s BMI measurement occurred prior to the child’s. BMI from visits that occurred after bariatric surgery were excluded. Additional characteristics were insurance status (private, public, and self-pay/unknown), age at BMI measurement, sex, race/ethnicity (non-Hispanic White, non-Hispanic Black, non-Hispanic Asian, Hispanic, and other/unknown) and presence of comorbidities using the Charlson Comorbidity Index¹⁸ for both parent and child.

Obesity was calculated for both parents and children as a dichotomous variable. BMI percentile was calculated using the CDC calculation¹⁹ for any individual whose BMI was from a visit when they were less than 20 years old. An individual was considered obese if their BMI percentile was greater than or equal to the 95th percentile or their raw BMI (age 20+ years) greater than 30 kg/m².

To model the relationship between parent–child obesity, we used logistic regression models. Models were adjusted for child sex, race/ethnicity, age, insurance status, and Charlson comorbidity index. To demonstrate the strength of capturing intergenerational data, we simplified the data structure in 3 ways and applied appropriate modeling strategies.

First, we randomly sampled one parent and child pair per family ID; the independent nature of these parent–child dyads allowed us to use a simple logistic regression model to estimate the association between parent and child obesity. The second data structure included one randomly sampled parent per family and all their children. The corresponding model was a mixed-effects logistic regression model with a random parent intercept, adjusted for the same covariates as the previous model. The third data structure included all the parents and children in each family; the corresponding model included a random intercept for family ID. While we would have preferred to use 2 random effects in this setting, one for parent and one for family, due to the dichotomous outcome and a singular fit using 2 random effects, we dropped the random parent effect. For all models, we reported conditional odds ratios, which represent the odds of obesity for children with obese versus nonobese parents. We further reported the intra-class correlation coefficient (ICC)²⁰ from the 2 mixed effects models, which represents the proportion of variance that can be explained by the grouping used (by parent in model 2 and by family in model 3).

Statistical analysis was performed in SAS (version 9.4) and R (version 4.1.1 or higher).

RESULTS

Of the 1.5 million unique individuals who had at least one EC in the EDW, we matched or inferred 1 846 036 relationships between 1 157 454 unique individuals, which make up 448 278 families. An additional 74 764 individuals were excluded for conflicting relationships (192 227 conflicting relationships). An individual was only excluded for conflicting relationships if they had no other relationships without conflict. Median family size was 2 with the largest family having 32 members. While not all families spanned multiple generations, 247 families (0.06%) included 4 generations or more (Table 1).

Table 1.

Descriptive statistics of Pythonic RIFTEHR output in Northwestern Medicine Enterprise Data Warehouse (NMEDW)

Variable	N	%
n	1 157 454
Relationships	1 846 036
Provided	841 957	45.61
Opposite	628 287	34.03
Inferred	375 792	20.35
Families	448 278
Family size, median [IQR]	2	[2, 3]
Maximum family size, # persons	32
Maximum family size, # relationships	314
Degree of relationship
First (ie, child, parent)	986 962	53.46
Second (ie, grandparent)	123 762	6.7
Third (ie, great-grandparent)	13 174	0.71
Fourth (ie, great-great-grandchild)	992	0.05
None (ie, spouse, sister-in-law)	653 326	35.39
Unknown (ie, parent/parent-in-law)	67 820	3.67
Number of consecutive generations (by family)
Zero	197 775	44.1
One	221 786	49.5
2	25 500	5.7
3	2970	0.66
4 or more	247	0.06

IQR: interquartile range; RIFTEHR: relationship inference from the electronic health record.

We compared the gold-standard, known mother–child relationships within the EDW, to those found using P-RIFTEHR to calculate the sensitivity of our algorithm. Of the 84 995 mother–child pairings provided in the EDW, 5145 matched pairs were also in the P-RIFTEHR output (Figure 1). Of those, 4894 matching pairs were correctly identified as mother–child, resulting in 95.1% sensitivity. Because not every mother in the NM system gave birth at a NM hospital, we are not able to calculate specificity or positive predictive values.

P-RIFTEHR identified 374 038 parent–child dyads (from 222 283 families) for whom we then pulled additional information from the electronic medical records. 116 363 dyads (from 80 442 families) had complete data (including parent BMI, child BMI, child gender, race, insurance status, age, and Charlson Comorbidity Index) and were included in our proof-of-concept analysis (Table 2). This sample included 93 535 parents and 101 373 unique children. The majority of parents and children were non-Hispanic White (70% for both), and female (69% of parents, 60% of children). Parents were older (mean [SD] age 58 [18] years vs 30 [20] years) and had a higher comorbidity index (3.34) compared to the children (1.06) (Table 2). This subsample is similar to the overall population in the NMEDW, which is 65% White or Caucasian and 71% Non-Hispanic (when excluding those missing race and ethnicity), and majority female (Figure 3).

Table 2.

Characteristics of BMI analysis sample

	Parent	Child
Characteristic	N (%) or mean (SD)	N (%) or mean (SD)
n	93 535	101 373
Female	64 261 (68.70)	61 160 (60.33)
Race/ethnicity
Non-Hispanic White	65 872 (70.42)	71 243 (70.28)
Non-Hispanic Black	6646 (7.11)	7714 (7.61)
Non-Hispanic Asian	3460 (3.70)	3802 (3.75)
Hispanic	9720 (10.39)	11 223 (11.07)
Other/Unknown	7837 (8.38)	7391 (7.29)
Age, years	58.15 (18.19)	29.90 (19.62)
Insurance status
Self-pay/Unknown	19 976 (21.36)	8997 (8.88)
Public	30 850 (32.98)	19 717 (19.45)
Private	42 709 (45.66)	72 659 (71.67)
BMI, kg/m2	28.47 (7.09)	25.69 (8.44)
BMI percentilea	73.88 (27.10)	61.59 (30.31)
Obese	31 843 (34.04)	28 332 (27.95)
Charlson comorbidity index	3.34 (3.39)	1.06 (1.81)

BMI: body mass index; SD: standard deviation.

^aIf <20 years old at time of BMI measurement (N_parent=243, N_child=34 546).

Figure 3.

NMEDW Patient population descriptive statistics. NMEDW: Northwestern Medicine Electronic Data Warehouse.

We estimated the relationship between parent and child obesity to compare our family structures to those of known family cohort studies in 3 separate models (Table 3). In model 1, one parent–child pair per family ID was randomly sampled resulting in a sample size of 80 442 pairs. Model 2 includes all children for one randomly sampled parent per family (80 442 parents, 99 370 dyads). Model 3 uses all available data (116 363 pairs). All models showed that children with obese parents were more than twice as likely to be obese compared with children with parents who were not obese. These associations were further strengthened by adjustment in all 3 models (Table 3). ICC was calculated for models 2 and 3. For model 2, the ICC of 0.23 means that 23% of the variability in child obesity can be explained by clustering the observations by parent. Likewise, the ICC of 0.36 in model 3 can be interpreted as 36% of the variability in child obesity can be explained by clustering the observations by family.

Table 3.

Association between parent obesity and child obesity

Covariates	Model 1: Child obesityOR (95% CI)		Model 2: Child obesityOR (95% CI)		Model 3: Child obesityOR (95% CI)
	N = 80 442 pairs		N = 99 370 pairs		N = 116 363 pairs
	Unadjusted	Adjusted	Unadjusted	Adjusted	Unadjusted	Adjusted
Parent obesity	2.3 (2.23–2.37)	2.53 (2.45–2.62)	2.35 (2.27–2.43)	2.57 (2.49–2.66)	2.22 (2.15–2.30)	2.47 (2.38–2.55)
Child age, years	—	1.03 (1.03–1.03)	—	1.03 (1.03–1.03)	—	1.03 (1.03–1.03)
Child sex	—		—		—
Female		Ref.		Ref.		Ref.
Male		0.95 (0.92–0.99)		0.96 (0.93–0.99)		0.97 (0.94–1.00)
Child insurance	—		—		—
Self-pay/Unknown		Ref.		Ref.		Ref.
Public		1.17 (1.10–1.25)		1.22 (1.14–1.30)		1.23 (1.15–1.31)
Private		1.03 (0.97–1.09)		1.04 (0.98–1.10)		1.03 (0.97–1.09)
Child Race/ethnicity	—		—		—
Non-Hispanic White		Ref.		Ref.		Ref.
Non-Hispanic Black		1.94 (1.83–2.05)		1.96 (1.85–2.08)		2.04 (1.92–2.17)
Non-Hispanic Asian		0.72 (0.65–0.79)		0.71 (0.65–0.79)		0.7 (0.63–0.77)
Hispanic		2.05 (1.95–2.15)		2.11 (2.01–2.22)		2.19 (2.08–2.31)
Other/Unknown		0.98 (0.92–1.05)		1.0 (0.94–1.07)		0.98 (0.91–1.04)
Child Charlson Comorbidity Index	—	0.99 (0.98–1.00)	—	0.99 (0.98–1.00)	—	0.98 (0.97–1.00)
ICC	—	—	0.23	0.19	0.36	0.32

CI: confidence interval; ICC: intra-class correlation; OR: odds ratio.

DISCUSSION

P-RIFTEHR can identify family members in the EHR. It performs as well as its parent algorithm in sensitivity (95.1% for P-RIFTEHR compared to 92.9% and 96.8% for RIFTEHR)⁸ is simpler to execute with fewer dependencies, and runs efficiently. Estimates of intergenerational obesity from family tree structures estimated by the P-RIFTEHR algorithm are consistent with previously published research using family cohorts.

While intergenerational obesity is shown here, this approach can be used to answer other questions with data available in the EHR. EHR data repositories are rich resources, housing a vast array of observations, measurements, lab values, clinician notes, and billing information spanning across time. Previous iterations of this work have been used to model screening rates for celiac disease and the heritability of 500 disease phenotypes.^8,21

There is a large body of literature studying the association between parent and child obesity both in the US and internationally.^{10–14,16,22} Prior studies that have estimated the correlation between parent and child obesity have found the correlation to be significantly different from zero, even when stratifying by or adjusting for race, gender, and often socioeconomic status.^12,13,23 Further, Wang et al performed a meta-analysis of 27 studies that examined the association between parent and child obesity.¹⁰ They found a strong association between parent and child obesity (pooled OR: 2.22; 95% CI, 2.09–2.36), which is compatible with estimates from our study.

The strength of this study is in the NMEDW, a large resource containing a vast number of individuals from across the Chicagoland area seeking care at a large integrated health system. The very nature of the P-RIFTEHR algorithm is a limitation as we cannot confirm that the family members linked by the algorithm are biologically related: many people list close family friends, aunts, or uncles as ECs, and there is no way to disentangle the truth in P-RIFTEHR. Therefore, there are likely spurious relationships that cannot be avoided which may bias results towards the null. However, when utilizing these data for social determinants of health research and as a proxy for shared environment, genetic relationships are not necessary. Changes made to the original RIFTEHR program’s algorithm were designed to increase the accuracy of results from our dataset. However, it is likely that different datasets (ie, different communities) display unique characteristics that would benefit from local customizations in the algorithm. Family structures with unique sizes or characteristics may potentially be reidentified, even if all the other variables in the dataset do not contain any direct or indirect identifiers. Consideration should be taken before family linkage data are made available for additional research within a given institution.

There are significant bioethical concerns about P-RIFTEHR and its parent algorithm, RIFTEHR.⁸ We have addressed these concerns to the best of our abilities in an attached ethical review Supplementary Material S1. At its core, this is an algorithm—its uses are varied and deserve scrutiny. Here, data were used to link individuals based on common identifiers, and after linkages were inferred, individuals were not contacted, inferences were not disclosed, and family inferences were not entered into the medical record. Genetic relationships were not assumed. No associations were researched that have not already been studied using traditional cohorts and relationships were only used for an assessment of disease risk at the population level. Regardless we look forward to participating in the ongoing discussion on the practical and ethical implications of EHR-based research.

CONCLUSIONS

We demonstrated 3 ways to model intergenerational obesity, but future research could span many more topics, delving more deeply into specific risk factors, looking only at mother–child pairs, or modeling the relationship between additional generations beyond parent–child. Algorithm testing in other health care systems, as different populations will invariably have unique characteristics, may yield opportunities for algorithm optimization or future research questions.

In conclusion, P-RIFTEHR estimates familial relationships using EHR data, providing a method to overcome a current limitation of most EHR systems and to strengthen future population health research.

FUNDING

Fernanda Polubriaginof was funded in part through the NIH/NCI Cancer Center Support Grant P30 CA008748.

AUTHOR CONTRIBUTIONS

AK managed development of the P-RIFTEHR algorithm, performed statistical analysis, and prepared the text. FG, DT updated and wrote the P-RIFTEHR algorithm. CW updated and wrote sections of the P-RIFTEHR algorithm and drafted the methods section of the manuscript. AL, MB ran the P-RIFTEHR algorithm in the NMEDW. LP advised on the statistical analysis plan and its execution and provided critical review of the manuscript. FP is responsible for conception of original RIFTEHR algorithm and consulted on its use and updates. DS, FA, NA directed the study’s implementation.

ETHICS APPROVAL

This study was approved by the Northwestern University Institutional Review Board, study number STU00210068.

SUPPLEMENTARY MATERIAL

Supplementary material is available at Journal of the American Medical Informatics Association online.

CONFLICT OF INTEREST STATEMENT

None declared.