Assessment of the Validity of Administrative Data for Gestational Diabetes Ascertainment

Abstract

BACKGROUND:

Administrative data, including International Classification of Diseases codes and birth certificate records, are often used for retrospective gestational diabetes research investigations to describe associations of gestational diabetes with perinatal complications and long-term outcomes, and to determine gestational diabetes prevalence. Research investigating the validity of using International Classification of Diseases codes and birth certificates for gestational diabetes ascertainment shows varying degrees of reliability.

OBJECTIVE:

This study aimed to evaluate the accuracy of both International Classification of Diseases codes and birth certificate diagnosis for gestational diabetes ascertainment in a large hospital-based cohort of pregnant individuals, using laboratory criteria for gestational diabetes mellitus as the reference.

STUDY DESIGN:

We studied individuals who received prenatal care at an academic hospital and affiliated community health centers between 1998 and 2016. In the setting of universal 2-step screening for gestational diabetes, pregnant individuals were classified as having gestational diabetes if ≥2 oral glucose tolerance test values met or exceeded National Diabetes Data Group thresholds. We calculated the sensitivity, specificity, positive predictive value, and negative predictive value for International Classification of Diseases code and birth certificate ascertainment of gestational diabetes, and their exact binomial 95% confidence intervals.

RESULTS:

In a cohort of 51,059 pregnancies with complete glucose screening, 1303 (2.6%) met National Diabetes Data Group laboratory criteria for gestational diabetes. Gestational diabetes International Classification of Diseases codes had moderate sensitivity of 70.5% (95% confidence interval, 67.9−72.9), high specificity of 99.3% (95% confidence interval, 99.3−99.4), a positive predictive value of 73.3% (95% confidence interval, 70.8−75.8), and a negative predictive value of 99.2% (95% confidence interval, 99.1−99.3). In the 46,512 pregnancies linked to birth certificate data, birth certificate diagnosis had moderate sensitivity (66.3% [95% confidence interval, 63.6−69.0]), high specificity (98.9% [95% confidence interval, 98.8−99.0]), moderate positive predictive value (62.1% [95% confidence interval, 59.8−64.4]), and high negative predictive value (99.1% [95% confidence interval, 99.0−99.2]).

CONCLUSION:

Ascertainment of gestational diabetes using administrative data, including International Classification of Diseases codes or birth certificates, has moderate sensitivity, moderate positive predictive value, high specificity, and high negative predictive value. Our findings provide context for interpreting the validity of studies that depend on administrative data for ascertainment of gestational diabetes and comparing them with prospective studies that use laboratory-based gestational diabetes criteria.

Introduction

Gestational diabetes mellitus (GDM), or hyperglycemia first diagnosed in pregnancy, is associated with adverse pregnancy outcomes, long-term risk of chronic disease in affected individuals and their neonates, and higher health care utilization.¹⁻⁵ GDM is diagnosed using oral glucose tolerance tests (OGTTs) performed during routine prenatal care. Prospective observational studies and randomized trials often ascertain GDM status using standard laboratory criteria applied to the results of these OGTTs. In contrast, administrative data, including International Classification of Diseases (ICD) codes and birth certificate records, are often used for retrospective GDM research investigations to describe associations of GDM with perinatal complications⁶⁻⁸ and long-term outcomes,⁹ and to determine GDM prevalence.^10,11 Yet, research investigating the validity of using ICD codes and birth certificates for GDM ascertainment shows varying degrees of reliability.¹²⁻¹⁶

Although previous studies have shown that ICD codes for diabetes mellitus ascertainment outside of pregnancy perform adequately, especially when combined with other clinical data,^17,18 they have performed less well for identifying diagnoses in pregnancy.^16,19,20 Birth certificate data have also been shown to perform imperfectly for ascertainment of diagnoses in pregnancy.^12,15,16 A limitation of previous literature assessing the validity of administrative data for GDM ascertainment is that few studies used gold-standard laboratory results from OGTTs as the GDM reference; those that did had limited sample sizes (<400 cases) or were performed outside the United States, where medical practice and administrative procedures differ.^13,14,21,22

We evaluated the accuracy of both ICD codes and birth certificate diagnosis for GDM ascertainment in a large hospital-based cohort of pregnant individuals, using laboratory criteria for GDM as the reference.²³ We hypothesized that administrative data would have moderate sensitivity and high specificity when compared with the laboratory reference.

Methods

Study population and gestational diabetes mellitus screening

This study included individuals who received prenatal care at an academic medical center and affiliated community health centers between 1998 and 2016. The data were downloaded directly from an obstetrical electronic medical record where values were entered prospectively by clinicians. Additional maternal glucose laboratory values were obtained from our health system’s research data repository for data completion purposes. The repository is a centralized repository for our health system that holds data from hospital clinical systems (including the academic medical center). It contains a diverse collection of data such as diagnostic codes and laboratory results. Approximately 1% of charts (N=547) were reviewed manually to confirm maternal laboratory glucose values after outlier values or missing data were identified. Birth certificate data were obtained and linked to the electronic health record data to investigate the accuracy of this alternative administrative data source for ascertaining GDM status. The health system’s institutional review board (IRB) approved the study and waived the requirement for informed consent. The Massachusetts Department of Public Health IRB approved the linkage and use of the birth certificate data.

Universal screening for GDM was practiced at our academic medical center and affiliated health centers during the period of data collection. Patients underwent a GDM screening test using a nonfasting 50-g glucose loading test (GLT) at approximately 28 weeks’ gestation followed by a diagnostic fasting 100-g 3-hour OGTT if the GLT result was ≥140 mg/dL. Pregnancies were classified as having GDM if ≥2 OGTT values met or exceeded thresholds. Our analyses used National Diabetes Data Group (NDDG) criteria²³ (thresholds for abnormal: fasting, ≥105 mg/dL; 1 hour, ≥190 mg/dL; 2 hours, ≥165 mg/ dL; 3 hours, ≥145 mg/dL) that were in clinical use at the medical center during the study period. In a sensitivity analysis, we used the Carpenter−Coustan (CC)²³ criteria (thresholds for abnormal: fasting, ≥95 mg/dL; 1 hour, ≥180 mg/dL; 2 hours, ≥155 mg/dL; 3 hours ≥140 mg/dL) to define GDM.

Individuals who did not have GLT data or whose GLT was performed before 22 weeks’ gestation were excluded from the analysis. Because patients known to have preexisting diabetes mellitus were not knowingly screened for GDM, their pregnancies were not included. If an individual had an abnormal GLT and did not have an OGTT or had an incomplete OGTT, they were considered to have incomplete screening and GDM status was not defined, leading to exclusion from our analyses. Review of a subset of medical charts showed that reasons for not completing the OGTT included transfer of care out of the hospital’s obstetrical program, inability to tolerate or declining the OGTT, or having a GLT result of ≥200 mg/dL (in which case patients were generally treated as having GDM without OGTT confirmation).

International Classification of Diseases codes

We obtained ICD codes from the health system’s research data repository from the date of the first prenatal visit up to the delivery date for each pregnancy. The primary analysis included abnormal glucose tolerance of the mother complicating pregnancy, childbirth, or the puerperium (648.8) from ICD-9 and GDM (O24.4) from ICD-10, recorded in administrative data from inpatient and outpatient encounters. A secondary analysis included additional codes and their subcodes that represent all possible diabetes mellitus and abnormal glucose diagnoses. From ICD-9, this included diabetes mellitus (250), diabetes mellitus complicating pregnancy, childbirth, or the puerperium (648.0), and abnormal glucose (790.2). For ICD-10, we included all types of preexisting diabetes mellitus in pregnancy, childbirth, and the puerperium (O24.0, O24.1, O24.3, O24.8), and all types of diabetes mellitus (E08-E13), abnormal glucose (R73.0), and hyperglycemia (R73.9).

Birth certificate data

We obtained birth certificate records through the state’s Department of Public Health and linked them to the medical center’s clinical records. Pregnancies and births were linked by medical record number and validated by matching data from our health system’s research data repository and clinical records.

Additional data

Maternal age, body mass index (BMI), race and ethnicity, and infant birthweight were obtained from the obstetrical electronic medical record. Additional race and ethnicity information from our health system’s research data repository and birth certificate records were incorporated to fill in missing data. Medication records were pulled from a research data repository query for a subset of participants who received care after 2007. Large-for-gestational-age birthweight was defined as birthweight >90th birthweight percentile for gestational age.²⁴

Analyses

We described characteristics and pregnancy outcomes of individuals who were classified as having GDM on the basis of NDDG laboratory criteria, ICD code, or birth certificate. We separately examined individuals who had discordant GDM status between ICD code and laboratory data. Characteristics are presented as mean±standard deviation for continuous variables and number (percent-age) for categorical variables.

For both ICD and birth certificate records, we calculated the sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) for ICD code−based ascertainment of GDM, along with their exact binomial 95% confidence intervals (CIs).¹⁹ NDDG criteria were used as the primary reference method for GDM classification, and CC criteria were used in a sensitivity analysis.

Differences in characteristics between pregnancies identified by ICD code without meeting NDDG criteria and pregnancies meeting NDDG laboratory criteria were compared using the Pearson chi-square test with Yates’ continuity correction for the categorical variables and the Welch 2-sample t-test for the continuous variables.

Statistical analyses were performed using R (version 3.6.1; R Core Team, Vienna, Austria) and SAS (version 9.4; SAS Institute, Cary, NC).

Chart review to explore gestational diabetes mellitus ascertainment discrepancies

A subset of charts was reviewed by a physician author to investigate discrepancies between ICD codes and laboratory values. From the pregnancies that were identified with a GDM ICD code but did not meet NDDG criteria, 50 random charts were selected and reviewed. From pregnancies that met NDDG criteria but did not have a GDM ICD code, another 50 random charts were selected and reviewed.

Results

Participant characteristics

After removing pregnancies with incomplete glucose screening data from the 65,955 pregnancies in the population, a cohort of 51,059 (77.4%) pregnancies was used for the ICD code analysis (Supplemental Figure 1). Characteristics of pregnancies with complete maternal glucose screening included in the analysis were similar to those of the excluded pregnancies with incomplete screening (Supplemental Table 1). In the ICD code analysis cohort, 1303 (2.6%) pregnancies met NDDG criteria according to GLT and OGTT values. In the same cohort, 1252 (2.5%) had GDM ICD codes, and 1854 (3.6%) pregnancies had at least 1 of the ICD codes on the expanded list including all abnormal glucose diagnoses (Supplemental Table 2). ICD codes indicating preexisting type 1 and type 2 diabetes mellitus (O24.0, O24.1, O24.3, O24.8, E08, E09, E10, E11) were not represented in the cohort (Supplemental Table 2).

After removing records with discrepancies in mother’s date of birth and infant sex, 53,802 neonates from 52,405 pregnancies were linked to birth certificate data (Supplemental Figure 1). From these linked records, a final cohort of 46,512 pregnancies with fully screened glucose values was used in our birth certificate analysis. In the birth certificate cohort, 1205 (2.6%) pregnancies met NDDG criteria according to GLT and OGTT values and 1286 (2.8%) were identified with GDM using birth certificates.

Compared with pregnancies that did not meet NDDG laboratory criteria for GDM, pregnancies that did meet these criteria occurred in individuals who were older, had higher BMI, were less likely to have private health insurance, and were less likely to identify as non-Hispanic White (Table 1). These pregnancies also had a higher proportion of large-for-gestational-age infants (Table 1).

TABLE 1.

Characteristics of study cohort by gestational diabetes mellitus status

Characteristic	No GDM by laboratory criteriaa (N=49,756)	GDM by laboratory criteriaa (N=1303)	GDM by ICD code (N=1252)	GDM by birth certificate (N=1286)
	Mean±SD or N (%)
Age (y)	30.8±5.9	32.7±5.5	32.7±5.4	32.3±5.5
Body mass index (kg/m2)	25.5±5.3	28.7±6.4	28.8 ±6.5	28.8±6.4
Race/ethnicityb
Hispanic/Latina	6161 (12.5%)	186 (14.3%)	157 (12.5%)	197 (15.4%)
Non-Hispanic Asian	4254 (8.6%)	202 (15.5%)	185 (14.8%)	183 (14.3%)
Non-Hispanic Black	3219 (6.5%)	114(8.7%)	114(9.1%)	133 (10.4%)
Non-Hispanic White	29,901 (60.7%)	575(44.1%)	591 (47.2%)	556 (43.3%)
Multiracial/none of the above	5709 (11.6%)	210(16.3%)	194 (15.6%)	213(16.6%)
Insurance type
Private	32,120 (64.6%)	719(55.2%)	723 (57.7%)	712 (55.4%)
Public	12,778 (25.7%)	441 (33.8%)	400(31.9%)	437 (34.0%)
Limited/none	4858 (9.8%)	143 (10.9%)	129 (10.3%)	137 (10.7%)
Marital statusc
Married/partner	35,049 (70.4%)	909 (69.8%)	879 (70.2%)	887 (69.0%)
Singled	15,137(28.4%)	422 (29.4%)	352 (29.0%)	387 (30.1%)
Large-for-gestational-age birthweighte	3895 (8.1%)	178 (13.9%)	186 (14.9%)	180 (14.0%)

GDM, gestational diabetes mellitus; ICD, International Classification of Diseases; SD, standard deviation.

^aNational Diabetes Data Group (NDDG) criteria were used as the laboratory reference, given that they were in clinical use during the study period;

^bFor race/ethnicity, there are 7 missing from those with no GDM by laboratory criteria, 1 missing from those with GDM by laboratory criteria, 0 missing from those with GDM by ICD code, and 1 missing from those with GDM by birth certificate;

^cFor marital status, there are 570 other/unknown from those with no GDM by laboratory criteria, 12 other/unknown from those with GDM by laboratory criteria, 11 other/unknown from those with GDM by ICD code, and 12 other/unknown from those with GDM by birth certificate;

^dSingle category includes single, divorced, separated, and widowed;

^eDefined as >90th birthweight percentile from Oken et al.²⁴

International Classification of Diseases code analysis with laboratory criteria reference

Using NDDG laboratory criteria as the reference, we observed that GDM ICD codes had moderate sensitivity of 70.5% (95% CI, 67.9−72.9) and high specificity of 99.3% (95% CI, 99.3−99.4) (Figure; Table 2). This resulted in a PPV of 73.3% (95% CI, 70.8−75.8) and an NPV of 99.2% (95% CI, 99.1−99.3). Thus, 29.5% of cases of GDM by NDDG criteria were missed by ICD codes, and among the 1252 pregnancies coded with a GDM ICD code, 26.7% did not have GDM according to NDDG criteria. Among the 49,807 pregnancies without any GDM ICD code, only 0.77% met NDDG criteria. When limiting to only GDM ICD codes from the delivery date, the sensitivity was much lower at 31.0% (95% CI, 33.6−28.5), with similar specificity (99.8% [95% CI, 99.8−99.9]), PPV (81.6% [95% CI, 77.9 −84.9]), and NPV (98.2% [95% CI, 98.1−98.3]).

FIGURE.

Performance of GDM administrative data ascertainment compared with laboratory criteria

GDM ascertained by ICD code (black) shows moderate sensitivity (70.5%), high specificity (99.3%), moderate PPV (73.3%), and high NPV (99.2%). GDM by birth certificates (light gray) shows a similar trend, with slightly lower sensitivity (66.3%), high specificity (98.9%), slightly lower PPV (62.1%), and high NPV (99.1%).

GDM, gestational diabetes mellitus; ICD, International Classification of Diseases; NPV, negative predicive value; PPV, positive predictive value.

TABLE 2.

Gestational diabetes mellitus diagnosis by administrative data vs laboratory criteria

ICD code analysis
	GDM by NDDG criteria	No GDM by NDDG criteria
	N (%)
N	1303	49,756
GDM ICD code (ICD-9: 648.8, ICD-10: O24.4)	918 (70.5%)	334 (0.67%)
No GDM ICD code	385 (29.5%)	49,422 (99.3%)
Birth certificate analysis
	GDM by NDDG criteria	No GDM by NDDG criteria
	N (%)
N	1205	45,307
GDM by birth certificate	799 (66.3%)	487 (1.1%)
No GDM by birth certificate	406 (33.7%)	44,820 (98.9%)

GDM, gestational diabetes mellitus; ICD, International Classification of Diseases; NDDG, National Diabetes Data Group.

Characteristics of individuals with GDM ascertained by ICD code were similar to those of individuals with GDM ascertained by NDDG laboratory criteria (Table 1).

With the NDDG laboratory criteria as the reference, the expanded list of diabetes mellitus ICD codes (Supplemental Table 2) had a sensitivity that was slightly higher (74.3% [95% CI, 71.8−76.6]) than that of the GDM-specific codes and a similar specificity (99.0% [95% CI, 98.8−99.1]) (Supplemental Figure 2). The PPV was 65.5% (95% CI, 63.1−68.0), lower than that of the GDM-specific codes, and the NPV was similar (99.3% [95% CI, 99.2 −99.4]) (Figure).

Using CC criteria as the reference, 2097 (4.1%) participants had GDM. The sensitivity of GDM-specific diagnostic codes for CC GDM was lower than that achieved using the NDDG criteria reference (46.9% [95% CI, 44.7 −49.0]). Specificity was slightly higher (99.5% [95% CI, 99.4−99.5]). This resulted in higher PPV (78.6% [95% CI, 76.2−80.8]) and slightly lower NPV (97.8% [95% CI, 97.6−97.9]) than that observed using the NDDG reference.

Birth certificate analysis with laboratory criteria reference

In the 46,512 pregnancies linked to birth certificate data, 1205 (2.6%) met NDDG criteria. Birth certificates identified 1286 pregnancies (2.8%) as having GDM. Similar to the findings from the primary analysis, using NDDG as the reference, birth certificates had moderate sensitivity (66.3% [95% CI, 63.6−69.0]), high specificity (98.9% [95% CI, 98.8−99.0]), moderate PPV (62.1% [95% CI, 59.8−64.4]), and high NPV (99.1% [95% CI, 99.0−99.2]) (Figure; Table 2). However, the sensitivity and PPV were lower with birth certificates than with ICD codes. Thus, 33.7% of the pregnancies that met NDDG criteria were not identified from the birth certificate data as having GDM, and 37.9% of pregnancies that had GDM listed on the birth certificate did not meet laboratory criteria.

Analysis of pregnancies with discordant gestational diabetes mellitus ascertainment

Among the 1303 pregnancies that met NDDG criteria, 385 pregnancies (29.5%) were not coded with a GDM ICD code (Table 2). These pregnancies had similar characteristics to those of pregnancies that met NDDG criteria (Table 3), with the most notable differences being lower GLT and OGTT glucose values and a lower proportion treated with GDM medications (Table 3). We randomly selected 50 individuals who had NDDG GDM without a GDM ICD code for chart review. Of the 52 pregnancies occurring in these individuals, 51 (98.1%) had confirmed and treated GDM, affirming that glucose laboratory values identified true cases that were missed using ICD code ascertainment.

TABLE 3.

Comparison of characteristics of individuals with gestational diabetes mellitus ascertained by different methods

Characteristic	GDM by laboratory criteriaa (N=1303) Mean±SD or N (%)	GDM by ICD code (N=1252) Mean±SD or N (%)	GDM by laboratory criteriaa but no ICD code (N=385) Mean±SD or N (%)	GDM by ICD code but not laboratory criteriaa (N=334) Mean±SD or N (%)	P value
Age (y)	32.7±5.5	32.7±5.4	32.6±5.5	32.6±5.1	.94
Body mass index (kg/m2)	28.7±6.4	28.8±6.5	28.3±6.4	28.3±6.6	.27
Race/ethnicityb					.002
Hispanic/Latina	186(14.3%)	157 (12.5%)	56 (14.5%)	27(8.1%)
Non-Hispanic Asian	202 (15.5%)	185 (14.8%)	61 (15.8%)	44 (13.2%)
Non-Hispanic Black	114(8.7%)	114(9.1%)	28 (7.3%)	29 (8.7%)
Non-Hispanic White	575(44.1%)	591 (47.2%)	174(45.2%)	190(56.9%)
Multiracial/none of the above	5709 (11.6%)	210(16.3%)	58 (15.3%)	42 (12.7%)
Insurance type					.03
Private	719(55.2%)	723 (57.7%)	211 (54.8%)	215 (64.4%)
Public	441 (33.8%)	400 (31.9%)	131 (34.0%)	90 (26.9%)
Limited/none	143 (10.9%)	129 (10.3%)	43 (11.2%)	29 (8.7%)
Marital statusc					.76
Married/partner	909 (69.8%)	879 (70.2%)	268 (69.6%)	238 (71.3%)
Singled	422 (29.4%)	352 (29.0%)	114(29.6%)	94(28.1%)
Large for gestational agee	178(13.9%)	186 (14.9%)	37 (9.6%)	45 (13.5%)	.89
Glucose loading testf	168.4±12.1	161.7±27.7	165.2±18.2	140.3±29.5	<.001
Oral glucose tolerance test results
Fasting glucose	94.2±16.9	93.3±17.0	91.9±15.2	85.5±11.8	<.001
1-h glucose	207.4±26.9	201.1 ±33.1	203.8±25.7	168.3±35.2	<.001
2-h glucose	190.1 ±27.8	181.5±35.2	187.1 ±25.9	140.8±30.5	<.001
3-h glucose	130.7±37.5	125.9±38.8	128.5±35.6	101.7±30.8	<.001
GDM medications (post-2007)	195 (33.6%)	201 (35.4%)	21 (12.7%)	27 (17.8%)	.001

GDM, gestational diabetes mellitus; ICD, International Classification of Diseases; SD, standard deviation.

p-value compares characteristics to those with GDM by laboratory criteria

^aNational Diabetes Data Group criteria were used as the laboratory reference, given that they were in clinical use during the study period;

^bFor race/ethnicity, there are 7 missing from those with no GDM by laboratory criteria, 1 missing from those with GDM by laboratory criteria, 1 missing from those with GDM by laboratory criteria with no ICD code, and 0 missing from those with GDM by ICD code without meeting laboratory criteria;

^cFor marital status, there are 12 other/unknown from those with GDM by laboratory criteria, 11 other/unknown from those with GDM by ICD code, 3 other/unknown from those with GDM by laboratory criteria with no ICD code, and 2 other/unknown from those with GDM by ICD code without meeting laboratory criteria;

^dSingle category includes single, divorced, separated, and widowed;

^eDefined as >90th birthweight percentile from Oken et al²⁴;

^fNonfasting 50-g glucose loading test taken between 24 and 28 weeks of gestation.

Among the 48,962 pregnancies that did not meet NDDG criteria, there were 334 pregnancies (0.67%) that were given a GDM ICD code. Compared with all pregnancies that met NDDG criteria, these pregnancies were more likely to be in individuals with private insurance and who identified as non-Hispanic White (Table 3). As expected, these pregnancies had lower GLT and OGTT glucose values compared with pregnancies that met NDDG criteria; accordingly, a lower percentage of these pregnancies were treated with GDM medications (Table 3). There were 66 (19.8%) pregnancies with a GDM ICD code that did not meet NDDG criteria for GDM but met the CC criteria for GDM.¹⁷ We randomly selected 50 individuals for chart review from the 334 pregnancies that had a GDM ICD code without meeting NDDG criteria. We found that ICD coding captured some clinician diagnoses and treatment of GDM that could not be verified by laboratory glucose values (n=13; 26%). Specific examples include GDM diagnosed outside of our hospital system, patients treated as having GDM despite a normal OGTT because glucose values were close to meeting NDDG criteria, or patients treated as having GDM on the basis of a history of GDM in a previous pregnancy when not meeting laboratory criteria during the current pregnancy. There were 2 pregnancies (4%) in which glucosuria was noted in the third trimester after normal GDM screening; these patients were subsequently treated for GDM. There were 7 pregnancies with ICD codes (14%) that were noted to have a history of GDM but were not treated for GDM in the relevant pregnancy. Other pregnancies had some abnormal glucose laboratory values (usually a GLT ≥140 mg/dL) but were not treated as having GDM (n=13; 16%). In addition, 15 pregnancies (30%) did not have any evidence of GDM or abnormal glucose testing on chart review, despite having a GDM ICD code.

Comment

Principal findings

In this retrospective analysis of 51,059 pregnancies from an academic medical center, we evaluated the performance of diagnosis codes from ICD-9 and ICD-10 for ascertainment of GDM by comparison with gold-standard laboratory criteria. We found that ICD codes had moderate sensitivity (70.5%) and moderate PPV (73.3%), with high specificity (99.3%) and NPV (99.2%) for ascertainment of GDM diagnosed by NDDG laboratory criteria that were in clinical use during the study period. In comparison with ICD codes, birth certificates had a slightly lower sensitivity (66.3%) and PPV (62.1%) for ascertaining GDM but retained a high specificity (98.9%) and NPV (99.1%). Pregnancies ascertained as having GDM by laboratory criteria, ICD codes, and birth certificates shared many characteristics. However, pregnancies that were given a GDM ICD code without meeting NDDG laboratory criteria were more likely to be in individuals with private insurance or of non-Hispanic White race. Our findings provide context for future United States −based studies that rely on administrative data for GDM ascertainment. Our results may help to explain potential differences in findings between prospective studies of GDM that use laboratory-based definitions and retrospective studies of GDM that rely on administrative data for ascertainment.

Results in the context of what is known

Previous smaller studies demonstrated inaccuracies in GDM ascertainment when relying on administrative data alone. Nicklas et al¹³ evaluated the performance discharge codes from delivery records at another academic medical center for ascertainment of GDM, using chart review as the reference. This study examined cases identified as having GDM, and unlike the present study, did not evaluate whether the absence of an administrative code could be used to exclude a GDM diagnosis. In this previous study, ICD codes were indicative of clinician diagnoses in 98% of 371 identified cases, but 36% of these cases did not meet laboratory criteria for GDM despite the presence of a clinical diagnosis (PPV, ~64%).¹³ Similar to this study, we found that 27% of the pregnancies with a GDM ICD code did not meet criteria according to laboratory values, corresponding to a PPV of 73%; our study was additionally able to assess the sensitivity, specificity, and NPV of administrative codes for GDM. A systematic review investigating the accuracy of birth certificate data for identifying GDM found that the sensitivities ranged from 46% to 83%, with specificities consistently >98%.¹² These results are also consistent with the findings from our analyses using birth certificate data. Similar to the present study, these investigators found that the hospital data performed better than birth certificates.¹⁶

In contrast to our results, a smaller study examining a total of 3654 deliveries with medical record and billing code information at a different academic medical center from 2016 to 2018 found that ICD-10 diagnosis codes performed well for identifying GDM, with high sensitivity (94.7% [95% CI, 91.5−97.9]) and specificity (99.1% [95% CI, 98.8 −99.4]), moderately high PPV (86% [95% CI, 81−90]), and high NPV (99.7% [95% CI, 99.5−99.9]).¹⁴ However, unlike our study, in which we used strict laboratory values as the reference method for ascertaining GDM, this study had a more liberal definition of GDM that included, for example, a chart diagnosis when some laboratory values were missing and diagnosis based on home self−blood glucose monitoring.¹⁴ Another study examining GDM ICD codes using laboratory values as the reference in 58,338 women in Canada also found that ICD-10 diagnosis codes performed well for identifying GDM, with high sensitivity (92% CI, 91 −93) and specificity (97% CI, 97−98), and moderate PPV (57.0% CI, 55 −59).²² However, this study was performed outside the United States, reflecting different billing and coding practices for GDM from those used in the United States. These differences in medical practice environment, diagnostic criteria, and methodology may account for the difference in our results from these previous studies.

Clinical implications

Retrospective studies depend on administrative data to ascertain GDM cases,⁶⁻¹¹ particularly when a large sample size is necessary. Obtaining accurate GDM ascertainment in studies is clinically relevant, as these large retrospective studies assess adverse pregnancy outcomes and long-term maternal risk of diabetes mellitus and cardiometabolic syndromes associated with GDM during pregnancy. Our findings show the strengths of ascertainment of GDM by leveraging objective laboratory data from electronic medical records in the context of near-universal GDM screening in the United States.

In addition, previous studies suggest that other conditions in pregnancy are prone to misclassification by administrative data.^16,19,20 In the United States, reimbursement may be less tied to diagnostic coding in the obstetrical environment (compared with outside of pregnancy) because of billing occurring through global maternity packages, which provide a one-time fee for all prenatal and inpatient maternity care. Studies that examined preeclampsia and eclampsia,¹⁹ maternal obesity,¹⁶ and obstetrical complications (trauma, lacerations, and infection)²⁰ found suboptimal performance of ICD codes, with low sensitivity (15.0%−16.2%), high specificity (98.7%−99.0%), and moderate PPV (54.4%−75.2%) and NPV (57.6%−93.9%). Our findings suggest that the sensitivity and PPV of ICD codes are higher for GDM than for some other pregnancy diagnoses.^16,19,20

Research implications

Our findings provide context for interpreting the validity of large retrospective studies on GDM in the United States and comparing them with prospective studies that use laboratory-based GDM definitions. Using ICD codes or birth certificate data for GDM ascertainment had moderate sensitivity. This could lead to an unreliable estimation of GDM prevalence if administrative data are used to ascertain the diagnosis. Although specificity of ICD codes for GDM diagnosis was high, approximately 1 in 4 cases identified by a GDM ICD code could not be confirmed with the laboratory values; approximately one-third of these cases had clinician diagnoses of GDM. GDM cases ascertained via administrative data had many similar characteristics to those of cases that met NDDG criteria, but there was enrichment for private insurance and White race in cases given an ICD code that did not meet laboratory definitions.

Strengths and limitations

Strengths of our study include the size of our cohort, availability of both ICD code and linked birth certificate administrative data in the same individuals, and the availability of laboratory data to use as the GDM ascertainment reference, analogous to what is typically used in prospective studies. In addition, because this was a hospital-based cohort, pregnancies had medical charts available for further validation and explanation of findings. Our study had several limitations. First, the study population was from a single medical center, which could limit the broad application of these results to other centers because coding and GDM screening practices could vary at different locations. Despite this, our results are largely consistent with previous studies conducted in other clinical settings. Second, at the time of the data collection, the NDDG criteria were used at our medical center for diagnosis of GDM, but currently CC criteria are more commonly used; thus, the results may not be completely transferable to studies that use more recently collected health data. As a result of using the NDDG criteria, the prevalence of GDM in our sample was lower than would be expected when using CC criteria. In addition, ICD-9 was used during most of the study period, whereas ICD-10 is used currently. Because there was a transition to ICD-10 in 2015, some data may have been disrupted in that year. Finally, we only included pregnancies with complete glucose screening data (77.4%), which means that GDM may have been incompletely ascertained, but the excluded and included participants seemed to have similar characteristics.

Conclusions

These findings offer insight on the interpretation of the validity of large retrospective studies on GDM performed in the United States. GDM ascertainment using administrative data had moderate sensitivity, which could lead to an unreliable estimation of the prevalence. GDM cases ascertained by ICD codes had enrichment for private insurance and White race, reflecting clinician diagnoses that did not correspond with strict laboratory-based definitions. Given the near-universal GDM screening in the United States, future research should consider ascertainment of GDM by using objective laboratory data from electronic medical records.

AJOG MFM at a Glance.

Why was this study conducted?

This study was conducted to evaluate the accuracy of both International Classification of Diseases (ICD) codes and birth certificate diagnosis in gestational diabetes mellitus (GDM) ascertainment in a large hospital-based cohort of pregnant individuals, using GDM laboratory criteria as the reference.

Key findings

Ascertainment of GDM using administrative data, including ICD codes or birth certificates, has moderate sensitivity, moderate positive predictive value, high specificity, and high negative predictive value.

What does this add to what is known?

Previous research investigating the validity of using administrative data for GDM ascertainment did not use laboratory values as the reference, used a limited number of cases, or was conducted outside of the context of the US medical system. Our findings provide context for interpreting the validity of studies that depend on administrative data and comparing them with prospective studies that use laboratory-based GDM definitions.