Abstract
Introduction
Early birth is often recommended for “poorly controlled” diabetes; however, no guidelines define the glycemic threshold that necessitates delivery. We use natural language processing (NLP) of electronic health records to identify individuals described by healthcare professionals as having “poor glucose control” and to examine the factors and outcomes associated with this categorization
Research design and methods
We completed a retrospective cohort study of pregnant individuals with pre-existing and gestational diabetes mellitus from 2018 to 2019. NLP identified prespecified terms indicating “poor glucose control” in clinical notes, and a cohort analysis compared those with and without “poor glucose control” language. Clinical characteristics, objective glucose measures, and neonatal and maternal outcomes were statistically compared.
Results
1433 individuals met inclusion criteria, and 143 (10%) were described as having “poor glycemic control.” After adjusting for diabetes type, pregnant individuals of color (adjusted OR (aOR) 2.4, 95% CI 1.63 to 3.57, p<0.001), individuals on public insurance (aOR 3.22, 95% CI 2.2 to 4.74, p<0.001), and non-English/non-Spanish speaking individuals (aOR 2.07, 95% CI 1.22 to 3.4, p=0.005) had higher odds of being categorized as having “poor glucose control” than control groups. This designation was often applied in the absence of objective markers of glycemia. While some individuals categorized with “poor glucose control” experienced earlier births and higher rates of neonatal complications, these differences were less pronounced when comparing individuals with A1c≤6.5%.
Conclusions
Pregnant individuals of color, those on public insurance, and non-English/non-Spanish speakers are more likely to be categorized as having “poor glycemic control.” Little objective data supported this categorization.
Keywords: Type 2 Diabetes; Diabetes, Gestational; Patient-Centered Care; Pregnancy
WHAT IS ALREADY KNOWN ON THIS TOPIC.
WHAT THIS STUDY ADDS
Approximately 10% of individuals with diabetes are described by providers as having “poor glucose control.”
Individuals from marginalized communities are more likely to be classified by healthcare professionals as having “poor glucose control” despite similar glycemic measures.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
This study highlights differences in how clinicians identify individuals with “poor glucose control” and identifies a need for further clinical guidelines to limit potential bias in categorization.
Introduction
Pre-existing and gestational diabetes mellitus (GDM) are significant comorbidities affecting approximately 0.9% and 6.0% of pregnancies in the USA, respectively.1 Individuals with perinatal diabetes are at an increased risk of several adverse maternal and neonatal outcomes.2 3 Recommended birth timing is based on diabetes type, with birth typically recommended between 39 weeks 0 days and 39 weeks 6 days for type 1, type 2, and medication-requiring GDM.4 5 However, for those with “poorly controlled” pre-existing diabetes, healthcare professionals can recommend birth between 36 weeks 0 days and 38 weeks 6 days.4 For those with “poorly controlled” medication-requiring GDM, birth can be considered between 37 weeks 0 days and 38 weeks 6 days.5
Multiple consensus statements6,8 recommend avoiding terms such as “poor glucose control” due to their association with increased distress among individuals with diabetes.9 Nevertheless, this language remains common in obstetric practice and continues to influence clinical decision-making, despite the absence of clear guidance on the degree of hyperglycemia or related clinical factors that justify early birth. In the absence of clear guidance, healthcare professionals often rely on their clinical experience and discretion to guide classification and birth timing decisions. Objective measurements such as hemoglobin A1c, self-monitored glucose values, and ultrasound findings associated with diabetes may be used to aid decision-making.10 However, without specific guidelines, diagnosis of “poorly controlled” glucose and subsequent delivery recommendations may be influenced by the potential biases of individual practitioners.
To identify the existence of subjective and objective factors associated with healthcare professional-determined categorization of “poor glucose control”, we use natural language processing (NLP) and clinical notes within the electronic health records (EHR) to examine demographic factors, objective clinical measures, and perinatal outcomes associated with “poor glucose control” classification. We hypothesized that healthcare professionals categorize individuals as having “poor glucose control” based on limited objective data and that individuals from historically marginalized racial/ethnic and low-income communities are more often categorized as having “poorly controlled diabetes.” Additionally, we hypothesized that individuals described as having “poor glucose control” would have a higher rate of adverse perinatal outcomes.
Research design and methods
Study setting
We performed a retrospective cohort study of individuals with perinatal diabetes who gave birth between January 2018 and December 2019 within the MHealth-Fairview system and reported using the Strengthening the Reporting of Observational Studies in Epidemiology guidelines.11 This time frame follows changes in American College of Obstetricians and Gynecologists (ACOG) practice guidelines,4 5 publication of consensus statements on language in diabetes,6 7 12 and before the COVID-19 pandemic, which we hypothesized would impact care delivery. MHealth-Fairview is an academic-community health system partnership encompassing a tertiary care academic hospital in an urban setting and six satellite hospitals throughout Minnesota. Individuals with diabetes in pregnancy were identified using the University of Minnesota Obstetrics Measures (UMOMs) Database,10 which is a clinical outcomes database containing over 100 000 births by over 73 000 pregnant individuals from 2011 to 2022 that links to data found within the EHR (University of Minnesota IRB approval #STUDY00012822). The EHR includes clinical notes, flow sheets, and scanned prenatal records from outpatient offices.
Population
Pregnant individuals with pre-existing diabetes (type 1 diabetes mellitus (T1DM) and T2DM) or GDM in an index singleton pregnancy were included. The International Classification of Diseases (ICD), 9th and 10th revision codes were used to identify individuals diagnosed with diabetes during pregnancy (online supplemental table 1). Pre-existing diabetes was diagnosed before pregnancy or by standard criteria prior to 15 weeks’ gestation,13 while GDM was diagnosed using a two-step process in alignment with ACOG recommendations.5 Individuals who did not require pharmacologic therapies (ie, insulin, metformin, or glyburide) for glucose management were categorized with diet-managed (GDMA1) diabetes. Those who required pharmacologic therapies were classified with medication-managed (GDMA2) diabetes. Manual verification of individuals’ classification was conducted by reviewing diagnoses indicated in documentation during admission for birth.
Individuals without diabetes, with multifetal gestation, and with fetal anomalies or known chromosomal abnormalities were excluded. Fetal anomalies were defined as prenatal congenital anomalies that resulted in neonatal intensive care unit (NICU) admission or postnatal surgical intervention.
Identifying cohorts
Using NLP-patient information extraction for research, a scalable, extensible, and secure system for processing, indexing, and searching clinical notes at the University of Minnesota, we identified clinician notes in the EHR containing prespecified terms suggestive of “poor glucose control.”14 We assessed for the presence of prespecified terms thought to reflect clinical assessment of individual glucose levels in the 40 weeks prior to birth. Specifically, we identified the phrases “poor glycemic control,” “poorly controlled glucose,” and “poor control of diabetes” along with “poor glucose control” variations and synonyms, including: “poor” glucose control, “bad” glucose control, “suboptimal” glucose control, “poorly controlled” glucose, and “uncontrolled” diabetes.
Once identified by the NLP algorithm, prespecified terms alongside the 50 preceding and following characters were extracted (henceforth termed “snippets”). These snippets, identifiers, and note details (ie, provider type, date of entry) were compiled into a separate database. The NLP algorithm then reanalyzed the initial extraction for false positive documents (ie, snippets pulled but confirmed by the algorithm not to contain prespecified terms). The remaining snippets then underwent a classification stage. For snippets containing multiple specified terms, each term was classified and analyzed separately (online supplemental figure 1).
Once processed, two investigators (SAW and AG) reviewed the remaining snippets to define the context of language use. For the term “poorly controlled diabetes,” several versions were identified from phrases including “in the setting of” or “for women with.” These were excluded from the analysis as they were found to represent templated phrasing in the counseling portion of the documentation. A similar process was undertaken for “uncontrolled diabetes” snippets found in nutritional screening documentation, which were found to be templated notes used by diabetes educators within the health system. These terms in these contexts were excluded since they were not thought to represent clinican assessments of a specific individual’s glucose levels.
Individuals were divided into two cohorts: those “with poor glucose language” and those “without poor glucose language” documented in their medical records. Individuals with any variant of poor glucose language in EHR were included in the “Individuals with ‘poor glucose control’ language” cohort. All other individuals were categorized into the “Individuals without ‘poor glucose control’ language” cohort.
Demographic and clinical outcomes data
Demographic data for pregnant individuals with diabetes were extracted from the EHR. Race, language, and insurance status were provided by individuals during their most recent healthcare encounter. If no race or language was identified, the metrics for these individuals were categorized as missing. Insurance status was categorized as public (provided by state or federal government agencies), private (provided by private entities), or uninsured. If no insurance status was identified, individuals were classified as uninsured. Prepregnancy body mass index (BMI) (kg/m2) was determined as the last recorded BMI before the first documentation of an individual’s pregnancy. Parity was based on the obstetric history provided during the index pregnancy. Manual chart review was completed to try to identify any missing clinical data (ie, BMI or birth weight). We report missing values remaining despite this review in the tables.
The type/specialty of healthcare professionals using “poor glucose control” language was summarized. Healthcare professional types included physicians, mid-level practitioners (certified nurse midwives, physician assistants, and advanced practice professionals), diabetes educators, and others. Physicians included licensed and trainee practitioners with Doctor of Medicine and Doctor of Osteopathic Medicine degrees. Mid-level healthcare professionals included healthcare workers with postsecondary school training, including certified midwives, nurse practitioners, and physicians’ assistants. Within this population’s healthcare system, diabetes educators are registered dietitians and nurses with advanced education in diabetes management who serve as the primary point of contact for initiation of diabetes management in pregnant individuals.
Data on glucose levels 2 weeks before birth were identified in clinical notes, scanned logs, continuous glucose monitor data copied to the EHR, and messages to clinicians. We identified abnormal glucose levels for fasting (>95 mg/dL), 1-hour postprandial (>140 mg/dL), and 2-hour postprandial values (>120 mg/dL). Individuals with at least 50% glucose levels elevated above the normal range 2 weeks before birth were categorized as having abnormal glucose levels. This was based on institutional policies for initiating or titrating insulin in pregnancy and aligns with our prior studies investigating early birth for “poor glucose control” indications.10 Cases in which glucose ranges were reported in the EHR but the percentage of abnormal values could not be specifically calculated were reported as “not able to be assessed.” Hemoglobin A1c was identified from the clinical record if obtained within 3 months of birth.
The primary outcomes were NICU admission and NICU length of stay (LOS). NICU admission was defined as any instance in the immediate postpartum period where an infant was admitted to the neonatal ICU. Further, LOS was defined as the number of days a neonate was admitted to the ICU before discharge or demise.
Secondary outcomes included gestational age (GA) at birth, neonatal birth weight, neonatal hypoglycemia, respiratory distress syndrome (RDS), maternal hypertensive disorders of pregnancy, mode of birth, and shoulder dystocia. Neonatal weight was defined as birth weight measured in grams. Large for GA (LGA) status was defined as birth weight>90th percentile for GA at birth and small for GA (SGA) status was defined as birth weight<10th percentile.15 Neonatal hypoglycemia was defined as any glucose levels<40 mg/dL over first 48 hours of life, as documented in the neonate’s EHR, consistent with our institution’s clinical treatment guidelines. RDS was determined based on ICD-10 codes identified in the neonate’s EHR. Hypertensive disorders of pregnancy (gestational hypertension, pre-eclampsia (with and without severe features, and superimposed pre-eclampsia) were defined by ACOG guidelines.16 The mode of birth was delineated as vaginal, operative vaginal, or cesarean birth. Shoulder dystocia is described when the neonate’s anterior shoulder is caught above the maternal pubic bone.
Statistical analysis
Student’s t-tests and Wilcoxon rank-sum tests were used to investigate the association between “poor glucose control” language and continuous variables. Χ2 and Fisher’s exact tests were used for categorical variables. Unadjusted and adjusted logistic regression models were used to investigate the effect of various factors (race, insurance status, parity, and DM medication) on “poor glucose control” language. For adjusted analyses, individuals with missing data for any of the included covariates were excluded from the analysis (ie, complete case analysis). Similarly, logistic regression models were used to investigate the effect of “poor glucose control” language on maternal and neonatal outcomes. ORs and 95% CI were obtained and adjusted for diabetes type and GA at birth as noted. All reported p values are two sided, and a significance level of 0.05 was used. Statistical analyses were performed using R (V.4.1.2, R Core Team).
Results
Between 2018 and 2019, 1599 births from individuals with a diagnosis of diabetes in pregnancy were identified (online supplemental figure 2). Individuals with known fetal anomalies (n=43) or not found to have diabetes on manual chart review (n=123) were excluded. Thus, 1433 individuals met the inclusion criteria, of which 143 (10%) were found to have “poor glucose control” language documented in their medical records.
Of the 22 194 notes identified within the UMOMs database between 2018 and 2019, 6237 were documented for the study’s cohort. 396 notes were found to have “poor glucose control” language (online supplemental figure 3A). The most frequently used terms included “uncontrolled diabetes” (42.2%), “poorly controlled diabetes” (31.8%), and “poor glucose control” (21%). When evaluating note distribution by type of healthcare professional, the majority were written by physicians (63.1%) and mid-level professionals (18.2%) (online supplemental figure 3B). “Poor glucose control” language was primarily used by physicians within the obstetrics/gynecology (62.6%), medicine (12.9%), and maternal–fetal medicine specialties (11.1%) (online supplemental figure 3C).
When analyzing cohort characteristics, pregnant individuals of color, those on public insurance, those speaking languages other than English or Spanish, and those requiring insulin had a higher percentage categorized as having “poor glucose control” (table 1). When adjusted for diabetes type, pregnant individuals of color had higher odds of being classified as having “poor glucose control” compared with white individuals (adjusted OR (aOR) 2.4, 95% CI 1.63 to 3.57, p<0.001; table 2). Those on public insurance had approximately three times the odds of being described as having “poor glucose control” compared with those with private insurance (aOR 3.22, 95% CI 2.2 to 4.74, p<0.001; table 2). Individuals with a primary spoken language other than English or Spanish had two times the odds of being categorized with “poor glucose control” (aOR 2.07, 95% CI 1.22 to 2.4, p=0.0005). Finally, those not requiring medications had approximately 85% lower odds of being described as having “poor glucose control” compared with those on insulin (aOR 0.18, 95% CI 0.09 to 0.33, p<0.001; table 2).
Table 1. Baseline demographics.
| Demographic | Individuals with “poor glucose control” language (n=143) | Individuals without “poor glucose control” language (n=1290) | P value* |
|---|---|---|---|
| Age, mean (SD) | 33.6 (5.6) | 33.0 (5.0) | 0.219 |
| Race†, n (%) | <0.001 | ||
| White | 58 (7.2%) | 747 (92.8%) | |
| Black | 53 (22.7%) | 180 (77.3%) | |
| Asian | 10 (4.2%) | 229 (95.8%) | |
| American Indian or Alaska Native | 7 (28.0%) | 18 (72.0%) | |
| Native Hawaiian or Pacific Islander | 1 (25.0%) | 3 (75.0%) | |
| More than one race | 2 (8.0%) | 23 (92.0%) | |
| Prepregnancy BMI‡ (mg/kg†), mean (SD) | 32.5 (6.4) | 30.4 (7.6) | 0.009 |
| Language, n (%) | 0.017 | ||
| English | 109 (9.0%) | 1100 (91.0%) | |
| Spanish | 11 (16.2%) | 57 (83.8%) | |
| Other | 23 (14.7%) | 133 (85.3%) | |
| Insurance type, n (%) | <0.001 | ||
| Uninsured | 5 (7.9%) | 58 (92.1%) | |
| Public | 90 (16.1%) | 470 (83.9%) | |
| Private | 48 (5.9%) | 762 (94.1%) | |
| Parity, n (%) | 0.134 | ||
| Nulliparous | 44 (8.4%) | 479 (91.6%) | |
| Multiparous | 99 (10.9%) | 811 (89.1%) | |
| Diabetes type, n (%) | <0.001 | ||
| T1DM | 15 (24.2%) | 47 (75.8%) | |
| T2DM | 37 (24.7%) | 113 (75.3%) | |
| GDMA1 | 26 (4.0%) | 624 (96.0%) | |
| GDMA2 | 65 (11.4%) | 506 (88.6%) | |
| Diabetes medication, n (%) | <0.001 | ||
| None | 21 (3.1%) | 661 (96.9%) | |
| Insulin | 111 (16.5%) | 560 (83.5%) | |
| Oral medication | 11 (13.8%) | 69 (86.2%) |
P value is for Student’s t-test for continuous variables. Χ2 or Fisher’s exact tests were used for categorical variables.
Race is missing for 12 individuals with “poor glucose control” language and for 90 individuals without “poor glucose control” language.
BMI is missing for 39 individuals with “poor glucose control” language and for 601 individuals without “poor glucose control” language.
BMI, body mass index; GDMA1, Gestational Diabetes, A1 (GDM not requiring medication); GDMA2, Gestational Diabetes, A2 (GDM requiring medication) ; T1DM, type 1 diabetes mellitus; T2DM, type 2 diabetes mellitus.
Table 2. Baseline demographics ORs.
| Demographic Characteristics |
Unadjusted* | Adjusted† | ||
|---|---|---|---|---|
| OR (95% CI) | P value* | OR (95% CI) | P value* | |
| Race‡, n (%) | ||||
| White | Ref | Ref | ||
| Persons of color | 2.08 (1.44 to 3) | <0.001 | 2.4 (1.63 to 3.57) | <0.001 |
| Insurance type, n (%) | ||||
| Uninsured | 1.37 (0.46 to 3.27) | 0.521 | 1.34 (0.44 to 3.28) | 0.561 |
| Public | 3.04 (2.11 to 4.42) | <0.001 | 3.22 (2.2 to 4.74) | <0.001 |
| Private | Ref | Ref | ||
| Language, n (%) | ||||
| Spanish | 1.95 (0.94 to 3.69) | 0.053 | 1.86 (0.88 to 3.6) | 0.082 |
| Other | 1.75 (1.05 to 2.79) | 0.024 | 2.07 (1.22 to 3.4) | 0.005 |
| English | Ref | Ref | ||
| Diabetes medication, n (%) | ||||
| None | 0.16 (0.1 to 0.25) | <0.001 | 0.18 (0.09 to 0.33) | <0.001 |
| Insulin | Ref | Ref | ||
| Oral medication | 0.8 (0.39 to 1.51) | 0.523 | 0.85 (0.41 to 1.61) | 0.634 |
Unadjusted logistic regression models were used to investigate the effect of demographics on “poor glucose control” language.
Adjusted logistic regression models were used adjusting for diabetes type.
Race was missing for 12 individuals with “poor glucose control” language and for 90 individuals without “poor glucose control” language.
We then examined clinical factors that may have been used to classify individuals with “poor glucose control.” Individuals categorized with “poor glucose control” had glucose values documented and A1c assessed more often than those without that categorization. However, only 53.1% of individuals categorized with “poor glucose control” had any glucose documented in the EHR 2 weeks preceding birth, and only 33.6% had an A1c obtained within 3 months of birth (table 3). In terms of recorded glucose values, more individuals without “poor glucose control” language were found to have >50% abnormal glucose values reported before birth (77.2% vs 36.8%, p<0.001) compared with those classified with “poor glucose control.” Median A1c levels were higher in individuals categorized with “poor glucose control” compared with those without poor glucose language (6.3 mg/dL vs 5.4 mg/dL). However, 60.4% of those classified as having “poor glucose control” had an A1c≤6.5%. Since white individuals were less likely to be described as having “poor glucose control” than individuals of color, a subgroup analysis of these factors was conducted and found similar rates of documented glucoses, as well as reported glucose values, and A1c within 3 months of birth in both cohorts (online supplemental table 2).
Table 3. Objective factors associated with categorization of “poor glucose control”.
| Variable | Individuals with “poor glucose control” language (n=143) | Individuals without “poor glucose control” language (n=1290) | P value* |
|---|---|---|---|
| Documented glucose 2 weeks prior to birth, n (%) | 76 (53.1%) | 403 (31.2%) | <0.001 |
| >50% abnormal glucose values, n (%) | <0.001 | ||
| Yes | 28 (36.8%) | 311 (77.2%) | |
| No | 32 (42.1%) | 44 (10.9%) | |
| Unable to assess | 16 (21.1%) | 48 (11.9%) | |
| Hemoglobin A1C taken 3 months prior to birth, n (%) | 48 (33.6%) | 158 (12.2%) | <0.001 |
| Median (Q1, Q3) | 6.3 (5.9, 7.6) | 5.4 (5.2, 5.8) | – |
| >6.5% | 19 (39.6%) | 14 (8.9%) | – |
| ≤6.5% | 29 (60.4%) | 144 (91.1%) |
P value is for χ2 or Fisher’s exact tests.
When examining birth outcomes, individuals categorized with “poor glucose control” gave birth earlier than their counterparts (37.4 vs 39.0 weeks, p<0.001), had higher rates of shoulder dystocia, and were more likely to deliver an LGA neonate (table 4). When adjusted for diabetes type, the odds of delivering LGA infants were approximately two times higher for individuals categorized as having “poor glucose control” (aOR 1.95, 95% CI 1.24 to 3, p=0.003; online supplemental table 3). When adjusted for diabetes type and GA at birth, the odds of shoulder dystocia were approximately six times higher (aOR 6.15, 95% CI 2.18 to 15.9, p<0.001; online supplemental table 3).
Table 4. Comparison of neonatal outcomes by “poor glucose control” language.
| Variable | Individuals with “poor glucose control” language (n=143) | Individuals without “poor glucose control” language (n=1290) | P value* |
|---|---|---|---|
| Mode of birth, n (%) | 0.084 | ||
| Vaginal | 64 (44.8%) | 686 (53.2%) | |
| C-section | 77 (53.8%) | 568 (44.0%) | |
| Assisted | 2 (1.4%) | 36 (2.8%) | |
| Gestational age at birth, median (Q1, Q3) | 37.4 (36.3, 38.4) | 39.0 (37.7, 39.4) | <0.001 |
| Birth weight† (g), median (Q1, Q3) | 3371.5 (2948.8, 3684.2) | 3330.0 (2980.2, 3657.2) | 0.682 |
| Birthweight categories†, n (%) | |||
| Small for gestational age | 5 (3.7%) | 74 (5.8%) | <0.001 |
| Appropriate for gestational age | 94 (70.1%) | 1025 (80.8%) | |
| Large for gestational age | 35 (26.1%) | 170 (13.4%) | |
| Neonatal intensive care unit (NICU) admission, n (%) | 52 (36.4%) | 186 (14.4%) | <0.001 |
| NICU length of stay, median (Q1, Q3) | 10.0 (3.8, 18.9) | 9.4 (3.0, 19.1) | 0.900 |
| Neonatal hypoglycemia, n (%) | 78 (54.5%) | 539 (41.8%) | 0.003 |
| Respiratory distress syndrome, n (%) | 32 (22.4%) | 137 (10.6%) | <0.001 |
| Shoulder dystocia, n (%) | 7 (4.9%) | 20 (1.6%) | 0.014 |
P value is for Wilcoxon rank-sum test for continuous variables. Χ2 or Fisher’s exact tests were used for categorical variables.
Birth weight was missing for nine individuals with “poor glucose control” language and 21 individuals without “poor glucose control” language.
In terms of NICU admission, more individuals categorized as having “poor glucose control” had infants admitted to the NICU compared with their counterparts (36.4% vs 14.4%, p<0.001; table 4), though no difference was seen in NICU LOS. When adjusted for diabetes type and GA at birth, neonates of individuals categorized with “poor glucose control” had approximately 70% higher odds of NICU admission (aOR 1.72, 95% CI 1.07 to 2.74, p=0.024; online supplemental table 3). While infants born to individuals categorized with “poor glucose control” were found to have higher rates of neonatal hypoglycemia and RDS, no differences were seen when adjusting for diabetes type and GA at birth (online supplemental table 3).
Given the number of individuals classified as having “poor glucose control” who had A1c≤6.5% within 3 months of birth, we compared neonatal outcomes in this subgroup. Infants of individuals with a normal A1c but categorized with “poor glucose control” had an earlier GA at birth (37.1 vs 38.7 weeks, p<0.001) and were more likely to be admitted to the NICU (55.2% vs 17.4%, p<0.001; online supplemental table 4) than those without “poor glucose control” language. However, no significant differences were seen in neonatal weight categorization, diagnosis of neonatal hypoglycemia, RDS, or shoulder dystocia.
Finally, we assessed how often individuals categorized as having “poor glucose control” gave birth prior to the ACOG-recommended 39th week. Of the 143 individuals categorized with poor glucose control, 117 (81.8%) experienced birth less than 39 weeks’ gestation. Reviewing indications for birth <39 weeks in this group, 71 (60.7%) of individuals had birth clinically recommended solely for “poor glucose control” indications. Overall, these recommendations were made with little objective data on hyperglycemia: the median A1c was 5.9% and approximately 40% had more than 50% of glucose values within goal in the 2 weeks prior to birth (online supplemental table 5).
Conclusions
Using NLP algorithms to identify terminology associated with “poor glucose control” in clinical notes, we evaluated the demographic factors, markers of glycemia, and maternal–neonatal outcomes associated with health care professional descriptions of “poor glucose control” in pregnancy. “Poor glucose control” and “uncontrolled diabetes” were the most frequently used terms, and physicians used these descriptions more than other healthcare professionals. Individuals classified with “poor glucose control” were more likely to be birthing individuals of color, use public insurance, or require insulin therapy. These individuals gave birth a week earlier, and their neonates were more likely to be LGA and admitted to the NICU. While individuals categorized by healthcare professionals as having “poor glucose control” had worse outcomes in several domains, we identified that this clinical classification was often made in the absence of any objective data suggestive of hyperglycemia or even in the presence of normal glucose and A1c values.
The correlation between elevated maternal glucose levels and poor perinatal clinical outcomes is well documented.17 To reduce the risk of some of these adverse events, such as stillbirth, early birth has been recommended for individuals with “poor glucose control.”4 5 However, there are no established criteria for categorizing individuals as having “poor glucose control” necessitating early intervention. Thus, healthcare professionals rely on their clinical judgment to guide management, such as the timing of birth recommendations.
While clinical judgment is an essential component of care, for many individuals categorized by healthcare professionals as having “poor glucose control,” no objective markers of glucose levels were identified in the EHR. Individuals classified as having “poor glucose control” had increased complications not solely related to GA at birth, such as LGA status and shoulder dystocia, suggesting clinical judgment does accurately identify individuals at increased risk of some complication. However, we also identified individuals categorized as having “poor glucose control” who had normal A1c within 3 months of birth and did not experience increased complications. This suggests that some individuals may be incorrectly identified as having “poor glucose control” despite having normal objective markers of glucose.
No guidelines currently identify a glycemic marker to guide early birth recommendations. In our study, we selected blood glucose from self-blood glucose monitoring or continuous glucose monitoring data and A1c as indicators of maternal glycemia, as these are the commonly used markers and easily accessible in the EHR. Notably, no consensus guideline on elevated self-monitored glucose levels representing “poor glucose control” exists. However, a threshold greater than 50% of readings above goal was used based on our health system’s recommendations for initiating or adjusting insulin in pregnancy. We identified that only 53% of individuals categorized as having “poor glucose control” had any reported glucose values documented in the EHR (from clinical notes or scanned documents) and that only 37% of those individuals with documented glucose had >50% of values above the target range. Surprisingly, 77% of individuals without “poor glucose” categorization who had clinically documented glucose values had more than 50% of values above the target range. This suggests that “poor glucose control” categorization was not necessarily equitably related to reported glucose levels.
A1c has been recommended by the American Diabetes Association (ADA) as an adjunct measure of glycemia for individuals with pre-existing diabetes.18 A1c measurements tend to decrease throughout gestation due to increases in red blood cell turnover and may be difficult to interpret in the presence of maternal anemia.19,21 However, several studies have identified A1c elevation at term and increasing A1c levels throughout pregnancy as associated with increased risk of perinatal complications.22,24 In our study, only 53% of individuals categorized as having “poor glucose control” and 31% without “poor glucose control” had any A1c documented within 3 months of birth. While the median A1c was higher in the group categorized with “poor glucose” control, over 60% had an A1c≤6.5% (normal range). Additionally, 8.9% in the group without “poor glucose control” language had an A1c>6.5%.
Our study further revealed that many individuals categorized with “poor glucose control” were those of marginalized populations (ie, black, those requiring public insurance, non-English speakers). This categorization was made at increased rates in individuals of color compared with white individuals despite no objective differences in glucose values. This indicates that some clinicians may be classifying marginalized individuals with “suboptimal glycemic control” without supportive evidence of hyperglycemia. Racial differences in A1c have been reported early in pregnancy. However, both non-Hispanic black and white birthing individuals were equally able to achieve A1c targets≤6.5% despite increases in the Social Vulnerability Index among black individuals.25 It is well established that black birthing individuals and those on public insurance are more likely to experience preterm birth.26,28 While the etiology for this is multifactorial, the categorization of black, brown, and low-income birthing individuals with “poor glucose control” could potentially further perpetuate poor neonatal health outcomes, contributing to long-lasting health burdens on marginalized birthing individuals and their families.
Based on increased rates of LGA status and shoulder dystocia in those categorized as having “poor glucose control,” we suspect that some individuals were correctly identified by clinicians as having elevated glucose levels. However, many individuals were categorized by obstetricians as having “poor glucose control” despite the presence of normal objective markers of glycemia. Thus, a potential risk of this designation is unnecessary late preterm/early term birth, resulting in increased NICU admission, which can increase costs of care and adverse outcomes unnecessarily.1029,31
Strengths and limitations
Several strengths of this study are noted. This study includes NLP algorithms and manual verification of clinical contexts to systematically identify the use of “poor glucose control language” and its variants in EHR documentation. This was done in the context of a large community-academic health system with over 10 000 births per year across seven hospitals spanning urban, suburban, and rural settings with care provided by multiple provider types in multiple practice settings, enhancing generalizability. Of note, within our system, maternal–fetal medicine specialists and endocrinologists primarily play a consultative role with most individuals remaining within the care of their primary obstetrics clinician. Although there is potential for variation in care patterns within our system, a system-wide ACOG-aligned and ADA-aligned protocol is in place for the care of individuals with diabetes in pregnancy.
There are several limitations to this study. Given the observational design, we can only identify associations between “poor glucose control” language use and individual characteristics and clinical outcomes, not causation. Further, due to limitations of data availability (including prepregnancy BMI and Area Deprivation Index), we were not able to control for all factors that may contribute to clinician language use. Additionally, language was only identified if found within clinical notes of the EHR; thus, we may have missed some individuals if language was only used in scanned outpatient documents. However, because of this potential concern, we did not quantify outcomes by the number of documented events; rather, we looked at whether any documentation was included in the EHR in any phase of pregnancy care. Additionally, while our EHR contains notes, laboratory results, electronic messages, and scanned records, and we made every effort to review these documents when assessing reported glucose levels, it is possible that we did not identify all documented glucose values. However, we believe the potential impact of this on reported findings is small. Finally, while we made efforts to identify objective markers clinicians may have used to inform their assessment of glucose levels, we cannot fully assess whether bias may have directly contributed to differential use of this terminology among marginalized groups.
In summary, pregnant individuals of color, those on public insurance, and those speaking a language other than English or Spanish are more likely to be categorized as having “poor glycemic control.” While some individuals classified as having “poor glucose control” did have glucose levels above goal, a number were identified as having “poor glucose control” without any supporting objective data. Additional clinical guidelines are needed from ACOG and other leading organizations to more clearly outline the degree of hyperglycemia and other clinical findings to support recommendations for early birth. Clearer guidelines may support less biased clinical decision-making regarding the timing of birth among individuals with diabetes and decrease potential complications.
Supplementary material
Footnotes
Funding: Research reported in this publication was supported by the National Institutes of Health (NIH), grant P30 CA77598, utilizing the Biostatistics Core shared resource of the Masonic Cancer Center, University of Minnesota, and by the National Center for Advancing Translational Sciences of the NIH, award number UM1TR004405. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. The investigators also received support from the Department of Obstetrics, Gynecology and Women’s Health at the University of Minnesota. The funders had no role in study design, data collection, data analysis, data interpretation, or manuscript writing. The funders had no role in the decision to submit the paper.
Provenance and peer review: Not commissioned; externally peer reviewed.
Patient consent for publication: Not applicable.
Ethics approval: This study involves human participants and was approved by the institutional review board (IRB) of the University of Minnesota (#STUDY00016980). Our large clinical database contains only individuals who do not opt out of research, and all data were deidentified in analysis.
Data availability free text: The datasets generated during and/or analyzed in the current study are available from the corresponding author on reasonable request.
Data availability statement
Data are available upon reasonable request.
References
- 1.Deputy NP, Kim SY, Conrey EJ, et al. Prevalence and Changes in Preexisting Diabetes and Gestational Diabetes Among Women Who Had a Live Birth - United States, 2012-2016. MMWR Morb Mortal Wkly Rep. 2018;67:1201–7. doi: 10.15585/mmwr.mm6743a2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Battarbee AN, Venkatesh KK, Aliaga S, et al. The association of pregestational and gestational diabetes with severe neonatal morbidity and mortality. J Perinatol. 2020;40:232–9. doi: 10.1038/s41372-019-0516-5. [DOI] [PubMed] [Google Scholar]
- 3.Ye W, Luo C, Huang J, et al. Gestational diabetes mellitus and adverse pregnancy outcomes: systematic review and meta-analysis. BMJ. 2022;377:e067946. doi: 10.1136/bmj-2021-067946. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.ACOG Practice Bulletin No. 201: Pregestational Diabetes Mellitus. Obstetrics & Gynecology. 2018;132:e228–48. doi: 10.1097/AOG.0000000000002960. [DOI] [PubMed] [Google Scholar]
- 5.ACOG Practice Bulletin No. 190: Gestational Diabetes Mellitus. Obstetrics & Gynecology. 2018;131:e49–64. doi: 10.1097/AOG.0000000000002501. [DOI] [PubMed] [Google Scholar]
- 6.Dickinson JK, Guzman SJ, Maryniuk MD, et al. The Use of Language in Diabetes Care and Education. Diabetes Care. 2017;40:1790–9. doi: 10.2337/dci17-0041. [DOI] [PubMed] [Google Scholar]
- 7.Cooper A, Kanumilli N, Hill J, et al. Language matters. Addressing the use of language in the care of people with diabetes: position statement of the English Advisory Group. Diabet Med . 2018;35:1630–4. doi: 10.1111/dme.13705. [DOI] [PubMed] [Google Scholar]
- 8.J S, T C S, T D, et al. Our language matters: Improving communication with and about people with diabetes. A position statement by Diabetes Australia. Diabetes Res Clin Pract. 2021;173:108655. doi: 10.1016/j.diabres.2021.108655. [DOI] [PubMed] [Google Scholar]
- 9.Dickinson JK, Guzman SJ, Wooldridge JS. The Emotional Impact of Negative Language in People With Diabetes: A Descriptive Study Using a Semantic Differential Scale. Sci Diabetes Self Manag Care . 2023;49:193–205. doi: 10.1177/26350106231168326. [DOI] [PubMed] [Google Scholar]
- 10.Nashif SK, Mahr RM, Tessier KM, et al. Neonatal outcomes and rationale for timing of birth in perinatal diabetes: a retrospective cohort study. Am J Obstet Gynecol MFM. 2023;5:101129. doi: 10.1016/j.ajogmf.2023.101129. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.von Elm E, Altman DG, Egger M, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. J Clin Epidemiol. 2008;61:344–9. doi: 10.1016/j.jclinepi.2007.11.008. [DOI] [PubMed] [Google Scholar]
- 12.P. Goddu A, O’Conor KJ, Lanzkron S, et al. Do Words Matter? Stigmatizing Language and the Transmission of Bias in the Medical Record. J GEN INTERN MED. 2018;33:685–91. doi: 10.1007/s11606-017-4289-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.ElSayed NA, Aleppo G, Bannuru RR, et al. 2. Diagnosis and Classification of Diabetes:Standards of Care in Diabetes—2024. Diabetes Care. 2024;47:S20–42. doi: 10.2337/dc24-S002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.McEwan R, Melton GB, Knoll BC, et al. NLP-PIER: A Scalable Natural Language Processing, Indexing, and Searching Architecture for Clinical Notes. AMIA Jt Summits Transl Sci Proc. 2016;2016:150–9. [PMC free article] [PubMed] [Google Scholar]
- 15.Fenton TR, Kim JH. A systematic review and meta-analysis to revise the Fenton growth chart for preterm infants. BMC Pediatr. 2013;13:59. doi: 10.1186/1471-2431-13-59. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Gestational Hypertension and Preeclampsia. Obstetrics & Gynecology. 2020;135:e237–60. doi: 10.1097/AOG.0000000000003891. [DOI] [PubMed] [Google Scholar]
- 17.HAPO Study Cooperative Research Group. Metzger BE, Lowe LP, et al. Hyperglycemia and adverse pregnancy outcomes. N Engl J Med. 2008;358:1991–2002. doi: 10.1056/NEJMoa0707943. [DOI] [PubMed] [Google Scholar]
- 18.ElSayed NA, McCoy RG, Aleppo G, et al. 15. Management of Diabetes in Pregnancy: Standards of Care in Diabetes—2025. Diabetes Care. 2025;48:S306–20. doi: 10.2337/dc25-S015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Mosca A, Paleari R, Dalfrà MG, et al. Reference intervals for hemoglobin A1c in pregnant women: data from an Italian multicenter study. Clin Chem. 2006;52:1138–43. doi: 10.1373/clinchem.2005.064899. [DOI] [PubMed] [Google Scholar]
- 20.Nielsen LR, Ekbom P, Damm P, et al. HbA1c levels are significantly lower in early and late pregnancy. Diabetes Care. 2004;27:1200–1. doi: 10.2337/diacare.27.5.1200. [DOI] [PubMed] [Google Scholar]
- 21.Edelson PK, James KE, Leong A, et al. Longitudinal Changes in the Relationship Between Hemoglobin A1c and Glucose Tolerance Across Pregnancy and Postpartum. J Clin Endocrinol Metab. 2020;105:e1999–2007. doi: 10.1210/clinem/dgaa053. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Koefoed AS, Knorr S, Fuglsang J, et al. Hemoglobin A1c Trajectories During Pregnancy and Adverse Outcomes in Women With Type 2 Diabetes: A Danish National Population-Based Cohort Study. Diabetes Care. 2024;47:1211–9. doi: 10.2337/dc23-2304. [DOI] [PubMed] [Google Scholar]
- 23.Kiefer MK, Finneran MM, Ware CA, et al. Association of change in haemoglobin A1c with adverse perinatal outcomes in women with pregestational diabetes. Diabet Med . 2022;39:e14822. doi: 10.1111/dme.14822. [DOI] [PubMed] [Google Scholar]
- 24.Finneran MM, Kiefer MK, Ware CA, et al. The use of longitudinal hemoglobin A1c values to predict adverse obstetric and neonatal outcomes in pregnancies complicated by pregestational diabetes. Am J Obstet Gynecol MFM. 2020;2:100069. doi: 10.1016/j.ajogmf.2019.100069. [DOI] [PubMed] [Google Scholar]
- 25.Venkatesh KK, Fareed N, Kiefer MK, et al. Differences in Hemoglobin A1c during Pregnancy between Non-Hispanic Black versus White Women with Prepregnancy Diabetes. Am J Perinatol. 2022;39:1279–87. doi: 10.1055/a-1788-5600. [DOI] [PubMed] [Google Scholar]
- 26.Jelliffe-Pawlowski LL, Baer RJ, Oltman S, et al. Risk and Protective Factors for Preterm Birth Among Racial, Ethnic, and Socioeconomic Groups in California. JAMA Netw Open . 2024;7:e2435887. doi: 10.1001/jamanetworkopen.2024.35887. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Johnson JD, Green CA, Vladutiu CJ, et al. Racial Disparities in Prematurity Persist among Women of High Socioeconomic Status. Am J Obstet Gynecol MFM . 2020;2:100104. doi: 10.1016/j.ajogmf.2020.100104. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Manuck TA. Racial and ethnic differences in preterm birth: A complex, multifactorial problem. Semin Perinatol. 2017;41:511–8. doi: 10.1053/j.semperi.2017.08.010. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Watson KQ, Kannan A, Sobhani NC. Impact of timing of delivery for type 2 diabetes on perinatal outcomes. Diabetes Epidemiology and Management . 2024;14:100196. doi: 10.1016/j.deman.2024.100196. [DOI] [Google Scholar]
- 30.Austin BA, McFarling KM, Likins B, et al. Impact on Neonatal Outcomes with Late Preterm and Early Term Delivery in Women with Diabetes. Am J Perinatol. 2024;41:122–6. doi: 10.1055/s-0043-1774311. [DOI] [PubMed] [Google Scholar]
- 31.Petrou S. Health economic aspects of late preterm and early term birth. Semin Fetal Neonatal Med. 2019;24:18–26. doi: 10.1016/j.siny.2018.09.004. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data are available upon reasonable request.