Development of a Risk-Scoring System for Prediction of Blood Transfusion During Hospitalization for Delivery

An internally validated, practical, and data-driven risk-scoring system predicts blood transfusion during hospitalization for delivery.

Abstract

OBJECTIVE:

To develop and internally validate a practical and data-driven risk-scoring system to predict blood transfusion during hospitalization for delivery in a contemporary U.S. cohort.

METHODS:

This was a secondary analysis of a multicenter cohort of patients who delivered on randomly selected days at 17 U.S. hospitals (2019–2020). Patients with placenta accreta spectrum were excluded. The primary outcome was any blood transfusion during hospitalization for delivery. Candidate risk factors for transfusion were selected based on relevant literature. A multivariable logistic regression model was developed and internally validated using stratified k-fold (k=5) cross validation with stepwise backward elimination that used significance level of 0.05. Each risk factor included in the final model was assigned a point value by dividing the log of the odds ratio (OR) by the log of the OR of the factor with the lowest value. The summed points for an individual generate a numeric risk score predictive of transfusion. Performance of the risk score to predict transfusion was assessed using the area under the receiver operating curve (AUC).

RESULTS:

Of 21,780 included individuals, 2.5% (n=545) received a blood transfusion. Factors associated with the highest risk for transfusion in the final model included thrombocytopenia, and placental abruption or significant antepartum bleeding. Risk score outputs among patients in the cohort ranged from 0 to 17 (maximum possible 26) with a corresponding predicted risk for transfusion from 1.0% to 84.4%. The AUC for prediction of transfusion in the validation subsample was 0.81 (95% CI, 0.76–0.85).

CONCLUSION:

We developed a clinically applicable numeric risk score to predict blood transfusion during hospitalization for delivery. Future work should externally validate this risk-scoring system.

Postpartum hemorrhage is a leading cause of maternal morbidity and mortality worldwide.^1,2 Significant effort has been put into developing postpartum hemorrhage and blood transfusion risk prediction tools to inform clinical preparedness (eg, intravenous access, uterotonics, blood product availability) in an effort to reduce hemorrhage-associated morbidity.^1,3 The California Maternal Quality Care Collaborative (CMQCC) risk assessment tool is the most widely used example in the United States.⁴ The tool considers patient-level factors (eg, leiomyomas), antepartum factors (eg, placenta previa), and intrapartum events (eg, prolonged labor, chorioamnionitis) to categorize individuals into three groups (low-risk, medium-risk, or high-risk) for blood transfusion during hospitalization for delivery with corresponding recommendations for blood preparedness (ie, type and hold, type and screen, type and cross).⁴

Despite widespread use, the CMQCC assessment tool is limited in its ability to distinguish risk in individuals given that they are categorized discretely into only three groups. For example, 7–10% of transfusions occur in the low-risk group.^5,6 As alternatives to the CMQCC risk assessment tool, regression-based and machine learning models for prediction of hemorrhage and blood transfusion have been developed.^7–9 Such models afford consideration of a wider range of variables to predict transfusion. However, clinical application is limited without a user-friendly interface for labor and delivery units or cut points to inform blood product preparedness.

An alternative to the CMQCC risk tool and regression models is the use of a weighted risk score.^10,11 Such an approach allows for a data-driven weight to be provided for risk factors while producing a more continuous risk assessment for ease of clinical application. Therefore, we aimed to develop and internally validate a practical risk-scoring system to predict blood transfusion during hospitalization for delivery in a contemporary U.S. cohort.

METHODS

This was a secondary analysis of the Eunice Kennedy Shriver National Institute of Child Health and Human Development Maternal-Fetal Medicine Units Network’s GRAVID (Gestational Research Assessments for COVID-19) study.^12,13 In brief, the GRAVID study was a retrospective cohort study of pregnant and postpartum individuals with confirmed severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection or who delivered without such an infection on a randomly selected day at one of 17 U.S. hospitals from March 2019 to December 2020. Trained perinatal research personnel completed medical record abstraction for demographics, medical and obstetric history, and outcomes through 6 weeks postpartum. The study methodology and findings have previously been published.^12,13

For this analysis, individuals who delivered on randomly selected days (four random days per month from March 2019 to December 2019, eight random days per month from March 2020 to May 2020, and four random days per month from June 2020 to December 2020) at more than 20 weeks of gestation were included.¹² Individuals with antenatally suspected placenta accreta spectrum were excluded based on the known high risk for hemorrhage and transfusion and standard approach of blood product preparedness for these deliveries, and to ensure that the newly developed scoring system was calibrated to those whose risk was not as clearly extant.^14,15

The primary outcome was any blood transfusion during hospitalization for delivery. To develop the risk-scoring system, candidate risk factors for transfusion were selected based on relevant literature and clinical expertise.^{1,4–8,16,17} Those candidate risk factors considered are in Box 1.

Box 1. Candidate Risk Factors for Blood Transfusion During Delivery Hospitalization.

• Prior cesarean delivery or uterine surgery (eg, myomectomy)

• Multifetal gestation

• More than 4 prior vaginal deliveries

• Suspected triple I/chorioamnionitis (defined clinically)

• Thrombocytopenia (platelets less than 150,000/microliter)

• Anemia (identified by iron supplementation)

• Polyhydramnios

• Preterm delivery (less than 37 wk of gestation)

• Postterm delivery (more than 41 wk of gestation)

• Preeclampsia with severe features or HELLP syndrome, as defined by the American College of Obstetricians and Gynecologists

• Abruption or significant antepartum bleeding*

• Placenta previa

• Fetal death

• General anesthesia

• Prolonged labor (more than 24 h)

• Unscheduled cesarean delivery (eg, occurring intrapartum)

• Scheduled cesarean delivery

• Suspected macrosomia

• Antepartum blood transfusion (before hospitalization for delivery)

A multivariable logistic regression model was developed and internally validated using stratified k-fold (k=5) cross validation with stepwise backward elimination that used a significance level of 0.05. Due to the low event rate, a stratified resampling method was used to improve the reliability and accuracy of the model. Each group was randomly divided into five equal parts among all patients included in the analysis. Using backward elimination, a logistic regression model was developed using four of the five parts. Validation for each of the five models was determined from the one-fifth part of the cohort that was not included in fitting each model. Among the developed models, the final model is the one with the highest c-statistic (area under the receiver operating curve [AUC]) where the factors in the final model appear in at least four of the five models.

Following the methods of Leonard et al¹⁸ and Sullivan et al,¹⁹ regression model outputs were converted to a numeric score for included risk factors. Each risk factor included in the final regression model using data from the full cohort was assigned a point value by dividing the log of the odds ratio (OR) by the log of the OR of the risk factor with the lowest value. This uses the lowest statistically significant OR as the reference, where using it as the denominator allows for risk factors with higher ORs to be proportionally valued with a higher point value. A total numeric risk score for transfusion for a specified individual was then calculated based on the sum of the point values for each risk factor for transfusion.

Univariable comparisons of baseline characteristics and candidate risk factors between individuals receiving a blood transfusion and those not receiving a blood transfusion were performed using χ², Fisher exact test, or Wilcoxon test, as appropriate. The adjusted risk difference with 95% CI was reported. This reflects the attributable risk for transfusion resulting from exposure to the specified risk factor in a theoretic population of 1,000 individuals.

The risk score for each risk factor included was reported. The total individual score range possible was reported based on outputs in this data set. Performance of the risk score to predict transfusion was assessed using the AUC for the validation subset and the full population.

P<.05 was considered statistically significant. No adjustment was made for multiple comparisons. As a secondary analysis, outcomes are considered exploratory. Analyses were completed using SAS 9.4. The GRAVID study was approved by the institutional review board of each participating institution.^12,13 This analysis was considered exempt from further institutional review board review. Reporting in this manuscript follows the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines.²⁰

RESULTS

After 58 exclusions for placenta accreta spectrum, 21,780 individuals were included in analysis and 2.5% (n=545) received a blood transfusion during their hospitalization for delivery. Individuals who received a blood transfusion were more likely to be non-Hispanic Black and nulliparous, to have conceived using assisted reproductive technology use, to have had an antepartum hospital admission, and to have had different primary prenatal care clinicians compared with those not receiving a transfusion (Table 1). Individuals receiving a blood transfusion were more likely to have or experience nearly all the candidate risk factors (Table 1).

Table 1.

Cohort Baseline Characteristics and Candidate Risk Factors by Blood Transfusion During Delivery Hospitalization

	Transfusion (n=545)	No Transfusion (n=21,235)	P
Characteristic
Maternal age (y)	30.1±6.2	29.8±5.8 (n=21,232)	.35
Race and ethnicity*			<.001
Hispanic	119 (21.8)	4,920 (23.2)
Non-Hispanic Black	151 (27.7)	4,541 (21.4)
Non-Hispanic White	203 (37.2)	9,401 (44.3)
None of the above or not reported	72 (13.2)	2,373 (11.2)
Private insurance	273/536 (50.9)	11,393/21,067 (54.1)	.15
Nulliparity	248/544 (45.6)	8,521/21,218 (40.2)	.01
BMI (kg/m2)	31.8 (27.6–36.8) (n=514)	31.1 (27.5–35.9) (n=20,601)	.06
Tobacco use	36 (6.6)	1,567 (7.4)	.49
Substance use	50 (9.2)	1,788 (8.4)	.53
Assisted reproductive technology	49/533 (9.2)	783/20,925 (3.7)	<.001
Antepartum admission	80 (14.7)	1,505 (7.1)	<.001
Primary prenatal care clinician			<.001
No prenatal care	31/541 (5.7)	272/21,114 (1.3)
Maternal–fetal medicine	87/541 (16.1)	2,244/21,114 (10.6)
Generalist	288/541 (53.2)	13,144/21,114 (62.3)
Family medicine practitioner	15/541 (2.8)	970/21,114 (4.6)
Midwife or nurse practitioner	77/541 (14.2)	3,572/21,114 (16.9)
Co-managed with maternal–fetal medicine	43/541 (8.0)	912/21,114 (4.3)
Risk factor
Prior cesarean or uterine surgery	144 (26.4)	3,907 (18.4)	<.001
Multifetal gestation	45 (8.3)	458 (2.2)	<.001
More than 4 prior vaginal deliveries	16 (2.9)	346 (1.6)	.02
Triple I/chorioamnionitis	47 (8.6)	904 (4.3)	<.001
Thrombocytopenia	50 (9.2)	193 (0.9)	<.001
Anemia	176 (32.3)	4,693 (22.1)	<.001
Polyhydramnios	3/236 (1.3)	81/9,197 (0.9)	.47
Preterm delivery (less than 37 wk)	161 (29.5)	2,718/21,225 (12.8)	<.001
Postterm delivery (more than 41 wk)	20 (3.7)	660/21,225 (3.1)	.46
Preeclampsia with severe features or HELLP syndrome	116 (21.3)	1,269 (6.0)	<.001
Abruption or significant antepartum bleeding	64 (11.7)	654/21,233 (3.1)	<.001
Placenta previa	16 (2.9)	89 (0.4)	<.001
Intrauterine fetal death	9 (1.7)	132 (0.6)	.003
General anesthesia	58 (10.6)	291/21,232 (1.4)	<.001
Prolonged labor (more than 24 h)	118/541 (21.8)	3,070/20,904 (14.7)	<.001
Unscheduled cesarean	106/466 (22.7)	1,162/18,650 (6.2)	<.001
Scheduled cesarean	79 (14.5)	2,583 (12.2)	.10
Suspected macrosomia	14 (2.6)	268 (1.3)	.01
Transfusion before delivery admission	6 (1.1)	13 (0.1)	<.001

BMI, body mass index; HELLP, hemolysis, elevated liver enzymes, and low platelet count.

Data are mean±SD, n (%), n/N (%), or median (interquartile range) unless otherwise specified.

^*Self-reported race and ethnicity were included to describe the baseline features of the study population.

The factors associated with the highest risk for transfusion during hospitalization for delivery in the adjusted multivariable model included thrombocytopenia (adjusted odds ratio [aOR] 5.2, 95% CI, 3.5–7.8; adjusted risk difference of 136/1,000 individuals, 95% CI, 81–191) and placental abruption or significant antepartum bleeding (aOR 3.9, 95% CI, 2.9–5.2; adjusted risk difference 53/1,000 individuals, 95% CI, 32–75; Table 2).

Table 2.

Adjusted Odds Ratio, Risk Difference Per 1,000, and Derived Risk Score for Factors Associated With Transfusion During Hospitalization for Delivery

Risk Factor	aOR (95% CI)	aRD/1,000 (95% CI)*	Score†
Thrombocytopenia‡	5.2 (3.5–7.8)	136 (81–191)	4
Abruption or significant antepartum bleeding	3.9 (2.9–5.2)	53 (32–75)	4
General anesthesia	3.2 (2.2–4.7)	92 (50–135)	3
Multifetal gestation	2.8 (1.9–4.2)	44 (16–72)	3
Preeclampsia with severe features or HELLP syndrome	2.7 (2.1–3.5)	39 (25–53)	3
Assisted reproductive technology	2.5 (1.8–3.6)	27 (10–44)	2
Unscheduled cesarean	2.5 (1.9–3.2)	33 (19–47)	2
Chorioamnionitis	2.2 (1.6–3.1)	23 (10–37)	2
Prior cesarean delivery or uterine surgery	1.7 (1.3–2.2)	12 (4–19)	1
Prolonged labor (more than 24 h)	1.6 (1.3–2.1)	9 (3–15)	1
Anemia	1.5 (1.2–1.8)	5 (1–10)	1

aOR, adjusted odds ratio; aRD, adjusted risk difference; HELLP, hemolysis, elevated liver enzymes, and low platelet count.

^*The aRD is the attributable risk for transfusion resulting from exposure to the specified risk factor in a theoretic population of 1,000 individuals.

^†The risk score is calculated by dividing the log of the odds ratio (OR) for the risk factor by the log of the OR of the factor with the lowest value in the model. For example, the risk score for multifetal gestation is calculated by dividing the log of the OR for multifetal gestation (2.83) by the log of the OR of the factor with the lowest value (anemia: 1.45) to produce a risk score of 3 (rounded from 2.8).

^‡Platelets less than 100,000/microliter.

The OR of the factor with the lowest value in the final model was anemia (aOR 1.5, 95% CI, 1.2–1.8). The numeric scores for individual risk factors ranged from 1 to 4, with a maximum possible score for an individual of 26 (ie, if all risk factors were present). When the model was applied to individuals in this cohort, total numeric risk scores ranged from 0 to 17, corresponding to a predicted risk for transfusion from 1.0% to 84.4%.

The AUC for the risk score in prediction of transfusion in the validation subsample was 0.81 (95% CI, 0.76–0.85). Using the full cohort, the AUC was 0.77 (95% CI, 0.74–0.79; Fig. 1). The corresponding estimated risk for transfusion based on the score index across the full range of potential scores was 1.0% (score=0) to 99.3% (score=26; Fig. 2).

Fig. 1. Receiver operating characteristic curve for the developed risk-scoring system to predict blood transfusion during hospitalization for delivery. The area under the receiver operating curve (AUC) was 0.77 (95% CI, 0.74–0.79) for the full cohort and 0.81 (95% CI, 0.76–0.85) for the validation subset.

Bruno. Blood Transfusion Risk-Scoring System. O&G Open 2025.

Fig. 2. Estimated blood transfusion risk (%) based on numeric risk score. The y-axis estimates the blood transfusion risk (%) by the x-axis numeric risk score for an individual (range 0–26). The minimum risk (score=0) is estimated at 1% and ranges to maximum risk (score=26) of 99.3%.

Bruno. Blood Transfusion Risk-Scoring System. O&G Open 2025.

DISCUSSION

In this secondary analysis of a U.S. cohort, we developed and internally validated a numeric risk score to predict blood transfusion during hospitalization for delivery. The final model includes 11 risk factors that may be assessed during hospitalization for delivery and produces a clinically applicable estimate of an individual's risk for blood transfusion. The model was internally validated with an AUC of 0.81, suggesting good predictive capability.

There are a multitude of postpartum hemorrhage and blood transfusion risk prediction models available for clinical use in the United States. The most widely used—the CMQCC tool—uses clinically available risk factors to produce categorical outputs (eg, low-risk, medium-risk, and high-risk) for transfusion, which allows for an easy-to-use interface for labor and delivery units and clinicians.⁴ However, the CMQCC tool, and others like it, were created using expert consensus and are limited to broad categories of risk rather than producing an individualized transfusion risk output. Although regression models for prediction of transfusion resolve some of these issues, they do not provide a user-friendly interface.^7–9 In contrast, our numeric risk score provides 11 easy-to-evaluate risk factors that may be assessed at hospitalization for delivery to produce an individualized risk estimate for transfusion.

For example, a hypothetical 26-year-old patient, G2P1001, at 39 2/7 weeks of gestation who presented for delivery with a history of one prior cesarean delivery and anemia would have a calculated risk score of 2 (1 point for prior cesarean and 1 point for anemia). This correlates to an estimated risk of blood transfusion of 2.1%. If labor became prolonged (longer than 24 hours) and chorioamnionitis developed, the individualized numeric risk score would increase to four (additional 1 point for prolonged labor and 1 point for chorioamnionitis) and a transfusion risk of 4.3%. The model balances the ease-of-use through straightforward risk factor assessment and a specific, practical, and individualized transfusion risk output.

This type of approach has been used based on data from a single institution.¹⁰ Magee-Womens Hospital developed an institution-specific risk prediction tool for blood transfusion using risk-factor weighting and laboratory results. In a retrospective cohort study of 21,843 deliveries at Magee-Womens Hospital, Hacker et al¹⁰ found the institution-specific weighted scale had moderate agreement with the CMQCC tool in predicting blood transfusion during delivery admission (Cohen's kappa 0.41, P<.001). This analysis extends this prior work by including nationally representative data from 17 institutions.

The present study has several limitations. Although broad laboratory findings were available from this data source (eg, thrombocytopenia), specific laboratory results (eg, exact platelet count or hemoglobin level) at time of hospitalization for delivery were unavailable. Accurate prediction of blood transfusion during hospitalization for delivery does not equate to the certainty of reduced maternal morbidity. Future work will need to inform whether the use of this risk model, or others, alters ability to predict or prevent maternal morbidity. The developed model applies to the hospitalization for delivery and does not extend to the risk for delayed postpartum hemorrhage or transfusion after hospital discharge. The primary outcome was any blood transfusion; therefore, the risk model does not predict volume (ie, number of units) of blood needed.

This study also has multiple strengths. We used a weighted risk score approach to develop a practical model that produces an individualized risk for transfusion during hospitalization for delivery. The model was internally validated. The GRAVID data were collected by trained perinatal research staff using medical record abstraction following a detailed protocol to define variables and outcomes improving accuracy and ascertainment of risk factors and outcomes. The GRAVID data set provided a contemporary, large, geographically and racially diverse cohort for the development of this model, increasing the generalizability of our findings.

In summary, we developed a numeric risk score tool that considers 11 easy-to-assess risk factors during hospitalization for delivery to produce an individualized predicted risk of blood transfusion regardless of mode of delivery. The model was internally validated with good predictive capability (AUC 0.81). Additional research is needed to externally validate the model and compare its clinical utility and performance with other models.