Prenatal Diagnosis Codes Predict Maternal Morbidity, Aid in Risk Stratification
Key findings
- Using data from a prospective cohort of 17,835 women, researchers at Massachusetts General Hospital studied the ability of ICD-10 diagnosis codes generated from outpatient prenatal encounters to predict severe maternal morbidity (SMM)
- Machine learning techniques were used to compare five methods of selecting which of 140 ICD-10 codes should be included in the prediction model
- The final model identified nine distinct ICD-10 codes that are relevant to the risk of SMM (e.g., iron deficiency anemia, abnormal placenta and prior uterine scar)
- In the testing set, women with a predicted probability of SMM in the highest decile had a rate of SMM, more than 3.5-fold higher than other women: 11.0% vs. 3.1% (P<0.001)
- In an external validation set of 11,013 women, those in the highest decile had a rate of SMM, more than double that of other women: 7.5% vs. 3.0% (P<0.001)
The rate of severe maternal morbidity (SMM) in the U.S. has more than doubled over the last 20 years. A public health strategy called regionalization of care is being promoted for pregnant women at high risk of complications so that they're directed to centers with appropriate capabilities for diagnosis and treatment.
Subscribe to the latest updates from OB/GYN Advances in Motion
However, most risk stratification tools used in obstetrics are based on information generated from the delivery encounter. That limits their ability to identify the best delivery location in advance.
Mark A. Clapp, MD, MPH, a specialist in the Maternal-Fetal Medicine Program in the Department of Obstetrics and Gynecology at Massachusetts General Hospital, Roy Perlis, MD, MSc, director of the Center for Quantitative Health in the Department of Psychiatry at Mass General, and colleagues have created a model for predicting SMM that incorporates ICD-10 codes generated during outpatient prenatal encounters. They describe it in the Journal of Perinatology.
Study Methods
The researchers studied a prospective cohort of 17,835 women who delivered viable infants at Mass General between July 1, 2016, and December 31, 2019. They randomly assigned 11,816 women (66%) to the training set and 6,019 (34%) to the testing set.
The team consulted electronic medical records and extracted ICD-10 codes generated from the prenatal clinic, laboratory and radiology visits. 140 codes were recorded at least once during the prenatal period and had an overall frequency between 1% and 20%. Machine learning techniques were used to compare five methods of selecting which of those 140 codes to include in the prediction model.
The Final Model
The model selected identified nine ICD-10 codes that predicted significantly greater adjusted odds of SMM:
- D50—Iron deficiency anemia (OR, 2.58)
- D68—Other coagulation defects (OR, 2.39)
- N97—Female infertility (OR, 1.98)
- O30—Multiple gestation (OR, 2.43)
- O34—Maternal care for abnormality of pelvic organs, including care for a uterine scar from previous surgery (OR, 1.46)
- O43—Placental disorders, including morbidly adherent placenta (OR, 2.30)
- O44—Placenta previa (OR, 2.71)
- R79—Other abnormal findings of blood chemistry (OR, 2.18)
- Z98—Other postprocedural states, including a uterine scar from previous surgery (OR, 1.49)
Model Performance
The model was similarly discriminative in the training and testing sets, with the area under the receiver operating curve (AUC) of 0.676 and 0.678, respectively. Women in the testing set who were in the highest decile of predicted probabilities of SMM had a rate of SMM more than 3.5-fold higher than other women: 11.0% vs. 3.1% (P<0.001).
The model was validated using data on 11,013 women who delivered at another large academic medical center in the same health system during the same time period. The AUC was 0.611. Women in the highest decile had a rate of SMM more than double that of other women: 7.5% vs. 3.0% (P<0.001).
367 of the 11,816 women in the training set (3.1%) had SMM. Another way to think of the model's performance is that the top decile provided a positive predictive value above that population-based figure of 3.1%:
- Positive predictive value—11% in the testing set and 7.5% in the validation set
- Negative predictive value—96.9% in the testing set and 97% in the validation set
Utility of the Model
If validated at other centers, this model could:
- In large academic medical centers—guide intrapartum management to reduce SMM in women at high risk (e.g., predelivery consultation with anesthesia, consultation with a maternal–fetal medicine specialist, extra care taken with communications among team members and during care transitions)
- In smaller or community-based hospitals—encourage referral of women at high risk to specialty care
- In all hospitals—replace providers' subjective assessment of SMM risk (which is subject to unconscious bias) with quantitative assessment, potentially reducing disparities in SMM risk assessment and treatment by race
view original journal article Subscription may be required
Learn more about the Maternal-Fetal Medicine Program
Refer a patient to the the Department of OB/GYN