OBJECTIVES:

Discovering, in real time, who are the pregnant women of a health care provider and what their probable trimester/month of pregnancy are is an important information to offer an integrated program of perinatal care. In this study, we propose a machine learning solution capable of reading a lot of clinical and demographic data from patients and discovering this information.

METHODS:

For a total of 61,179 patients (women aged 15 to 50 years), medical claim data were analyzed, such as: consultations, exams, procedures, medical specialties, materials used, complexity and emergency indicators, ICD-10 codes, open text fields (from various electronic medical records), age, place of care, among other clinical and demographic data. The model´s architecture was built from a supervised gradient boosting framework that uses tree based learning algorithms, called LightGBM.

RESULTS:

The database used for training and validating the model was well balanced, with 29,945 non-pregnant women (controls) and 31,234 pregnant women, all properly labeled with the pregnancy indicator and the month of pregnancy. In the test dataset, which corresponds to 20% of the database (12,236 women), the accuracy of the “pregnant/non-pregnant classifier” was 98%, with a sensitivity of 98% and specificity of 99%. To detect the probable month/trimester of pregnancy, the LightGBM framework was used in “regressor mode”, and we obtained an R Square value of 86.9, which indicates that 86.9% of the data fits the regression model. A standard deviation of plus or minus 1.2 month was chosen, for which the overall accuracy of the model was 83%.

CONCLUSIONS:

This study led to the development of a tool to identify pregnancy and probable trimester/month of pregnancy, using some healthcare plan administrative databases as input. The health plan intends to use this tool in its integrated perinatal care program, whose main objectives are to provide better monitoring and health care targeted at pregnants.