A Novel and Precise Profiling Tool to Predict Gestational Diabetes

Abstract

Background:

There is a trend in healthcare for developing models for predictions of disease to enable early intervention and improve outcome.

Instrument:

We present the use of artificial intelligence algorithms that were developed by Gynisus Ltd. using mathematical algorithms.

Experience:

Data were retrospectively collected on pregnant women that delivered at a single institution. Hundreds of parameters were collected and found to have different importance and correlation with the likelihood to develop gestational diabetes mellitus (GDM). We highlight 3 of 29 specific parameters that were important in pregestation and in early pregnancy, which have not been previously correlated with GDM.

Conclusion:

This predictive tool identified parameters that are not currently being used as predictors in GDM, even before pregnancy. This tool opens the possibility of intervening on patients identified at risk for GDM and its complications. Future prospective studies are needed.

Introduction

The rate of severe maternal morbidity in the United States has been trending up, which in turn has increased the healthcare expenditures.¹ A population-based cross-sectional study on New York City deliveries demonstrated that severe maternal morbidity almost doubled the cost of delivery and, furthermore, resulted in approximately $83 million in excess over five years.² Recently, there has been many models and algorithms developed and tested in order to screen early for disease processes in obstetrics that lead to morbidity in hopes of providing early interventions. A systematic review found 17 studies on prediction models for gestational diabetes.³ Another review found a total of 68 prediction models for preeclampsia within 70 studies including over 400 000 participants.⁴ Many of these models, however, were derived from baseline maternal characteristics used as predictor of preeclampsia.⁴

There is a trend for use of prediction models and algorithms based on mega data. A unique, screening algorithm that uses mathematical formulas may be useful as it would take into consideration not only risk factors known to be associated with disease processes but other factors that may have not been studied. Thus, these algorithms would provide us with a more precise tool to assess the risk in an individual patient by finding the causality and importance of each condition. In this article, we demonstrate how using an artificial intelligence platform (Safe Pregnancy Artificial Intelligence—SPAI), and its algorithmic capabilities, can be used to predict various conditions by diagnosing and predicting personal and important risk factors. We reveal below how it identified some predictors for gestational diabetes.

Method

GynIsUs, Ltd. developed an innovative artificial intelligence platform (SPAI). We trained the algorithmic capabilities based on data from various sources including data elements from the admission notes, prenatal and outpatient notes, and laboratory values in the electronic medical records from deliveries that occurred between 2001 and 2018 at Maimonides Medical Center, Brooklyn, NY. The project was approved by the institutional review board. Most of the patients had prenatal care and delivered in our hospital.

Inclusion criteria were women who had at least one pregnancy and data before and during the pregnancy. Exclusion criteria were data elements with nulls and duplications as well as data that contained contradictions. Women with the diagnosis of pregestational diabetes mellitus (Diabetes Mellitus Type 1, Diabetes Mellitus Type 2) were also excluded.

Our primary outcome was gestational diabetes. Gestational diabetes was defined as an abnormal three-hour glucose tolerance test based on Carpenter and Coustan’s criteria.^5,6

The development process and algorithms involved many steps. A unique patient profile was created by each data point receiving a weight, which was based on a reliability calculator. Next, each of the parameters received an importance score that was calculated precisely based on mathematical models. Using this method, we found what the importance is for each parameter in the decision of developing gestational diabetes, and also checked the various degrees of the relationship between the parameters and the probability of developing gestational diabetes. The probability, causality, and the reliability results were then combined to provide a result for developing gestational diabetes.

There were several metrics that were used to estimate the quality of the algorithm: specificity, sensitivity, accuracy, and loss. Sensitivity is defined as true positives, whereas specificity is defined as true negatives. Accuracy is equivalent to success rate, i.e., how many times an algorithm predicts correctly if a disease will develop. Loss is a number indicating how accurate the algorithm’s prediction was on a single example. If the algorithm’s prediction is perfect, the loss is zero; otherwise, the loss is greater. The goal of training a model is to find a set of weights and biases at a minimum loss (<0.5), on average, across all examples.

Experience

After exclusion criteria, there were a total of 99 851 pregnancies among 60 960 women who delivered between 2001 and 2018. Among these pregnancies, 11 699 (11.7%) pregnancies were complicated with diabetes of which 6732 (57.5%) pregnancies were complicated with gestational diabetes. Through this predictive model, we were able to predict gestational diabetes with 91.7% accuracy.

Among the included cohort, 41.6% of patients were obese (defined as body mass index >30 kg/m²). In terms of race, the patient population was diverse. Of the 36 053 patients with race documented, 55.5% were white, 21.7% were Asian, 8.9% were black/African American, and 8.4% were Hispanic, with the remaining numbers being many other races such as Middle Eastern and Mid East Indian.

We looked at the predictive value for developing gestational diabetes out of hundreds of parameters ranging from classic parameters such as maternal age, parity, and vital signs to all laboratory tests. SPAI algorithms identified the most informative parameters to perform a prediction of gestational diabetes even from the preconception phase, as shown in Table 1. These combinations of new parameters may shed light on mechanisms that are involved in the biological process of developing gestational diabetes. Furthermore, we used SPAI to provide some of these biological explanations (Figure 1).

Table 1.

The Most Important Medical Tests to Predict GDM at the Preconception Phase that Found by SPAI Algorithms.

Abbreviations: GDM, gestational diabetes mellitus; SPAI, Safe Pregnancy Artificial Intelligence.

Figure 1.

Presents one of the possible biological explanations by SPAI and showed the interactions between the most important parameters (Table 1) that were found by SPAI algorithms. All the parameters were collected at the preconception phase at the time that the prediction was performed. The flow diagram shows one of the possible feedback circles beginning from the perspective of the medical test that may be led a patient to develop gestational diabetes mellitus. SPAI, Safe Pregnancy Artificial Intelligence.

There were other important factors noted when developing this algorithm. For example, the probability and relations between religion was evaluated (Table 2). Patients that identified with Jewish or Anglican religions had the lowest probability of developing gestational diabetes, while patients that identified with Jehovah Witness and Seventh Day Adventist had the highest probability of developing gestational diabetes. Religion was not addressed in the literature as a confounder associated with gestational diabetes mellitus, which we identified in our population using this model. (Table 3) Religion may serve as a surrogate marker for a patient’s cultural background, which may be important in the prediction of certain diseases.

Table 2.

Religion Parameter and Probability of Gestational Diabetes.

Religion	Total number of pregnancies (N)	Number of pregnancies with GDM (N)	Number of pregnancies without GDM (N)	GDM probability (%)
Hindu	143	24	119	16.8
Pentecostal	175	25	150	14.3
Seventh Day Adventist	92	13	79	14.1
Muslim	7611	1062	6549	14.0
Protestant	61	7	54	11.5
Buddhist	901	103	798	11.4
Baptist	288	32	256	11.1
Catholic	6703	703	6000	10.5
Christian	4455	381	4074	8.6
Jehovah Witness	110	9	101	8.2
Russian Orthodox	100	8	92	8.0
Anglican	51	2	49	3.9
Jewish	22 705	738	21 967	3.3

Abbreviation: GDM, gestational diabetes mellitus.

Table 3.

Specificity and Sensitivity on Two Religion Domains.

Religion	Total number of pregnancies (Na)	Specificity	Sensitivity	Accuracy	Loss
Catholic	6703 (703)	98.1%	88.0%	93.0%	0.0693
Jewish	22 705 (738)	99.3%	96.5%	97.9%	0.0207

^aNumber with gestational diabetes.

Another example evaluated whether there were any relations between a combination of parameters and the probability of developing gestational diabetes. We looked at the number of previous pregnancies combined with the history of gestational diabetes as well as the birth order (e.g., the difference of probability between a patient with three prior pregnancies with history of gestational diabetes in her first pregnancy to another patient with three prior pregnancies with history of gestational diabetes in her last pregnancy). The results are shown in Table 4. We found that the birth order of the history of gestational diabetes affects the probability of having gestational diabetes in the current pregnancy. For example, a patient entering her fourth pregnancy with two prior pregnancies complicated by gestational diabetes where her last pregnancy did not have gestational diabetes had a lower probability of developing gestational diabetes compared to another patient with the same number of pregnancies and history of pregnancies complicated by gestational diabetes, but the last pregnancy was complicated by gestational diabetes (20.7% vs 45.5%, respectively).

Table 4.

Domain Combination of Birth Order with History of Gestational Diabetes.

Number of pregnancies (N)	Number of prior pregnancies with GDM (N)	Last pregnancy with GDM	Total number of pregnancies (N)	Probability to develop GDM in current pregnancy (%)
3	0	0	7940	6.1
3	1	0	556	17.8
3	1	1	583	25.4
3	2	1	178	42.7
4	0	0	3802	5.1
4	1	0	359	18.1
4	1	1	226	22.6
4	2	0	29	20.7
4	2	1	101	45.5
4	3	1	35	57.1

Abbreviation: GDM, gestational diabetes mellitus.

Discussion

The purpose of this article was to highlight the potential of mathematical algorithms to identify important elements for prediction of a disease, particularly in gestational diabetes, which otherwise would not be identified. We found that there were hundreds of parameters that had different importance, likelihoods, relation, and interaction between parameters and the probability of developing gestational diabetes. As shown in Table 1, thyroid function tests and vitamin B12 were found to have high importance for the development of gestational diabetes. In a recent study, higher TSH was found to be correlated with higher A1C levels, and higher vitamin B12 levels were correlated with lower A1C levels.⁷

We found a high prevalence of diabetes in pregnancy within this population (11.7% vs 7% reported in the United States⁵). The high prevalence was an advantage for identifying important, predictive factors. Retrospectively, we were able to detect differences between the parameters and ranked their performance to be able to predict at preconception and early gestation the risk of developing gestational diabetes. However, we described only a few examples above as the full data training of this algorithm is beyond the scope of this article.

There are other predicting models described in the literature for gestational diabetes. Rasanen and colleagues evaluated the association of glycosylated fibronectin, c-reactive protein, adiponectin, sex hormone–binding globulin, and placental lactogen in a case control study of pregnant women with gestational diabetes. They found that glycosylated fibronectin was associated with gestational diabetes independent of maternal age, parity, gestational age, and other biomarkers.⁸ Other early biomarkers tested early in pregnancy have been evaluated and found to be associated with gestational diabetes.^9,10 A systematic review found 14 studies on predictive models for gestational diabetes, but only three were externally validated.³ These studies demonstrate novel laboratory tests that are associated with gestational diabetes; however, adoption may lead to higher healthcare costs. The mathematical, prognostic model described in this article does not require extra tests because it was based on routine care in pregnancy and prepregnancy, and used methods for data completion in the absence of data. Although almost half of pregnancies are unintended in the United States,¹¹ many of the important laboratory tests found were similar to tests performed during the first trimester.

Currently, in the medical field, only models have been used for prediction of disease. SPAI has an artificial intelligence that uses algorithms. The main difference between models and algorithms is that an algorithm is a mathematical technique derived by statisticians and mathematicians for a particular task, for example, prediction, while model is a set of strict rules to follow. In addition, SPAI has the capability of pulling and analyzing data not only from structured and discreet data elements, but also from narrative and free text data elements.

In conclusion, we conveyed the concept of mathematical algorithms to identify important predictive data elements, which may be beneficial in addition to traditional statistical probabilities. This is a promising strategy as these algorithms may decrease unnecessary medical tests that can decrease the number of complications, and in turn decrease healthcare costs related to unnecessary medical tests. Furthermore, precise early detection of gestational diabetes allows opportunities for early interventions that in turn will decrease morbidity. We are in the process of collecting data on prospective cohorts in order to validate the capability of SPAI to predict gestational diabetes at preconception and early pregnancy in order to enable early intervention such as diet and lifestyle modifications.