Introduction

The World Health Organization (WHO) defines preterm labor as deliveries before 37 weeks of gestation. Premature birth (PB) is one of the most common causes of infant mortality and morbidity [ 1 ], leading to a high mortality rate, and in those who survived acts as one of the most important causes of neurological complications in the infant [ 2 , 3 ]. The majority of infants born with PB would require extensive care from specialists in the early years of life, imposing substantial costs on families and the health care system [ 3 , 4 ]. An estimated 15 million infants are born preterm every year which accounts for more than 10% of the total deliveries [ 5 ].

Various methods were investigated to predict PB in women with symptoms of preterm labor, such as measurement of cervical length and fetal fibronectin [ 6 ] and measurement of cervical length and cervicovaginal fluid cytokine levels [ 7 ]. For instance, high levels of maternal salivary cortisol in women with a threatening episode of preterm labor between 24^th to 31^st weeks of gestation can be a predictor of preterm delivery [ 8 ]. However, none of these techniques can objectively and accurately predict the timing of labor and prematurity of an infant, mainly due to their low positive predictive power [ 9 ].

Prevention of PB before the onset of preterm labor symptoms can reduce morbidity and mortality as well as the cost of treatment in this group of infants [ 10 ]. Several methods were introduced, such as the use of progesterone or cervical cerclage [ 11 , 12 ] to prevent PB. However, it is essential to first identify high-risk individuals to apply these methods.

Prediction of the onset of preterm labor is challenging due to affection by many parameters. Previous work employed machine learning (ML) to predict preterm labor. For example, Catley et al. predicted PB based on eight risk factors using an artificial neural network (ANN), but the results were not promising (sensitivity=55%) [ 13 ]. Mas-Cabo et al. conducted an ANN-based prediction of PB using electrogastrographic records. However, this approach may not be practical due to the high expenses of electrogastrography tests [ 14 ]. Moreover, Wlodarczyk et al. applied a convolutional neural network (CNN) on ultrasound images to predict PB, but this method required a sonography procedure that may not be available [ 15 ]. Among studies [ 13 - 15 ] on an ML based prediction of PB, either their performance was not promising [ 13 ], or they have applied expensive procedures that are not available in many medical centers, reducing substantially their usability [ 14 , 15 ]. This study aimed to propose an ML-based method based on routine tests and procedures for PB prediction. Efficient prediction of PB causes physicians to apply necessary treatments for the prevention of this condition, thereby greatly reducing mortality and morbidity.

Material and Methods

Population study

For this historical cohort study, data were collected from pregnant women admitted to Mahdieh Obstetrics and Gynecology Hospital in Tehran from 2018 to 2019. During the two years, 150 women who had a preterm delivery (delivery before the 37^th week of gestation) with complete data were selected as the case group, and 150 women with natural delivery were randomly selected as the control group.

Data from pregnant women, including baseline demographic data and all information about previous pregnancies, were collected through interviews and questionnaires, physical examinations, and laboratory measurements. A total of 13 parameters, affecting the success of intrauterine insemination (IUI), including occupation, age, education, race, weight at the first trimester of pregnancy, body mass index (BMI), order of menses, parity, multi-parity, history of preterm labor in previous pregnancies, the interval between pregnancies, gestational diabetes, and blood type, were collected and selected as the ANN’s inputs. The current study was approved by the Ethics Committee of Shahid Beheshti University of Medical Sciences in Iran and all patients’ data were kept confidential. This was a historical cohort study and previously collected data (from routine medical tests for pregnant women) were obtained from patients’ files. Accordingly, no new procedure or test was conducted for this study.

ANN

An ANN classifier was designed with one hidden layer with 7 neurons, in which sigmoid activation functions were used in the input and hidden layer. However, a linear activation function was applied to the output layer. Pilot studies showed that the ANN outperformed linear and statistical classifiers, such as linear discriminant analysis (LDA) and Bayesian, respectively. The ANN was trained using the Levenberg-Marquardt algorithm. From the case and control groups, randomly, 15% of the subjects were selected for testing, 15% for validation, and 70% for training the ANN. The cross-entropy between the actual response and the ANN output was employed as the loss function. The ANN structure and training parameters were empirical (Figure 1).

Figure 1. The structure of the proposed artificial neural network (ANN) for premature birth (PB) prediction; the number of inputs is n=13 risk factors, the number of hidden neurons (triangular nodes) is 7, and the output node outputs the predicted class: normal (0) or PB (1).

Evaluation of the ANN performance

Training and testing of the model were repeated 100 times, and the average results were reported. The ANN performance was assessed using three parameters: classification accuracy, sensitivity, and specificity.

Results

The classification accuracy, sensitivity, and specificity of the test data were 79.03%, 73.45%, and 84.62%, respectively. Figure 2 shows the receiver operating characteristics (ROC) curve.

Figure 2. The receiver operating characteristics (ROC) curve of the proposed artificial neural network (ANN) for the test data

In the case and control groups, the majority of the subjects were housewives (about 94.3%). Moreover, 95.4% of the subjects in the case group and 74.3% in the control group were Iranian (the rest were Afghans). The mean weight at the first trimester of pregnancy in the case and control groups was 60.26±13.3 and 65.58±10.63, respectively. Menses was regular in 83% and 79.8% of the subjects in the case and control groups, respectively. In addition, 11.3%, 60.4%, and 28.3% of the subjects in the case group and 3.8%, 60.6%, and 34.9% in the control group were in the age ranges of under 18, 18-30, and above 30, respectively. In addition, 45.3%, 41.5%, and 13.2% of the subjects in the case group had their first to third pregnancies, respectively. However, in the control group, 34.9%, 33.9%, 16.5%, and 14.7% of the subjects gave birth for the first to the fourth time, respectively. Moreover, 94.3% and 97.2% of pregnancies in the case and control groups were single, and the rest were twins. Furthermore, 7.5% and 1% of the subjects in the case and control groups, respectively, had suffered from preterm delivery complications in their previous pregnancy. The mean distance between pregnancies in the case and control groups was 1.8 and 2.5 years, respectively. Gestational diabetes was not significantly different between the two groups (the difference was approximately 3.3%).

Discussion

In this study, an ANN was proposed to predict preterm delivery based on 13 risk factors. Several studies have previously investigated the prognosis of this complication. The advantage of the proposed approach is that it is non-invasive. In addition, the proposed method is very economical since no laboratory information was used. The classification accuracy of this study was 79.03%. However, increasing the number of subjects can significantly increase the performance of ML and therefore will be investigated in our future research. Moreover, the performance may improve by adding other risk factors as the ANN inputs, such as those suggested in [ 3 , 16 ], which will be also investigated in our future studies.

There is limited literature on the application of ML for the prediction of preterm delivery. Elaveyini et al. [ 17 ] achieved a classification accuracy of 70% for an ML-based prediction of PB, in which the low performance might be attributed to the small number of samples in their study (50 subjects). In 2019, Kwang-Sig Lee et al. [ 18 ] studied 504 pregnant women for ML-based prediction of preterm delivery with promising results, which were higher than that of the current study. However, as opposed to the current method, some of the risk factors [ 18 ] were noninvasive. Perhaps a larger statistical population and the involvement of additional risk factors as the ANN inputs, such as cervical length (which may have useful information for PB prediction) was the reason for the higher performance of the method in [ 18 ]. In addition to parameters, such as cervical length, some researchers have investigated the role of laboratory biomarkers in predicting preterm delivery [ 19 - 24 ]. The results of these studies indicate that their performance is not superior to the method presented in this research despite the imposition of cost and aggressive method.

The obtained results substantially outperformed those of Catley et al., in which 8 risk factors were used [ 13 ], due to the limited number of risk factors as well as the choice of risk factors. Mas-Cabo et al. used electrogastrographic records for PB prediction and achieved promising performance [ 14 ]. However, this approach is not practical, as it requires expensive lab tests, without any availability in many hospitals. Wlodarczyk et al. proposed using ultrasound images for CNN-based PB prediction [ 15 ], but despite the high computational complexity of this approach and the need for a sonography test, this method [ 15 ] achieved a lower sensitivity than that of the proposed approach.

Conclusion

This work proposed an ML-based approach for automated prediction of PB with the advantage of high practicality due to only 13 risk factors (as the model inputs) collected in a questionnaire, without the need for any lab tests. The proposed method can help physicians in the early prediction of PB and consequently may increase thinfant’s survival chance.

Acknowledgment

We would like to thank the participants and clinicians of this study.

Authors’ Contribution

M. Mirzamoradi conceived the idea and supervised the project. H. Mokhtari Torshizi analyzed the data and drafted the manuscript. M. Abaspour and A. Ebrahimi conducted data collection. A. Ameri supervised the project and edited the manuscript. All the authors read, modified, and approved the final version of the manuscript.

Ethical Approval

This research was approved by the ethics committee of Shahid Beheshti University of Medical Sciences in Iran (IR.SBMU.MSP.REC.1397.568).

Informed Consent

This was a historical cohort study and previously collected data (from routine medical tests for pregnant women) were obtained from patients’ files were used, and all patients’ data were kept confidential. Hence no new procedure or test was conducted for this study.

Funding

This research was supported by grant No. 13807 ().

Conflict of Interest

None

References

1. Sendeku FW, Beyene FY, Tesfu AA, Bante SA, Azeze GG. Preterm birth and its associated factors in Ethiopia: a systematic review and meta-analysis. Afr Health Sci. 2021; 21(3):1321-33. Publisher Full Text | DOI | PubMed
2. Beck S, Wojdyla D, Say L, Betran AP, Merialdi M, et al. The worldwide incidence of preterm birth: a systematic review of maternal mortality and morbidity. Bull World Health Organ. 2010; 88(1):31-8. Publisher Full Text | DOI | PubMed
3. Goldenberg RL, Culhane JF, Iams JD, Romero R. Epidemiology and causes of preterm birth. Lancet. 2008; 371(9606):75-84. Publisher Full Text | DOI | PubMed
4. Petrou S. The economic consequences of preterm birth during the first 10 years of life. BJOG. 2005; 112:10-5. DOI | PubMed
5. Liu L, Oza S, Hogan D, Chu Y, Perin J, Zhu J, et al. Global, regional, and national causes of under-5 mortality in 2000–15: an updated systematic analysis with implications for the Sustainable Development Goals. The Lancet. 2016; 388(10063):3027-35. DOI
6. Nikolova T, Bayev O, Nikolova N, Di Renzo GC. Comparison of a novel test for placental alpha microglobulin-1 with fetal fibronectin and cervical length measurement for the prediction of imminent spontaneous preterm delivery in patients with threatened preterm labor. J Perinat Med. 2015; 43(4):395-402. DOI | PubMed
7. Jung EY, Park JW, Ryu A, Lee SY, Cho SH, Park KH. Prediction of impending preterm delivery based on sonographic cervical length and different cytokine levels in cervicovaginal fluid in preterm labor. J Obstet Gynaecol Res. 2016; 42(2):158-65. DOI | PubMed
8. García-Blanco A, Diago V, Serrano De La Cruz V, Hervás D, Cháfer-Pericás C, Vento M. Can stress biomarkers predict preterm birth in women with threatened preterm labor?. Psychoneuroendocrinology. 2017; 83:19-24. DOI | PubMed
9. Euliano TY, Nguyen MT, Darmanjian S, McGorray SP, Euliano N, Onkala A, Gregg AR. Monitoring uterine activity during labor: a comparison of 3 methods. Am J Obstet Gynecol. 2013; 208(1):66e1-6. Publisher Full Text | DOI | PubMed
10. Frey HA, Klebanoff MA. The epidemiology, etiology, and costs of preterm birth. Semin Fetal Neonatal Med. 2016; 21(2):68-73. DOI | PubMed
11. Rundell K, Panchal B. Preterm Labor: Prevention and Management. Am Fam Physician. 2017; 95(6):366-72. PubMed
12. Choi SJ. Use of progesterone supplement therapy for prevention of preterm birth: review of literatures. Obstet Gynecol Sci. 2017; 60(5):405-20. Publisher Full Text | DOI | PubMed
13. Catley C, Frize M, Walker CR, Petriu DC. Predicting high-risk preterm birth using artificial neural networks. IEEE Trans Inf Technol Biomed. 2006; 10(3):540-9. DOI | PubMed
14. Mas-Cabo J, Prats-Boluda G, Garcia-Casado J, Alberola-Rubio J, Perales A, Ye-Lin Y. Design and assessment of a robust and generalizable ANN-based classifier for the prediction of premature birth by means of multichannel electrohysterographic records. Journal of Sensors. 2019; 2019(7):1-13. DOI
15. Włodarczyk T, Płotka S, Rokita P, Sochacki-Wójcicka N, Wójcicki J, Lipa M, Trzciński T. InMedical Ultrasound, and Preterm, Perinatal and Paediatric Image Analysis. Springer, Cham; 2020.
16. Yang L, Wang P, Jiang Y, Chen J. Studying the explanatory capacity of artificial neural networks for understanding environmental chemical quantitative structure-activity relationship models. J Chem Inf Model. 2005; 45(6):1804-11. DOI | PubMed
17. Lilliecreutz C, Larén J, Sydsjö G, Josefsson A. Effect of maternal stress during pregnancy on the risk for preterm birth. BMC Pregnancy Childbirth. 2016; 16:5. Publisher Full Text | DOI | PubMed
18. Elaveyini U, Devi SP, Rao KS. Neural networks prediction of preterm delivery with first trimester bleeding. Arch Gynecol Obstet. 2011; 283(5):971-9. DOI | PubMed
19. Lee KS, Ahn KH. Artificial Neural Network Analysis of Spontaneous Preterm Labor and Birth and Its Major Determinants. J Korean Med Sci. 2019; 34(16):e128. Publisher Full Text | DOI | PubMed
20. Huang L, Hou Q, Huang Y, Ye J, Huang S, et al. Serum multiple cytokines for the prediction of spontaneous preterm birth in asymptomatic women: A nested case-control study. Cytokine. 2019; 117:91-7. DOI | PubMed
21. Bachkangi P, Taylor AH, Bari M, Maccarrone M, Konje JC. Prediction of preterm labour from a single blood test: The role of the endocannabinoid system in predicting preterm birth in high-risk women. Eur J Obstet Gynecol Reprod Biol. 2019; 243:1-6. DOI | PubMed
22. Carlisle N, Chandiramani M, Carter J, Shennan AH. Reply: Evaluation of the quantitative fetal fibronectin test and PAMG-1 test for the prediction of spontaneous preterm birth in patients with signs and symptoms suggestive of preterm labor. J Matern Fetal Neonatal Med. 2020; 33(14):2505. DOI | PubMed
23. Radan AP, Aleksandra Polowy J, Heverhagen A, Simillion C, Baumann M, et al. Cervico-vaginal placental α-macroglobulin-1 combined with cervical length for the prediction of preterm birth in women with threatened preterm labor. Acta Obstet Gynecol Scand. 2020; 99(3):357-63. DOI | PubMed
24. Ijabi J, Moradi-Sardareh H, Afrisham R, Seifi F, Ijabi R. SKA2 gene - A novel biomarker for latent anxiety and preterm birth prediction. Eur J Obstet Gynecol Reprod Biol. 2019; 237:106-12. DOI | PubMed