A Robust Machine Learning Predictive Model for

Maternal Health Risk

Lokesh Pawar Janvi Malhotra Astha Sharma

CSE Department CSE Department CSE Department
Chandigarh University Chandigarh University
Chandigarh University
Mohali,India Mohali,India
[email protected] Mohali,India
[email protected]
[email protected]
Diya Devanshi Vaidya
CSE Department CSE Department
Chandigarh University Chandigarh University
Mohali, India Mohali,India
[email protected] [email protected]

Abstract—The wellness of women throughout pregnancy, during pregnancy and childbirth are very highly preventable.
delivery, and the postpartum period is referred to as maternal There should be alternative care for pregnant women,
health. especially in not-so-developed areas [3]. There are several
Maternity morbidity and mortality are important health factors that influence perinatal complications. Studies
indices because they indicate the availability of maternal and suggest that the first trimester of pregnancy is the best stage
other healthcare resources. Hemorrhage, infection, increased to predict and prevent pregnancy related complications. For
blood pressure, unsafe abortion, and premature labor are the instance, increasing obesity in women of childbearing age
leading causes of maternal morbidity and death. Machine leads to increased risk of diseases such as gestational
learning algorithms play a significant role in determining diabetes, large for gestational age (LGA), fetal macrosomia,
maternity health risks. In this paper, Traditional Machine and hypertensive syndromes in pregnancy. Also, a pattern of
Learning algorithms are applied for Maternal health risk decreased birth rates is observed in the developed countries
prediction. The performance of models has the scope of
over the years, predisposing women to adverse pregnancy
improvement so we proposed A robust and more reliable
machine learning model that works in worst, average and best
outcomes.
synerios and returns the robust performance by considering all To address this problem, emerging field of computer science
scenario’s performance. The proposed Robust Model turns out like Machine learning is used for decision-making,
to be the most efficient robust model among all with an abatement of medical errors, improving accuracy in the
accuracy of 70.21% which is quite satisfactory as compare to
prediction of various pregnancy-related complications [4].
traditionally applied ML Models.
Machine learning techniques can infer relevant and
Keywords—Machine Learning, Robust Model, Maternal, important connections between data from various data sets
Health risk that are difficult to correlate [5]. Hence, machine learning
can be used for determination of maternal health risk at
I. INTRODUCTION initial stages and preventive measures can be taken at the
time, reducing the risk of pregnancy related complications.
An unimaginable joy is what women feel when she Using some modern tools and related techniques we can
brings a new life to this world. Every mother holds the right find out unseen patterns, derive meaningful information and
to feel this joy. But for many women in this world, this can make some important decisions. Using complex ML
moment either becomes scary. Two-thirds of all maternal algorithms, we can build predictive models [6]. Predictive
deaths are caused by excessive bleeding, infections, and high causal analytics is required if you want a model that can
blood pressure during pregnancy. Pregnancy-related forecast the likelihood of a specific occurrence in the future.
complications is the biggest cause of death among girls who If you lend money on credit, you are concerned about your
are 15 to 19 years of age. Because teen girls are still evolving clients' capacity to make future credit payments on time.
their body, they are at greater risk if they become pregnant. Here, you may create a model that can do predictive
Child brides are less likely to receive proper medical care analytics on the customer's payment history to anticipate
while pregnant or to deliver in a health facility, so they are whether future payments will be made on time [7]. Machine
more prone to pregnancy-related complications [1]. The learning algorithms are the wisest option if you have
maternal Mortality rate is very high. Approximately 295 000 transactional data from a finance organization and need to
women died during and by the time of pregnancy and develop a model to predict future trends. This comes under
childbirth in 2017. Most of these deaths (94%) occurred in the supervised learning paradigm. The term "supervised"
low-resource environments, and most of them were highly refers to the fact that you already have data on which you
preventable [2]. Every moment during pregnancy should be have to train your computers. A fraud detection model can
monitored to ensure proper fetal growth as to ensure a also be developed using historical evidence of unauthorized
harmless delivery. As already discussed, most of deaths transactions [8].

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

 `
Machine Learning is the branch of science that studies how features of dataset are Age, Systolic Blood Pressure,
computers can learn without being explicitly programmed. Diastolic BP, Blood Sugar, Body Temperature, Heart Rate
ML is among the most intriguing technologies that have and one attribute is Risk Level. All these factors
ever been encountered in the history of technologies. As the significantly contribute to maternal mortality. All the
name implies, it provides the computer with a feature that features of the Dataset are described below:
makes it more human-like: The potential to learn. Machine
learning is being employed actively now, maybe in many
more locations than one would imagine [9]. Machine
Learning is a branch of artificial intelligence that focuses on
machine learning and generating predictions based on its
experience. Paper proposes an artificial neural network
optimization technique which is used in reactive AODV
routing protocol.[10]It allows computers or devices to
generate data-driven judgments rather than being expressly
programmed to perform a specific task. These programs or
algorithms are created in such a manner that when they are
directed to fresh data, they adapt and gradually improve. To
develop a model, a Machine Learning algorithm is
developed using a training data set. When fresh input data is
presented to the ML algorithm, it predicts based on the
model [11]. The concept behind Machine Learning is that
you teach and Train Machines via means of feeding them
records and defining features. Computers learn, grow, adapt,
and expand themselves whilst they are fed with new and
applicable records, without counting on specific Fig. 1: Features in Dataset
programming. Without records, there is little or nothing that
Machines can learn. The Machine observes the datasets,
identifies styles in it, learns robotically from the behaviors, In this paper, we worked on this dataset, by using different
and makes predictions [12]. Machine learning algorithms algorithms like Decision table, NaiveByesUpdateable,
develop a model using sample data, referred to as training MultilayerPerceptron, J48, LMT, Random Forest, REPTree,
data, to make predictions or judgments without being Begging, etc.
explicitly programmed to do so. Machine learning
algorithms are utilized in a broad range of applications,
including medicine, spam filtering, natural language II. LITERATURE REVIEW
processing, and human computer interaction, when
developing traditional algorithms to execute the required We are focusing on maternal health risks in this area, and
tasks would be difficult or impossible [13]. work has already been done on this topic by numerous
researchers; we are just emphasizing this with the aid of the
In 2013, it was reported that 90% of the entire world's data Literature Review given below.
was collected in the previous two years. This field is
growing so rapidly and is helping many industries. This • Ahmed et al. in paper IoT Based Risk Level
field has immense importance in many fields, be it health Prediction Model For Maternal Health identified
care, business, finance, cybersecurity and many more [14]. the problem of Risk level of pregnant women.
It plays a vital role in the healthcare industry. Data Science Ahmed et al. came up with the solution and The
led to a more efficient health care industry. A vast amount of analysis was done with data mining and machine
data is now available via everything from EMRs to clinical learning algorithm with statistical approach and
databases which help to diagnose disease faster, practice predicition model working on backend risk factor
prevention, and find out some new treatment options [15]. It analyzer was used for demonstration with the use
is without a doubt one of the most intriguing technologies of DT algorithm.
available, focusing on the construction of system programs
in such a manner that they can obtain relevant data and • Ran Neiger et al. in paper Long-Term Effects of
utilize it to constantly improve themselves without the need pregnancy Complications on Maternal Health
for human participation. The potential for computers to identified the problem of Long term maternal
learn on their own makes them more human-like. It starts complications caused by pregnancy to improve
with facts or observations, which may include instructions women’s health and decrease such risks, both
or firsthand experience [16]. women themselves and the medical team caring for
them need to be aware of these risks. Multiple
The most basic requirement for performing classification interventions including diet modifications, weight
using machine learning model is collection of data. The data loss, and increased physical activity appear to be
for this problem statement is taken from UCI Machine effective in decreasing these risks.
learning repository. The dataset belongs to the area ‘life’ in
the collection. The dataset has a total of 1014 instances with • Eugenia Arrieta Rodr ́ıguez et al. in paper Early
6 features. Daffodil International University, Dhaka, Prediction of Severe Maternal Morbidity Ibero-
Bangladesh is the original source of the dataset [17]. The American Conference on Artificial Intelligence
data is collected with the help of IoT based risk monitoring (NOV 2016) identified the problem of Risk level
system from various hospitals, maternal health cares, for Severe Maternal Morbidity in patients during
community clinics of the rural areas of Bangladesh. The six pregnancy. He came up with the solution using of
this technique, it is expected to
 `
build a tool for risk identification or risk independent (but similar) dataset. To put it another way, a
classification of a patient having SMM. The goal is to resilient algorithm is one whose testing error is near to its
provide a timely and adequate attention to each patient training error [18]. As machine learning is used for more
depending. sensitive jobs and on noisier and noisier data, it is becoming
on the risk level to be determined using KNN,NN,DT, increasingly crucial that the algorithms we create for ML are
ANN and LR algorithms. robust to worst-case noise. Even if we are not interested in
pure statistical study, we can use the notion of resilience in
• Williams et al. in Paper The Effect of Maternal Obesity learning systems. Our deep learners can be protected from
on the Offspring identified the Risk of maternal obesity irrelevant and potentially misleading information if we feed
and came up with the solution which expected that these them robust estimators. [19] Recent research has shown that
approaches will help decrease risk in fetus, infants and machine learning algorithms are subject to adversarial
mother and these decreased risk will be salient for one's perturbations, demonstrating that a modest and human-
life invisible input perturbation can totally affect the model
output. As a result, many real-world applications face
• Akhan Akbulut et al. in his paper Fetal health status
prediction based on maternal clinical history using
substantial security risks, making it necessary to formally
machine learning techniques identified About 60–70% evaluate the robustness of machine learning models. [20]
of the anomalies can be diagnosed via ultrasonography, For the robustness of our model, we first check the
while the remaining 30–40% can be diagnosed performance of the model from 0 to 10 folds. After applying
after childbirth. With the use of SVM, RF, KNN, DT, the 10 folds we obtain our results for the worst case, best
MLP, BN, LDA and LR case and the average case scenarios. Then we apply the
feature selection model and after that we train the dataset.
A. Maintaining the Integrity of the Specifications We calculate the average of all the cases, then the output
obtained after this becomes robust and reliable.
The template is used to format your paper and style the
text. All margins, column widths, line spaces, and text fonts
are prescribed; please do not alter them. You may note
peculiarities. For example, the head margin in this template
measures proportionately more than is customary. This
measurement and others are deliberate, using specifications
that anticipate your paper as one part of the entire
proceedings, and not as an independent document. Please do
not revise any of the current designations.
III. PROPOSED METHODOLOGY

Fig.3: K-Fold Cross Validation (Robust Model)

IV. RESULTS AND DISCUSSIONS

In this section, the performance metrics of the proposed ML
models are beinganalyzed. Accuracy, Error, ROC (Receiver
Operating Characteristics), Precision, Recall, True Positive
Fig.2: System Model Rate, False Positive Rate are the performance metrics used
for model evaluation. It is observed that the performance
This section is focusing on building a robust model to
predict maternal health risk. This work is categorized into improved for our proposed model when compared to the
four phases. In the first phase, we identify the problem and traditional ones.
then to address the problem, dataset from UCI machine
learning repository is identified and required pre-processing i) Confusion Matrix: Confusion Matrix is a
is done on it. In the imported dataset important features diagrammatic explanation of the performance of a
needs to be prioritized so in the second phase we perform machine learning classification model. It is divided
feature selection on it by using the Ginni index and find out into 4 parts. These are TP (True Positive), TN (True
the prominent features of our dataset. In the third phase we
Negative), FP (False Positive), FN (False
train and test it with the traditional Machine Learning
algorithm. In the fourth phase we compare the performance Negative).
of traditional Machine Learning algorithm with proposed
Robust Model based on K-fold cross validation.
The robustness of model is a method to identify quality
that describes how effective it is when evaluated on a new
 `

TP FN

FP TN

Fig.4: Confusion Matrix

• True Positive: It is outcome of the model when it

correctly predicts positive class as positive. [21] Fig.6: Error
TP = TP
TP+FN .....(1) ROC: ROC curve is a fantastic way for evaluation basis of
• False Positive: It is outcome of the model when it classification basis. ROC is a probability curve. AUC (Area
incorrectly predicts negative class as positive. Under Curve) is an area under curve. The greater the AUC,
FP = FP the better the performance of the model.
TN+FP .....(2)

Accuracy: This evaluation parameter is used in

classification problems to tell the percentage of
correct predictions a machine learning model
made. Accuracy is believed to be the best
evaluation metric for a classification problem. [22]

Accuracy = Correct Predictions.100

Total predictions

Accuracy = TP+FN
TP+FP+TN+FN.....(3)

Fig.7: ROC

Precision: Precision is the outcome of accuracy of a

positively predicted class.
Precision = FP …..(5)
TP+FP

Fig.5: Accuracy

Error: Error=1-accuracy …..(4) Fig.8: Precision

Table 2: Performance Metrics (percentage split)

 `
BioAssays. Artificial Intelligence and Machine Learning in 2D/3D
Medical Image Processing, CRC Press (pp. 171-178 ).
5. Christine B. Williams, Kusaynyonon C. Mackenzie, Sheila
Gahagan,2014 .The Effect of Maternal Obesity on the Offspring. Clin
Obstet Gynecol. 2014 ; 57(3): 508–515. doi.org/10.1097/
GRF.0000000000000043.
6. Yukiko Washio& Mara Humphreys, 2018.Maternal Behavioral
Health: Fertile Ground for Behavior Analysis. PerspectBehav Sci,
doi.org/10.1007/s40614-018-0143-z.
7. Lokesh Pawar, Pranshul Agrawal, Gurjot Kaur, Rohit Bajaj, 2021.
Table 3: Performance Metrics (K-fold Cross validation) Elevate Primary Tumor Detection Using Machine Learning.
Cognitive Behavior and Human Computer Interaction Based on
Machine Learning Algorithm,2021 (pp. 301-313). John Wiley &
Sons, Inc.
8. Nicole Racine (PhD), Andre Plamondon, PhD, SheriMadigan, PhD,
SheilaMcDonald, PhD, SuzanneTough, PhD. ”Maternal Adverse
Childhood Experiences and Infant Development” PEDIATRICS,
Volume 141, number 4, April 2018: e20172495, Nov 22,2017.
9. Zhoupeng Ren, Jun Zhu, Yanfang Gao, Qian Yin. ”Maternal exposure
to ambient PM 10 during pregnancy increases the risk of congenital
heart defects: Evidence from machine learning models.” Elsevier
B.V.2018
V. CONCLUSION AND FUTURE SCOPE 10. L Pawar et. al. "Review of Various Optimization techniques in
MANET Routing Protocols" IJSTER,Vol 4, Iss. 8,2015.
In our proposed model, we convert the data (obtained after 11. AkhanAkbulut, EgemenErtugrul, VarolTopcu “Fetal health status
performing percentage split) into consistent data (data prediction based on maternal clinical history using machine learning
obtained after performing cross validation) by using feature techniques” Elsevier B.V.(2018)(pp. 0169-2607)
selection. In this we perform the feature selection by using 12. Sebastian Rauschert, Phillip E. Melton. ”Machine Learning-Based
DNA Methylation Score for Fetal Exposure to Maternal Smoking:
Ginni index. By doing this we get to know about the Development and Validation in Samples Collected from Adolescents
important features of the data. This helps in making our and Adults” 15 september, 2020(pp. 097003-1-9)
model more Robust and reliable by using the K-fold cross 13. William Caicedo Torres, Juan Carlos Martinez Santos “Early
validation testing. Prediction of Severe Maternal Morbidity Using Machine Learning
Te c h n i q u e s . ” I b e r o - A m e r i c a n C o n f e r e n c e o n A r t i f i c i a l
When the data was not pre-processed, the results come out Intelligence( NOV 2016), doi.org/10.1007/978-3-319-47955-2-22
to be unsatisfactory, so we noticed that there is a scope of 14. Dinesh Kumar, Anuj Kumar Sharma, Rohit Bajaj, Lokesh Pawar,
improvement hence after that we processed the data, then 2021. Feature Optimized Machine Learning Framework for
the results obtained were more satisfactory as compared to Unbalanced Bioassays. Cognitive Behavior and Human Computer
Interaction Based on Machine Learning Algorithm, pp. 167-178, John
the traditional algorithm used. Wiley & Sons Inc.
Although our results are better than those obtained by using 15. MursalDawodi, Tomohisa Wada, Jawid Ahmad Baktash
traditional algorithms and unprocessed data as we have used “Applicability of ICT, Data Mining and Machine Learning to Reduce
feature selection, but always there is a scope of Maternal Mortality and Morbidity: Case Study Afghanistan” , 2020.
improvement. 16. KS Betts, S Kisely,RAlati “Predicting common maternal postpartum
complications: leveraging health administrative data and machine
The Concentration of Performance evaluation can also learning.” 20 february, 2019(pp.702-703)
include other domains like operational research, which helps 17. Pankaj Rahi, Sanjay P Sood, Rohit Bajaj, Yogesh Kumar, 2021. Air
in considering two to three parameters on a single quality monitoring for Smart eHealth system using firefly
algorithm. If WSAW score comes out to be good then it optimization and support vector machine. International Journal of
Information Technology, Springer Singapore, volume 13, issue 5, pg
helps in providing strength to the model. The results of this 1847-1859.
algorithm are quite satisfactory but we can use other domain 18. Rebecca R.S, Clark PhD”Three machine learning algorithms and their
operational research for more improvement. So further work utility in exploring risk factors associated with primary cesarean
can be carried out by using this alternate method to improve section in low-risk women: A methods paper” Wiley(2021)
(pp.559-565)
the model.
19. Mario W.L. Moreira, Joel J.P.C. Rodrigues, Francisco H.C Carvalho,
NavinChilamkurti, Jalal Al Muhtadi, Victor Denisov. ”Biomedical
REFERENCES data analytics in mobile-health environments for high-risk pregnancy
outcome prediction” Journal of Ambient Intelligence and Humanized
1. Marzi Ahmed and Mohammod Abul Kashem 2020. IoT Based Risk Computing, Journal of Ambient Intelligence and Humanized
Level Prediction Model For Maternal Health Care InThe Context Of Computing, doi.org/10.1007/s12652-019-01230-4
Bangladesh. 2nd International Conference on Sustainable 20. S. A. Hannan, A. V. Mane, R. R. Manza, and R. J. Ramteke,
Technologies for Industry 4.0 (STI), 19-20 December, Dhaka, doi.org/ Prediction of heart disease medical prescription using radial basis
10.1109/STI50764.2020.9350320 function, in 2010 IEEE International Conference on Computational
2. Ran Neiger,2017. Long-Term Effects of Pregnancy Complications on Intelligence and Computing Research, pp. 16 (2010). DOI: 10.1109/
Maternal Health: A Review Paper. J. Clin. Med. 2017, 6, 76; doi.org/ ICCIC.2010.5705900
10.3390/jcm6080076, pp 2-16. 21. Hwang J-w, Park S-J, Oh S-y, Chang S-A, Lee S-C, Park SW, et al.
3. Say, L., Chou, D., Gemmill, A., Tunçalp, Ö., Moller, A.B., Daniels, J., The risk factors that predict chronic hypertension after delivery in
Gülmezoğlu, A.M., Temmerman, M. and Alkema, L., 2014. Global women with a history of hypertensive disorders of pregnancy.
causes of maternal death: a WHO systematic analysis. The Lancet Medicine 2015;94:e1747.
Global Health,2014 Jun ;2(6):e323-33. doi.org/10.1016/ 22. Albright CM, Ali TN, Lopes V, Rouse DJ, Anderson BL. The Sepsis
S2214-109X(14)70227-X. 2(6), pp.e323-e333. in Obstetrics Score: a model to identify risk of morbidity from sepsis
4. Lokesh Pawar, Anuj Kumar Sharma, Dinesh Kumar, Rohit Bajaj, in pregnancy. Am J ObstetGynecol 2014;211:39. e1-8.
2021. Advanced Ensemble Machine Learning Model for Balanced
`