A machine learning technology to improve the risk of non-invasive prenatal tests

Abstract

BACKGROUND:

Timely and accurate diagnosis of genetic diseases can lead to proper action and prevention of irreparable events.

OBJECTIVE:

In this work we propose an integrated genetic-neural network (GNN) to improve the prediction risk of trisomy diseases including Down’s syndrome (T21), Edwards’ syndrome (T18) and Patau’s Syndrome (T13).

METHODS:

A dataset including 561 pregnant were created. In this integrated model, the structure and input parameters of the proposed multilayer feedforward network (MFN) were optimized.

RESULTS:

The results of execution of the GNN on the testing dataset showed that the developed model can be accurately classify the anomalies from healthy fetus with 97.58% accuracy rate, and 99.44% and 85.65% sensitivity, and specificity, respectively. In the proposed GNN model, the Levenberg Merquident (LM) algorithm, the Radial Basis (Radbas) function from various types of functions were selected by the proposed GA. Moreover, maternal age, Nuchal Translucency (NT), Crown-rump length (CRL), Pregnancy-associated plasma protein A (PAPP-A) were selected by the proposed GA as the most effective factors for classifying the healthfetuses from the cases with fetal disorders.

CONCLUSION:

The proposed computerized model increases the diagnostic performance of the physicians especially in the accurate detection of healthy fetus with non – invasive and low – cost treatments.

Keywords

Trisomy syndromes optimization artificial neural networks genetic algorithms

1. Background

Anomalies disorders such as down’s syndrome (T21), Edwards’ syndrome (T18) and Patau’s syndrome (T13) are among the most common anomalies diseases among pregnant women [1, 2, 3, 4, 5]. T21 is a set of physical, psychological and functional anomalies, such as a change in a distinctive appearance accompanying a defect in the vital organs of the body, such as the heart and the immune system [6]. T18, the result of trisomy in the 18 chromosome, is the second common autosomal trisomy syndrome with an outbreak of 1 in 6000 to 1 in 8000 of the live birth [7]. T13 is also a severe clinical disease resulting from the trisomy of 13 human chromosomes, and in cases of T13 syndrome, 82% of the cases will not live more than one month, and 85% do not live more than one year [8].

Trisomy detection methods are divided into two classes of invasive and non-invasive methods during pregnancy [9]. Passive methods usually have higher risk and more accuracy to non-invasive treatments. In the non-invasive treatments, first the risk of infection is examined, and in the case of high risk infection, the pregnant suggests aggressive methods to more confidence. In fact, non-invasive treatments play a role in determining the risk of disease and the need to carry out aggressive methods. The screening is among the usual non-invasive methods to determine the risk of trisomy during pregnancy.

Age, the difference between paired gonadotropin concentrations and Pregnancy-associated plasma protein A (PAPP-A), during week 16 of pregnancy and the Nuchal translucency (NT), Crown-rump length (CRL), as obtained by fetal sonography, in the week 11 to 14 of pregnancy, as well as the fetus’s DNA sampling through their mother’s blood, reduces the risk of syndrome cases [10]. In the non-invasive treatments with low risk of intervention, in non-invasive methods, despite the low risk of intervention, they have a diagnostic error in determining the risk of disease; so the use of invasive methods such as Amniocentesis is used as a more effective method to examine the healthy pregnancy in clinical specialist labs [11]. It is possible to recognize the condition of the fetal health through examination and the existence of DNA Karyotype in amniotic fluid. Usually this type of diagnosis is performed in the second quarter of pregnancy. The advantage of this method is its high accuracy (over 99%) and the disadvantages are the leakage of amniotic fluid, abortions, infection, lack of samples and the risk of fetus damage. The rate of abortions is reported between 0.05 to 1 percent [12]. High cost and time-consuming presentation of results are the disadvantages of those methods. Also, due to the high cost of definitive diagnostic methods for genetic anomalies, it has been observed in many cases that due to the financial inability of couples to perform these methods, they decided to terminate the pregnancy based on the screening test results of the first trimester. Furthermore, this issue can reduce the fertility rate and increase the rate of aging of the population. Predictions show that Iran will be the third country in the world in terms of population aging rate after the UAE and Bahrain [13]. In addition, the anxiety of pregnant women during pandemic diseases such as COVID-19 in going to screening centers and stressing them has caused negative effects on their referrals for screening tests. This has increased the risk of diagnosing fetal genetic diseases [14, 15, 16].

The application of artificial intelligence technologies in the area of medicine and human health, which is one of the most essential problems in everyone’s life, has grown considerably in recent decades [17]. Today’s human civilization is seeing the appearance of stronger decision support systems, more accurate diagnostic tools for illnesses, and even robot surgery, due to artificial intelligence and machine learning methods. Artificial intelligence makes use of current learning algorithms as computational techniques to enhance performance or generate correct prediction [18, 19]. Researchers now see ANNs, which use machine learning techniques to solve difficult problems in medicine and health, as an accurate tool for diagnosis and classification [20]. Akbarian et al. obtained an average accuracy of 90.9% in their research utilizing 45 variables affecting the pregnancy outcomes of lupus ectopic moms by testing on just 149 pregnant women and using ANNs [21].

Katlan et al. examined 97 pregnant women considered high-risk in screening tests for fetal health using an ANN model [22]. Feed-forward neural network is also a popular ANN algorithm that was used to classify and diagnosis of diseases in several studies [23, 24, 25]. The differences in the mentioned studies are mainly on the type of ANNs and the number of layers and neurons, and the architecture of the ANNs.

An ANN’s success is determined by its design and architecture. Typically, ANN optimization is performed by manually tuning the variables that define the network’s structure [26]. There are no clear and accurate guideline for ANN optimization, which is often accomplished via repetition, trial and error. This is generally tough and time-intensive, and it does not always result in the intended outcome [26]. In recent years, there has been a surge in the usage of ANN design optimization techniques, particularly GAs and hybrid models [27, 28, 29, 30, 31, 32]. There are two types of functional basis-based optimization algorithms: techniques based on reinforcement learning (RL) and methods based on evolutionary algorithms (EA). In RL-based techniques, choosing a component of the ANN design is called an action, and a succession of action sequences leads to the development of an ANN. Finally, the accuracy of the constructed ANN is taken into account as a metric. The search for the optimum architecture in EA-based techniques is based on mutations and the reassembly of the ANN architecture component [31]. In a sense, the designs with the best performance are chosen and utilized to continue the development process.

GA is an evolutionary algorithm widely used to optimize machine learning systems and provide an almost optimal solution with a random search technique [32, 33]. This algorithm works by modelling chromosome proliferation and recombination capabilities and mutations in a process similar to what occurs in the meiotic stage of cell proliferation. An initial set of solutions is created randomly to solve a problem called a population. Each solution is considered a chromosome. Chromosomes make a new generation each time the recombination and mutation cycles are repeated, and the function of each generation of chromosomes is evaluated. Chromosomes whose phenotypes perform best in the ANN are considered the next generation’s parents. Thus, in the next generations, the population will evolve. After several generations, the population converges to the best chromosome, providing the desired architecture of the ANN [34, 35]. In this study, an intelligent model was achieved by using an ANN technique and a GA optimization capacities to identify the risk of trisomy of T13, T18 and T23 during first trimester pregnancy.

2. Method

The current research was carried out in three phases. The research dataset was generated in the first step, and then the beneficial variables in detecting prenatal anomalies were discovered. The intelligent model was then developed to predict the genetic anomalies, and finally, the constructed model was evaluated. To evaluate the reliability of the model performance, the proposed GNN model was implemented 10 times on the datasets. Therefore, the average results were reported in every experiments.

2.1 Dataset

The dataset in this study consists of 15 to 45 year-old pregnant women who were referred to the Genetics Laboratory Center in Ahvaz city, center of province in southwest of Iran, for screening tests between the 11th and 14th weeks of pregnancy. The required information of 561 pregnant was observed through the study of files in the center’s archives, including maternal age, race, smoking during pregnancy, maternal diabetes mellitus, maternal preeclampsia, fetal heart rate, crown rump length (CRL), fetal nasal bone, human chorionic gonadotropin (hCG), and pregnancy-associated plasma protein A (PAPP-A). Based on the information in the mothers’ files, pregnancy outcomes and fetal health status were also retrieved. Data was collected in February 2018 to July 2019 period. The data included the initial results from the analysis of the NIPT according to International Society of Prenatal Diagnosis (ISPD) protocol were at three scales of low, moderate and high risks. The pregnancy outcomes and fetal health status for moderate and high risk were followed to second trimester of pregnancy to perform the additional tests.

2.2 The selection of effective diagnostic factors

After extracting the diagnostic factors from the scientific resource, and in consultation with the obstetrician, maternal age, maternal weight, smoking, pre-eclampsia, fetal heart rate, CRL, NT, free beta placental gonadotropin (Free $\beta$ hCG), and PPAP-A were chosen as the acceptable input variables for the integrated GNN system. Smoking and preeclampsia were ruled out in all instances owing to the same quantity.

2.3 Intelligent neuro-genetic integrated model development

In this study, among several machine learning methods, an integrated ANN system was developed as the classification model for identifying three prenatal anomalies. Various researches have demonstrated that ANNs outperform other models in the field of medical diagnosis in the majority of situations [36, 37, 38, 39, 40].

2.4 ANN

The proposed ANN model in this study is a multilayer feedforward network (MFN) as a kind of supervised ANNs. The MFN have an output layer, an input layer, and multi-layer of processing neurons (middle layer) in which information move unilaterally from the input layer neurons to the output layer neurons. In other words, the output of one layer becomes the input for the next. A transmission function (activation function) is necessary in ANNs to transform input impulses to output signals. The transfer function governs how neurons work. There are various transmission functions for activating ANN neurons, each with its own characteristics such as whether it is linear or nonlinear and how it functions [39].

Every communication in the ANN is associated with a weight. The output of the ith neuron in the middle layer is obtained from the following formula [41]:

$\displaystyle h_{i}=\sigma\left({\mathop{\sum}\limits_{j=1}^{N}V_{ij}x_{j}+T_{% i}^{\textit{hid}}}\right)$ (1)

Where $\sigma$ is the transfer function, $N$ is the number of input neurons, $V_{ij}$ weights, $x_{j}$ the input values of the input neurons and third are the initial conditions of the middle layer neurons.

In this work, to use the most appropriate transfer function for the input and middle layers, 10 types of transfer functions have been used (Table 1). It should be noted that in order to create an optimal ANN, the selection of appropriate transmission functions for each layer were done by a GA.

As mentioned, this network has multi-hidden layers. The number of neurons in each of these layers was considered to be between 1 and 20 neurons. In order to train the ANN, 10 types of training functions and 15 types of learning functions were used (Table 1). To create the optimal ANN structure, the training function, and the appropriate learning function were optimized by the GA.

ANN output is also a system for diagnosing healthy cases from genetic disorders T21, T18, and T13.

2.5 Genetic algorithm

The parameters that organize the ANN structure and the input factors were defined by haploid chromosomes containing the genes including the number of neurons in the hidden layers, transfer functions, training functions, learning functions and the presence or absence of one of the seven ANN input factors. Figure 1 shows the structure of the GA and its operation to optimize the ANN prediction performance in the GNN model. Furthermore, Table 1 shows the transfer and training functions for the hidden layers and the output layer to be used by the GA in the integrated GNN model.

Table 1
ANN training and transferring functions used in the GA optimization process

Transfer function graph

Transfer function

Training function

Hyperbolic tangent sigmoid transfer function (tansig)

Levenberg-Marquardt backpropagation (trainlm)

Log-sigmoid transfer function (logsig)

Scaled conjugate gradient backpropagation (trainscg)

Linear transfer function (purelin)

Resilient backpropagation (trainrp)

Saturating linear transfer function (satlin)

Conjugate gradient backpropagation with Powell-Beale restarts (traincgb)

Symmetric saturating linear transfer function (satlins)

Conjugate gradient backpropagation with Fletcher-Reeves updates (traincgf)

Positive linear transfer function (poslin)

Conjugate gradient backpropagation with Polak-Ribiére updates (traincgp)

Symmetric hard-limit transfer function (hardlims)

One-step secant backpropagation (trainoss)

Hard-limit transfer function (hardlim)

Gradient descent with momentum and adaptive learning rate backpropagation (traingdx)

Triangular basis transfer function (tribas)

Gradient descent with adaptive learning rate backpropagation (traingda)

Radial basis transfer function (radbas)

Gradient descent backpropagation (traingd)

In the GA chromosome evaluation, the existing data was trained by the ANN and each created ANN was evaluated by the mean squared error (MSE) through Eq. (2).

$\displaystyle\textit{MSE}=\frac{1}{n}\mathop{\sum}\limits_{i=1}^{n}({T_{i-}O_{% i}})^{2}$ (2)

Where $n$ is the number of studied cases, $O$ is the output of case $i$ , and $T$ is the expected target for case $i$ . Chromosomes were considered as higher-score which their decoding resulted in the networks with lower MSE and greater accuracy. With a certain number of iterations of the generation cycle, the ANN with the lowest MSE was finally determined as the final optimal network. The process of finding the optimal ANN structure by the GA is shown in Fig. 1. Figure 1 displays the process of GA including input parameters selection, training functions optimization and classification of trisomy disorders from the health fetus in the ANN model. Moreover, the description of GA parameters is shown in Table 2.

Figure 1.

The process of GNN optimization for prenatal screening test.

Table 2

Description of genetic algorithm parameters

GA parameters	Method
Population generation	120 randomized chromosomes
Parent selection	Rolette wheel method
Length of chromosome	14 genes
Crossover intersection	Two points
Probability ratio of two-point intersection	0.6
New generation strategy	Combination of offspring and current population

2.6 Data analysis

Three indicators evaluated the performance of the generated model; sensitivity, specificity, and accuracy, which are calculated using confusion matrix data. This matrix is commonly used to evaluate the performance of classification systems [42, 43, 44].

Sensitivity is the ability of a test to identify patient cases correctly and can be calculated with the following equation:

$\displaystyle\textit{Sensitivity}=\frac{TP}{TP+FN}$ (3)

Specificity is the ability of a test to identify healthy cases correctly and is obtained from the following relation:

$\displaystyle\textit{specificity}=\frac{TN}{TN+FP}$ (4)

Accuracy is the ratio of correctly identified patient cases to all patient cases, which is defined by the following equation:

$\displaystyle\textit{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN}$ (5)

3. Results

The results of an implementation of the integrated GNN model in the classification of healthy cases and cases with genetic abnormalities T21, T18, and T13 are presented as a confusion matrix (See Fig. 2).

To find the reliability of the model performance, the average of 10 times run of the optimum model was calculated. The results showed that on average, the designed model was able to correctly classify healthy and genetic disorders with 99.11% accuracy, with MSE equal to 0.89% for training set. In the test data, the accuracy was 97.58%, the specificity was 99.44% and the sensitivity in the best result was 92.60% and MSE was 2.43%. The results for test data, training data, and total data are shown separately in Fig. 2. Sensitivity, specificity, accuracy, and MSE of 10 times of expectation of the integrated GNN model are given in Table 3. Moreover, a snapshot of training performance of the GNN model in the experiment process is shown in Fig. 3. Furthermore, the comparison of training performance of the integrated model and the execution of the ANN model were showed in the Fig. 4. In addition, the comparison of performance of the ANN and GNN models is presented in Table 4.

Table 3
Performance of 10 times of execution of the GNN model

Iteration	Sensitivity (%)	Specificity (%)	Accuracy (%)	MSE (%)
1	85.16	99.81	99.11	0.89
2	88.89	100	99.47	0.53
3	81.48	100	99.11	0.89
4	88.89	99.63	99.11	0.89
5	81.48	100	98.93	0.89
6	81.48	99.81	99.28	1.07
7	92.60	99.63	98.56	0.71
8	85.19	99.25	99.11	1.43
9	85.19	99.81	99.29	0.89
10	86.2	100	100	0.71

Table 4

Performance of 10 times execution of the ANN in comparison with GNN model on the testing dataset

Model	Sensitivity (%)	Specificity (%)	Accuracy (%)	MSE (%)
ANN	80.42	91.71	88.30	12.70
GNN	85.65	99.44	97.58	2.43

3.1 Input factors

Data related to 10 times of execution of the developed model showed that with the optimization performed by the GA, the highest repetition of the presence of input factors related to maternal age, NT, and PAPP-A was observed in all model experiments. CRL, FHR, Free b-hCG, and maternal weight were also in the next ranks with 9, 7, 5, and 3 repetitions, respectively. The results for the number of iterations of the input parameters are shown in Table 5.

Table 5
Number of iterations of neural network input factors by GA

Input factors	Maternal age	Maternal weight	FHR	CRL	NT	Free b-hCG	PAPP_A
No of repeats	10	3	7	9	10	5	10

Figure 2.

Model performance for testing, training, and total datasets.

Figure 3.

A snapshot of training performance of the GNN model.

Figure 4.

A snapshot of training performance of the GNN and ANN model on the training dataset.

3.2 Transfer, training, and learning functions

Based on the GA selection, the radbas function with 4 repetitions and the logsig function with three repetitions had the highest presence in 10 model runs among the transfer functions used in this study. The tribas and purelin functions were performed twice and once, respectively. The logsig and tribas functions were both presents with three repetitions in the second layer. With two repetitions, the radbas function came next. Tansig and poslin functions were also run once each. The logsig function was repeated six times in the output layer, while the satlins and tansig functions were repeated three and one times, respectively.

The trainlm (LM) function was the most prevalent among the training functions used in this study, with 7 repetitions. The traincgb, traincfg, and trainrp functions were all run at once in MATLAB environment.

With four repetitions, the learnsom function had the most presence among the learning functions. The learnp and learnlv2 functions were repeated twice, and the learnis and learnk functions were repeated once each. Table 6 shows the results for the most repetition in the transfer, traning, and learning functions.

Table 6
Selective parameters of neural network structure by GA

	First layer’s transfer function	Second layer’s transfer function	Output layer’s transfer function	Training function Learning function
Most used function	Radbas-logsig-	Tribas-tansig	Logsig-satlins	Trainlm	Learnsom
No. of iterations	4-3	3-3	6-3	7	4

4. Discussion

In the realm of prenatal abnormalities detection, machine learning methods, particularly ANNs, have produced promising results [22, 35]. The optimum ANN architecture, on the other hand, may properly illustrate an ANN’s ability for early diagnosis.

This study demonstrated that an optimum ANN design for the early diagnosis of fetal anomalies may be obtained utilizing a GA.

Trimester screening facilities usually categorize the results as low, intermediate, and high-risks. The low-risk cases rate is defined 1 than 1100 individuals, cases in the high-risk rate is defined 1 than 50 people, and cases in the intermediate risk are between these two groups. Examination of the data used in this study revealed that none of the subjects were classified as low-risk when the screening test by common methods was used. Moreover, the diagnosis ratio of persons with genetic anomalies investigated in this study by common methods was 1.38 percent in instances with an intermediate risk diagnostic and 22.64% in high-risk diagnoses. Table 7 displays data on the number of cases diagnosed in risk regions, as well as the number of healthy and diseased individuals.

Table 7
Results of trimester screening test by common methods in different levels of genetic disorder risk

Rate of genetic abnormalities prediction (%)	Number of identified genetic disorder cases	Number of identified healthy cases	Risk range
0	0	237	Low-risk
38.8	3	215	Intermediate-risk
18.8	24	82	High-risk

As demonstrated in Table 6, there is a significant number of healthy patients at the intermediate and high-risk levels. Individuals identified in these two groups, particularly the high-risk group, by physicians and genetic counselors to perform other more definitive identification methods such as Noninvasive prenatal testing (NIPT), Chorionic villus sampling (CVS), and amniocentesis were referred to further ensure the outcome of pregnancy, which imposes costs and increases the risk of injury. In healthy situations, they urge abortion and damage to the mother.

Similar research has been conducted in recent years to identify Aneuploidy problems using machine learning and computational techniques [45]. ANN models were proposed for predicting pregnancy outcome using data from several pregnant datasets. Architecture with neurons in the input, the middle, and the output layers were accomplished by utilizing effective variables while the ANNs structure were tuned by trial and error technique to maximize network performance. The best result of applying the ANN developed in the testing set were showed 99.8% accuracy rate [9, 35, 46]. Furthermore, Soleimani et al. created a model based on an ANN. The proposed model was compared to the logistic regression model in the study. Determinants in predicting movement abnormalities in these infants were identified using a 14-factors coefficient analysis method. Finally, comparing the results of the two models revealed that the specificity of the ANN and regression model were 88% and 85%, respectively. Moreover, the specificity of ANN and regression models were 31.7% and 18.3%, respectively [47].

In another study, creating an adequate ANN for detecting prenatal genetic disorders was investigated with three big data sets, each with 51,001, 29,999, and 42329 records, respectively [9]. Various experiments were carried out on the given data in this study in order to determine the optimal ANN design. The overall anomaly detection rate in three data sets were 73.7%, 94.1%, and 85.2%, respectively [9]. Furthermore, the usage of feedback ANN models in the classification of Pataw, Edwards, Down, Turner, and Kleinfelter syndromes performed better than the feedforward ANN with the same structure in the intermediate layer with the accuracy of 98.8% [34]. Jamshidnezhad et al., in the similar study showed that the GA has a proper performance to optimize the ANN architecture for diagnosis of Down syndrome with the average accuracy, specificity and sensitivity rate of 99.91%, 99.72% and 90.91%, respectively [40].

The average results obtained in the current study from the model’s implementation revealed that the developed model, with an accuracy of 99.11%, specificity of 99.79% and sensitivity of 85.56%, and finally a mean square error of 0.89%, has a good performance compared to previous studies, particularly in diagnosing healthy cases from abnormal cases of Pataw, Edwards, and Down syndromes. The performance of the proposed hybrid model in comparison with the Jamshidnezhad et al. work showed the lower sensitivity rate of current study than the last one, while the current study implemented for diagnosis of three anomalies [40].

Moreover, when a fetus is diagnosed with a genetic disease, the parents usually elect to terminate the pregnancy. As a result, the function of specificity in comparison with sensitivity in proper decision-making appears to be more significant, according to definitions of sensitivity based on identifying the true instances of the disease and the specificity that identifies the true healthy cases. Because if a healthy fetus is misidentified as a patient and decides to terminate the pregnancy, an irreversible error will be made. As the suggested model has greater specificity than sensitivity, it may be utilized as a supplementary diagnostic technique in conjunction with other diagnostic methods.

4.1 Input factors

The developed model’s results revealed that maternal age, NT, PAPP-A, and CRL variables with the highest repetition in 10 times of execution had the most effects on the identification of genetic disorders. As a result, some characteristics may be more significant than others in diagnosing cases of genetic diseases. Necleus et al. employed ANNs to investigate the relevance and sensitivity of several factors used to predict chromosomal abnormalities. Separate testing for the HCG, PAPP-A, and CRL instances revealed that they contributed around 6% to each improvement through increasing system performance. Other factors, such as maternal age and the existence of the nasal septum and venous duct, each contributed 3%, 2.6%, and 0.8% to the system’s improvement, respectively [48].

The study shows that the NT and PAPP-A variables are more essential than other factors, but it also cites the age of the mother, which was one of the most commonly utilized input elements in the model, as a low-impact factor. As a result, additional research into the influence of these parameters in the identification of genetic disorders can help to improve the design of diagnostic models.

Trainlm according to LM function was the most popular function selected by the GA among the training functions in this study. Trainrp, traincfg, and traincgb are all other functions which were ranked in the lower position. Given this issue, it is predicted that limiting the options to these four functions and eliminating the GA’s examination of additional training functions will enhance the implementation speed of such models.

Among the learning functions in MATLAB environment, the learnsom function had the most success in selection via the GA, with 4 repeats. The learnsom function operates by recognizing chosen neurons and adjusting the weights of these neurons and other neurons in their proximity in order to come closer to the input vector. Similarly, Zeinley et al., recognize that in the implementation of a ANN with an intermediary layer, two learnsom and learngdm functions with four and three repetitions, respectively, are among the best learning functions in a study comparing ANN learning functions in runoff modeling [49].

4.2 Genetic algorithm

Recent studies that have used GAs to improve the structure of the ANN have focused on the number of hidden layers and neurons in these layers, and ultimately the number of neurons in the ANN input layer [50, 51, 52, 53]. However, in this study, the focus was on the ANN settings, ranging from the number of neurons in the hidden layers and the input layer to the kind of teaching and learning functions of each layer all were chosen using a the proposed GA.

In this study, because of the number of variables on each chromosome and the range of their modifications, the population of each generation of 120 chromosomes was evaluated in order to assess more genetic variety and, as a consequence, more options to reach the desired outcome. The number of repetitions of the generation production cycle was established by trying 50, 40, 30, 20, and 10 repetitions, eventually was settled on 20 repetitions. Because, repeating the generation production cycle more than 20 times not enhanced the outcomes while taking more processing time. In several comparable experiments, using smaller populations and more repeated generation cycles resulted in better performance [53, 54, 55]. However, it did not appear to be the case in this investigation. Furthermore, the number of variables in the cited studies, as well as the range of changes in these variables, are significantly lower than the variables in this research.

4.3 Screening during the COVID-19 pandemic

It is critical that expectant moms are referred to fetal screening test facilities on time and on schedule for diagnosis and counseling on fetal health [15]. Numerous studies demonstrate that during the COVID-19 pandemic, the proportion of pregnant women referred to screening facilities fell considerably [14, 15, 16]. Studies have also revealed higher stress in pregnant women conducting COVID-19 screening tests, which has resulted in fewer visits to fetal screening centers throughout pregnancy, particularly in the second and third weeks [16, 56]. As previously stated, according to Table 6, more than 38% (218 cases in 561 pregnant) and 18% (106 cases in 561 pregnant) of persons referred to laboratories for diagnosis and screening of fetal chromosome disorders, were identified with a moderate and high risk of genetic abnormalities, respectively, who need to the additional screening tests. As a consequence of the analysis of early screening tests using current technologies, expectant mothers are more likely to be sent to screening centers for additional testing and more accurate examination of the initial results. Some of these extra references are the result of computational and diagnostic mistakes in the first screening tests. The performance of the model, particularly in the accurate detection of healthy fetal in this study with an average specificity rate of over 99 percent, might avoid pregnant women from making many visits to screening facilities to verify the diagnosis.

5. Conclusion

The goal of this study was to develop an intelligent model for more accurate diagnosis of healthy instances of genetic disorders through screening tests using ANN method and GA optimization capabilities. The results of this study demonstrated that using GA to create the topology of the ANN have the proper performance of distinguishing healthy people from those with fetal abnormalities using information from first trimester screening tests. Moreover, the GA determined that the maternal age, PAPP_A and NT are the most effective factors on the identification of genetic disorders. Using this sort of integrated GNN prediction models as intelligent computer approaches can be utilized to reduce the risk of false positive error in the existing methods to help clinicians properly detect healthy fetal. Additional future studies are suggested on the performance of intelligent hybrid models such as PSO utilizing various datasets and comparing the findings with those of existing approaches.

Footnotes

Acknowledgments

The authors would like to thank the Fertility, Infertility and Perinatology Research Center of Ahvaz Jundishapur University of Medical Sciences and the vice chancellor of Research Affairs of Ahvaz Jundishapur University of Medical Sciences for their financial and administrative support to undertake this project. Research project No: FIRC-9709. Research ethics code: IR.AJUMS.REC.1397.826.

Conflict of interest

None to report.

References

Seyyed-Kavoosi

Younessi

Farhud

. Screening of fetal chromosome aneuploidies in the first and second trimester of 125,170 Iranian pregnant women. Iranian Journal of Public Health. 2015; 44(6): 791.

Ahangari

Doosti

Ahangari

BaradarnRafiee

GhayoorKarimiani

. Prenatal diagnosis of fetal aneuploidies using QF-PCR in 333 cases. Molecular Medicine Journal. 2016; 2(1): 38-42.

Neocleous

Schizas

Biehl

Petkov

, eds.Marker selection for the detection of trisomy 21 using generalized matrix learning vector quantization. 2017 International Joint Conference on Neural Networks (IJCNN); 2017 14–19 May 2017.

Vatankhah

Jalilvand

Sarkhosh

Azarmi

Mohseni

. Prevalence of congenital anomalies in Iran: A review article. Iranian Journal of Public Health. 2017; 46(6): 733.

Dolk

Loane

Garne

. The prevalence of congenital anomalies in Europe. Advances in Experimental Medicine and Biology. 2010; 686: 349.

Nanjiani

Hossain

Mahgoub

. Patau syndrome. J Neuropsychiatry Clin Neurosci. 2007; 19(2): 201-2.

Satgé

Nishi

Sirvent

Vekemans

. A tumor profile in Edwards syndrome (trisomy 18). American Journal of Medical Genetics Part C: Seminars in Medical Genetics. 2016; 172(3): 296-306.

Duarte

Menezes

Devens

Roth

Garcias

Martino-Roth

. Patau syndrome with a long survival. A case report. Genet Mol Res. 2004; 3(2): 288-92.

Neocleous

Nicolaides

Schizas

. Intelligent noninvasive diagnosis of aneuploidy: Raw values and highly imbalanced dataset. IEEE Journal of Biomedical and Health Informatics. 2016; 21(5): 1271-9.

10.

Nicolaides

. Screening for chromosomal defects. Ultrasound Obstet Gynecol. 2003; 21(4): 313-21.

11.

Wax

Cartin

Chard

Lucas

Pinette

. Noninvasive prenatal testing: Impact on genetic counseling, invasive prenatal diagnosis, and trisomy 21 detection. Journal of Clinical Ultrasound. 2015; 43(1): 1-6.

12.

Seeds

. Diagnostic mid trimester amniocentesis: How safe? American Journal of Obstetrics and Gynecology. 2004; 191(2): 607-15.

13.

Yavari

Basakha

Sadeghi

Naseri

. Economic aspects of ageing. Iranian Journal of Ageing. 2015; 10(1): 92-105.

14.

Farrell

Pierce

Collart

Craighead

Coleridge

Chien

Perni

Frankel

Ranzini

Edmonds

Rose

. The impact of the emergence of COVID-19 on women’s prenatal genetic testing decisions. Prenatal Diagnosis. 2021. doi: 10.1002/pd.5902.

15.

Deprest

Choolani

Chervenak

Farmer

Lagrou

Lopriore

McCullough

Olutoye

Simpson

Van Mieghem

Ryan

. Fetal Diagnosis and Therapy during the COVID-19 Pandemic: Guidance on Behalf of the International Fetal Medicine and Surgery Society. Fetal Diagn Ther. 2020; 47: 689-698. doi: 10.1159/000508254.

16.

Ozalp

Demir

Akbas

Kaya

Celik

Osmanagaoglu

. Effect of COVID-19 pandemic process on prenatal diagnostic procedures. The Journal of Maternal-Fetal & Neonatal, Medicine. 2020. doi: 10.1080/14767058.2020.1815190.

17.

Kabootarizadeh

Jamshidnezhad

Koohmareh

. Differential Diagnosis of Iron-Deficiency Anemia from β-Thalassemia Trait Using an Intelligent Model in Comparison with Discriminant Indexes. Acta Inform Med. 2019; 27: 78-84. doi: 10.5455/aim.2019.27.78-84.

18.

Jayalakshmi

Santhakumaran

, eds. A Novel Classification Method for Diagnosis of Diabetes Mellitus Using ANNs. 2010 International Conference on Data Storage and Data Engineering; 2010 9–10 Feb. 2010.

19.

Karabatak

Ince

. An expert system for detection of breast cancer based on association rules and neural network. Expert Systems with Applications. 2009; 36: 3465-9.

20.

Drew

Monson

JRT

. Artificial intelligence for clinicians. Journal of the Royal Society of Medicine. 1999; 92(3): 108-9.

21.

Akbarian

Paydar

, ostam Niakan Kalhori SR Sheikhtaheri

. Designing an ANN for prediction of pregnancy outcomes in women with systemic lupus erythematosus in Iran. Tehran University Medical Journal. 2015; 73(4): 251-9.

22.

Katlan

Tanacan

Orgul

Leblebicioglu

Beksac

. An Intelligent Prenatal Screening System for the Prediction of Trisomy-21 Using Triple Test Variables: The Hacettepe System. Gynecology Obstetrics & Reproductive Medicine. 2019; 25(2): 67-9.

23.

Drew

Monson

JRT

. Artificial intelligence for clinicians. Journal of the Royal Society of Medicine. 1999; 92(3): 108-9.

24.

Karabatak

Ince

. An expert system for detection of breast cancer based on association rules and neural network. Expert Systems with Applications. 2009; 36: 3465-9.

25.

Goodfellow

Vinyals

Saxe

. Qualitatively characterizing neural network optimization problems. arXiv preprint arXiv:14126544. 2014.

26.

Kapanova

Dimov

Sellier

. A genetic approach to automatic neural network architecture optimization. Neural Computing and Applications. 2018; 29(5): 1481-92.

27.

Ul Islam

Baharudin

Raza

Nallagownden

, eds. Optimization of neural network architecture using genetic algorithm for load forecasting. International Conference on Intelligent and Advanced Systems (ICIAS); 2014: IEEE.

28.

Ritchie

White

Parker

Hahn

Moore

. Optimization of neural network architecture using genetic programming improves detection and modeling of gene-gene interactions in studies of human diseases. BMC Bioinformatics. 2003; 4(1): 1-14.

29.

Boithias

El Mankibi

Michel

, eds. Genetic algorithms based optimization of ANN architecture for buildings’ indoor discomfort and energy consumption prediction. Building Simulation; 2012: Springer.

30.

Luo

Tian

Qin

Chen

Liu

T-Y

. Neural architecture optimization. arXiv preprint arXiv:180807233. 2018.

31.

Ding

. An optimizing BP neural network algorithm based on genetic algorithm. Artificial Intelligence Review. 2011; 36(2): 153-62.

32.

Leung

FH-F

Lam

H-K

Ling

S-H

Tam

PK-S

. Tuning of the structure and parameters of a neural network using an improved genetic algorithm. IEEE Transactions on Neural Networks. 2003; 14(1): 79-88.

33.

Heckerling

Gerber

Tape

Wigton

. Use of genetic algorithms for neural networks to predict community-acquired pneumonia. Artificial Intelligence in Medicine. 2004; 30(1): 71-84.

34.

Guo

Wang

Han

, eds. The enhanced genetic algorithms for the optimization design. 2010 3rd International Conference on Biomedical Engineering and Informatics; 2010: IEEE.

35.

Catic

Gurbeta

Kurtovic-Kozaric

Mehmedbasic

Badnjevic

. Application of Neural Networks for classification of Patau, Edwards, Down, Turner and Klinefelter Syndrome based on first trimester maternal serum screening data, ultrasonographic findings and patient demographics. BMC Medical Genomics. 2018; 11(1): 19.

36.

Kayaer

Yildirim

, eds. Medical diagnosis on Pima Indian diabetes using general regression neural networks. Proceedings of the International Conference on ANNs and Neural Information Processing (ICANN/ICONIP); 2003.

37.

Selker

Griffith

Patil

Long

d’Agostino

. A comparison of performance of mathematical predictive methods for medical diagnosis: identifying acute cardiac ischemia among emergency department patients. Journal of Investigative Medicine: The Official Publication of the American Federation for Clinical Research. 1995; 43(5): 468-76.

38.

Ayer

Chhatwal

Alagoz

Kahn

Jr Woods

Burnside

. Comparison of logistic regression and ANN models in breast cancer risk estimation. Radiographics. 2010; 30(1): 13-22.

39.

Mesgari

Tavusi

Mahmudi

Seyed Mahdi

. Evaluation of some transmission functions in ANNs for short-term prediction of minimum temperature (Case study: Sanandaj synoptic station). Journal of Agricultural Meteorology. 2020; 8(1): 40-50.

40.

Jamshidnezhad

Hosseini

Mohammadi-Asl

Mandana

. An intelligent prenatal screening system for the prediction of Trisomy-21. Informatics in Medicine Unlocked. 2021; 24(100625). doi: 10.1016/j.imu.2021.100625.

41.

Wang

S-C

. ANN. Interdisciplinary computing in java programming: Springer; 2003. 81-100.

42.

Grover

Toghi

, eds. MNIST Dataset Classification Utilizing k-NN Classifier with Modified Sliding-window Metric. Science and Information Conference; 2019: Springer.

43.

Troisi

Sarno

Martinelli

Di Carlo

Landolfi

Scala

, et al. A metabolomics-based approach for non-invasive diagnosis of chromosomal anomalies. Metabolomics. 2017; 13(11): 140.

44.

Al-Maqaleh

Abdullah

AMG

. Intelligent predictive system using classification techniques for heart disease diagnosis. International Journal of Computer Science Engineering (IJCSE). 2017; 6(6): 145-51.

45.

Hannah

Raamesh

. Early Prenatal Diagnosis of Down’s Syndrome-A Machine Learning Approach. Soft Computing for Problem Solving: Springer; 2020; 467-77.

46.

Langarizadeh

Ghazi Saeedi

Karam Niay Far

Hoseinpour

. Predicting premature birth in pregnant women via assisted reproductive technologies using neural network. Journal of Health Administration. 2016; 18(62): 42-51.

47.

Soleimani

Teymouri

Hemati

Biglarian

. ANN for risk assessment of neonatal movement disorders. Razi Journal of Medical Sciences. 2014; 20(115): 31-8.

48.

Neocleous

Nicolaides

Syngelaki

Neokleous

Loizou

Neocleous

, et al., eds. ANNs to investigate the importance and the sensitivity to various parameters used for the prediction of chromosomal abnormalities. IFIP International Conference on Artificial Intelligence Applications and Innovations; 2012: Springer.

49.

Zeinal

Hashemi

. Comparison of neural network learning functions in runoff modeling Iranian journal of Ecohydrology. 2016; 3(4 #a00379).

50.

Fadzil

Nor Ashidi Mat

Zakaria

Muhammad Khusairi

. Intelligent medical disease diagnosis using improved hybrid genetic algorithm-multilayer perceptron network. Journal of Medical Systems. 2013; 37(2): 9934.

51.

Sharma

Mukharjee

. Brain tumor segmentation using genetic algorithm and ANN fuzzy inference system (ANFIS). Advances in Computing and Information Technology: Springer; 2013; 329-39.

52.

Mohammed

Ghani

MKA

, Hamed

Abdullah

Burhanuddin

. A real time computer aided object detection of nasopharyngeal carcinoma using genetic algorithm and ANN based on Haar feature fear. Future Generation Computer Systems. 2018; 89: 539-47.

53.

Mantzaris

Anastassopoulos

Adamopoulos

. Genetic algorithm pruning of probabilistic neural networks in medical disease estimation. Neural Networks. 2011; 24(8): 831-5.

54.

Heckerling

Gerber

Tape

Wigton

. Use of genetic algorithms for neural networks to predict community-acquired pneumonia. Artificial Intelligence in Medicine. 2004; 30(1): 71-84.

55.

Karegowda

Manjunath

Jayaram

. Application of genetic algorithm optimized neural network connection weights for medical diagnosis of pima Indians diabetes. International Journal on Soft Computing. 2011; 2(2): 15-23.

56.

Effati-Daryani

Zarei

Mohammadi

Hemmati

Ghasemi

Mirghafourvand

. Depression, stress, anxiety and their predictors in Iranian pregnant women during the outbreak of COVID-19. BMC Psychol. 2020; 8(1): 99-109. doi: 10.1186/s40359-020-00464-8.

A machine learning technology to improve the risk of non-invasive prenatal tests

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSION:

Keywords

1. Background

2. Method

2.1 Dataset

2.2 The selection of effective diagnostic factors

2.3 Intelligent neuro-genetic integrated model development

2.4 ANN

Table 1 ANN training and transferring functions used in the GA optimization process

Table 3 Performance of 10 times of execution of the GNN model

Table 5 Number of iterations of neural network input factors by GA

Table 6 Selective parameters of neural network structure by GA

Table 7 Results of trimester screening test by common methods in different levels of genetic disorder risk

4.2 Genetic algorithm

4.3 Screening during the COVID-19 pandemic

5. Conclusion

Footnotes

Acknowledgments

Conflict of interest

References

Table 1
ANN training and transferring functions used in the GA optimization process

Table 3
Performance of 10 times of execution of the GNN model

Table 5
Number of iterations of neural network input factors by GA

Table 6
Selective parameters of neural network structure by GA

Table 7
Results of trimester screening test by common methods in different levels of genetic disorder risk