A Model of the First Trimester Evaluation of Foetal Movements and Their Outcomes via Explainable Artificial Intelligence: A Multicentric Study

ABSTRACT

Foetal outcomes with reduced foetal movements in the later pregnancy are widely reported. We intend to quantify early foetal movements (FMs) through a checklist and their foetal outcomes via explainable artificial intelligence. It is a prospective observational study of 356 foetuses in the first trimester, and we were able to screen only 230 foetuses for early foetal growth restriction (FGR). Of which 26 were FGR and 204 were normal and were identified from the dataset using non‐probability convenience sampling techniques. JASP 0.18.3, Jamovi 2.3.21, and Google Collaboratory were used to construct the predictive model. Ultrasound scores of more than 8 had favourable indicators of a normal foetus. CatBoost had the highest accuracy and recall of 87; the highest precision of 79 was given by random forest (RF), decision tree (DT), K‐nearest neighbour (KNN), and CatBoost; and the F1 score of 83 was given by CatBoost. The lowest Hamming loss of 0.13 was obtained via CatBoost. The highest Jaccard score of 0.87 was by CatBoost. The stacked model has an accuracy of 89, a precision of 79, and a recall of 83. Shapley additive explanations (SHAP), local interpretable model‐agnostic explanations (LIME), QLattice, and Anchor also provided good explanations. The created model can serve as a warning tool to obstetricians to make timely medical decisions.

356 pregnant women were recruited, and normative data on first‐trimester foetal movements (FMs) were obtained. Furthermore, we collected data from 230 postnatal patients. We also applied various explainable artificial intelligence algorithms to construct a reliable and optimal clinical predictive model for early foetal growth restriction. The created model can serve as a warning tool to obstetricians to make timely medical decisions.

1. Introduction

Foetal movements (FMs) are critical indicators of foetal health and development. Reduced foetal movements (RFMs) are recognised as nonspecific markers of foetal well‐being. A systematic review by Carroll et al. revealed that stillbirth, small for gestational age, assisted vaginal birth and caesarean section, and induction of labour, are increased with RFM. While RFM occurs in up to 40% of pregnant women, the incidence of stillbirth remains relatively low, affecting approximately 0.7% of live births [1, 2]. Pregnant women are counselled to monitor their FM and report if there are any reduced FMs [3]. Typically, FMs become perceptible between the sixteenth and twentieth weeks of pregnancy, with multigravid women and primigravid women generally between the sixteenth and eighteenth weeks of gestation. However, FM starts at the eighth–tenth weeks of gestation [4, 5], reflecting intricate neural connections' development. By the ninth–twelfth weeks postmenstrual age, the foetus demonstrates a rich repertoire of specific movement patterns called general movements (GM), startles, isolated limb movements, twitches, stretches, yawning, and breathing movements. These endogenously generated movement patterns observed in utero remain present after birth. A significant portion of early movement patterns persist and remain identifiable throughout prenatal and postnatal development. These findings suggest that the neural connections underlying these behaviours are likely formed in early pregnancy [6].

GMs emerge as early as the ninth–tenth weeks postmenstrual age, and it is improbable that structures beyond the brainstem are involved in their production [7, 8]. Central pattern generators in the brainstem, namely the dorsal medullary reticular formation and the nucleus tractus solitarius, produce movements, which are then influenced by sensory information from the reticular formation [9]. The involvement of GM spans all segments from the cervical to the lumbar spinal cord, suggesting that the neuronal structures responsible for generating these movements are likely located supraspinally. Therefore, foetal movements can be considered markers of the developing nervous system. Figure 1 shows the different foetal movements and different levels of neurological development in the foetus.

FIGURE 1.

Schematic representation of different foetal movements and the development of various neurological structures. (Prepared using Biorender‐institutional access)

The foetus conserves energy in response to prolonged hypoxia. The adaptive mechanism known as RFM reduces oxygen use. Other factors include maternal anxiety, alcohol, sedatives, corticosteroids, sleep patterns, intrauterine growth restriction, hypothyroidism, foetal anaemia, neurological or muscular abnormalities, and variations in amniotic fluid levels. Some studies infer excessive foetal activity may indicate malformations, such as anencephaly [10, 11]. Studies suggest twins may be more active than singletons [12, 13]. A retrospective study by Mor L et al. revealed that foetal growth restriction (FGR) with RFM had higher rates of adverse neonatal outcomes than those without RFM [14].

Modern engineering has the potential to progress perpetually, and modern medicine continues to advance and evolve through the intersection of engineering and medicine [15]. Artificial intelligence (AI) developments could revolutionise many aspects of healthcare, opening the door to more personalised, accurate, predictive, and accessible models [16]. Continuous developments in machine learning (ML) have led to the invention of valuable applications in clinical situations, such as detecting compromised foetal states and pregnancy‐related consequences in decision‐support systems [17, 18]. Algorithms for ML can be either supervised or unsupervised, while some studies also categorise other algorithms as reinforcements, as they use data to learn patterns and respond to their surroundings [19]. When applying ML algorithms to clinical data, it is important to compare their performance to the opinions and insights of medical experts [20]. It is impossible to develop a model without errors since medical datasets are naturally imperfect, frequently because of noise or recording errors. Predictions will always contain false positives and false negatives because these errors are typically random. Another significant issue is AI bias, which leads to systematic errors that deviate from what we expect the AI to predict. Explainability makes settling disputes between human specialists and AI systems possible regardless of who renders an error in judgement [21, 22].

Explainability also serves as a communication tool that facilitates consistent and intelligible information sharing between individuals and those making decisions. Models of explainable artificial intelligence (XAI) are intended to improve their efficacy and make them understandable to nonexperts in various contexts. XAI has been growing in popularity to increase openness and address the AI ‘black box’ puzzle [23]. Table 1 gives a concise overview of works performed with foetal movements.

TABLE 1.

gives a concise overview of works performed with foetal movements.

Reference	Objective	Algorithms used	XAI method
Kurjak A et al. [24]	Neurobehavioral movements in ultrasound images with KANET scoring in the third trimester	Convolutional neural networks (CNN)	Nil
Miyagi Y et al. [25]	Analysing foetal facial expression variations between weeks 27 and 37 of pregnancy objectively and quantitatively that seem to be connected to foetal brain activity	Mathematical computational processing; discrete fourier transform (DFT) and chaotic analysis	Nil
D Gopakumar [26]	To identify different patterns of cardiotocography (CTG) throughout pregnancy and IBM Watson classification algorithms to predict foetal health.	XGB Classifier, XGB Classifier enhancement, Random Forest Classifier, Random Forest Classifier enhancement	Nil
Liang S et al. [27]	Foetal movement signals recorded by accelerometer were used.	Kalman filtering, time domain and frequency domain wavelet domain feature extraction, Bayesian optimisation for the LightGBM model	Nil
Innab N et al. [28]	CTG results for expectant mothers with gestational ages that range from 29 to 42 weeks	Multiple ML	SHAP
Miyagi Y et al. [29]	4D ultrasound to detect foetal facial expressions between 27 and 37 weeks of pregnancy, which are believed to correspond to foetal brain activity.	Facial recognition	Nil
Vasung L et al. [30]	24 and 40 weeks of pregnancy, blood oxygenation level‐dependent (BOLD) images MRI method	3D convolutional neural network, 3D U‐Net‐based network	Nil

The studies in Table 1 suggest that foetal movements can be effective markers for evaluating the brain activity of the foetus. Therefore, we hypothesise that observing the pattern of foetal movements in early pregnancy is important. We quantified different types of foetal movements in first‐trimester anomaly scans (FTAS) and developed a scoring tool that can be used even in a low‐resource setting. The tool may act as a warning tool to predict foetal outcomes. To further enhance the tool, we have integrated XAI algorithms to train the model with different FMs and the developed scoring tool in FTAS to predict early foetal growth restriction.

The other contributions are as follows:

• The data were statistically analysed to understand different types of foetal movement variations and patterns.

• Twenty‐seven features from the first‐trimester anomaly scan were evaluated on a scale of 10. Therefore, 28 attributes were considered.

• ML models, including random forest (RF), logistic regression (LR), decision tree (DT), K‐nearest neighbour (KNN), and Catboost. The algorithms have also been combined using a new stacking architecture.

• XAI techniques such as Shapley additive explanations (SHAP), local interpretable model‐agnostic explanations (LIME), QLattice, and Anchor are used to interpret predictions based on FMs.

• Further discussion of crucial foetal movements as markers from an obstetrical perspective is needed.

2. Materials and Methods

2.1. Dataset Description

This is a prospective observational study. The scans were performed via a Voluson S8 and Voluson E8. Informed and signed consent was obtained from all participants during recruitment. The International Society of Ultrasound in Obstetrics & Gynaecology (ISUOG) guidelines were followed during the ultrasound examination [31]. As it involves vulnerable data, permission was obtained from the district health officer. Data collection began only after approval from Kasturba Medical College and Kasturba Hospital Institutional Ethics Committee (IEC1: 122/2022) and after the clinical trial registry (CTRI/2022/08/044770). The sample size was calculated by the proportional concept. The formula we used was n = 4PQ/D² from the review of literature. We have identified proportional value as 0.667 and the margin of error (d) as 5% with a 95% confidence interval. It was observed that the required sample size is 356 for the research study with gestational ages between 11 and 13 weeks and 6 days. 356 foetuses were evaluated for foetal movements in the first trimester, and the results are tabulated in Table 4. Furthermore, we were able to screen only 230 foetal outcomes, i.e., 64.60% of foetus due to socio‐cultural practices in our country. The outcome is early foetal growth restriction (FGR) prediction with foetal movements. Of which only 26 were FGR and 204 were normal, identified from the dataset using non‐probability convenience sampling techniques. Therefore, only those with complete details were included, and others were excluded. The primary outcome of the study is first‐trimester evaluation of foetal movements during routine ultrasound examination and its outcome.

TABLE 4.

presents the distribution of foetal movement.

Foetal movements	Median [Q1, Q3]
Startles	2 [2, 2]
Head anteflexion	2 [1, 2]
Head retroflexion	2 [1, 2]
Head rotation	1[0, 1]
Jaw opening	0 [0,2]
Swallowing	0 [0,0]
Sucking	0 [0,0]
Yawning	0 [0,0]
Munching	0 [0,0]
Hand to head	0 [0,0]
Hand to ear	0 [0,0]
Hand to face	2 [0,2]
Hand to eye	0 [0,0]
Hand to mouth	0 [0, 0]
Clenching	2 [1, 2]
Unclenching	2 [1, 2]
Simultaneous flexion and extension of upper limbs (UL)	2 [1, 2]
UL‐Right	0 [0, 0]
UL‐Left	0 [0, 0]
UL‐Hand to hand	1 [0, 2]
Simultaneous flexion and extension of lower limbs (LL)	2 [0, 2]
LL‐Right	0 [0, 0]
LL‐Left	0 [0, 0]
LL‐Pedalling	2 [1, 2]
Number of trunkal rotations	2 [2, 2]
Movement of the head, trunk, and limb movements in a complex pattern	2 [2, 2]
Breathing	0 [0, 0]

Ultrasound examinations were performed in accordance with the International Society of Ultrasound in Obstetrics and Gynaecology guidelines [31]. The duration of each scan was between 15 and 30 min, depending on the foetal position. Women with multiple pregnancies, morbid obesity, or those unwilling to participate were excluded. The scans were performed by the clinical professionals, who are experienced obstetricians with expertise in ultrasound. We have employed a 2D setting in the study because it is feasible and is used in the regular setting. Moreover, 3D 4D requires high‐end equipment [32, 33, 34, 35, 36]. The images generated in these settings are the reconstructed images, which may miss minute detailing during the analysis, as it is affected by the frame rate [37]. Figure 2 represents different types of foetal movements in a 12‐week, 2‐day‐old foetus. Table 2 presents important primary outcome variables of different types of foetal movements, and the explanations of different movements were given. Different movements were quantified and expressed as the median interquartile Range (IQR) and scored on a scale of 10 on the basis of the presence or absence of movements. The presence of the above movements was marked as 1, the absence was 0, and the total score was 10. Table 3 presents the different types of foetal movements in different planes.

FIGURE 2.

FM in a 12‐week 2 day‐old foetus. The 2D ultrasound image represents the different types of movements involving different parts of the foetus, as described in Table 2.

TABLE 2.

presents the explanation of different types of foetal movements.

Features	Explanation of dataset	Description of foetal movements
1. Startles	Number of startles observed	These are smooth, intricate movements that engage the entire body, lasting from a few seconds or manifesting as ongoing general movements.
2. Head anteflexion	Number of head anteflexion and extension	The to‐and‐fro motions of head and neck movements can be observed in the mid‐sagittal plane and in a neutral position, particularly during head parameter measurements.
3. Head retroflexion	Number of head retroflexion
4. Head rotation	Number of head rotations
5. Jaw opening	Number of jaw‐opening	The lower jaw exhibits movements such as sucking, yawning, munching, and swallowing, which are more distinctly observed in profile views of the foetus and when facial markers such as the nasal bone, orbits, and lips are scanned.
6. Swallowing	Number of swallowing movements
7. Sucking	Number of sucking movements
8. Yawning	Number of yawnings
9. Munching	Number of munching
10. Hand‐to‐head	Number of hand‐to‐head	Notable movements include those directed toward the midline of the face (such as the nose or mouth) and toward the lateral aspects (such as the ears).
11. Hand‐to‐ear	Number of hand‐to‐ear
12. Hand‐to‐face	Number of hand‐to‐face
13. Hand‐to‐eye	Number of hand‐to‐eye
14. Hand‐to‐mouth	Number of hand‐to‐mouth
15. Clenching	Number of clenching	Clenching and unclenching of fingers can be detected during ultrasound, with the digits of the hands identifiable as early as 11 weeks of gestation.
16. Unclenching	Number of unclenching
17. Flexion and extension of both upper limbs together	Number of upper limb movements involving both hands	These movements, such as flexion and extension of the upper limbs, can be observed in both the mid‐sagittal and longitudinal planes.
18. Upper limb right	Number of upper limb movements involving the right hand only
19. Upper limb left	Number of upper limb movements involving the left hand only
20. Hand to hand	Number of hand‐to‐hand movements
21. Flexion and extension of both lower limbs together	Number of lower limb movements involving both limbs with flexion and extension	Flexion and extension of both lower limbs can be noted at various angles throughout the scan, indicating healthy musculoskeletal development.
22. Right lower limbs	Number of movements in right lower limb
23. Left lower limbs	Number of movements in left lower limb
24. Pedalling	Number of pedalling movements involving active lower limb movements against the uterine space
25. Trunkal movements	Number of movements involving the trunk of the foetus	The trunk moves along with the rotation of the entire body, using the entire space of the uterus.
26. Complex rotation	Number of complex rotations in the entire space of the uterus	Movement involves the whole body in a complex, elegant, and patterned manner. Normal gross movements involve the whole body and can last from a few seconds to several minutes or longer. The limbs, trunk, and head exhibit variable yet smooth movements that fluctuate in speed and intensity, beginning and ending gradually (4, 9, 38)
27. Breathing	Number of breathing movements	Contraction and relaxation of the abdomen can be observed as a movement of the chest wall.
28. Ultrasound score	Composite score of the movements	Composite record of score on a scale of 10

TABLE 3.

presents the different types of foetal movements in different planes.

Mid sagittal plane	Coronal plane	Transverse plane
Startles Isolated head movements ·Anteflexion ·Retroflexion ·Rotation Mouthing movement ·Jaw Opening ·Swallowing ·Sucking ·Yawning ·Munching Hand movements ·Hand to head ·Hand to ear ·Hand to face ·Hand to eye ·Hand to mouth ·Hand to hand Finger movements ·Clenching ·Unclenching Upper limb movements ·Flexion and extension (Both) ·Isolated limb movements (Right and left) ·Hand to hand Lower limb movements ·Flexion and extension (both) ·Isolated limb movements (right and left) ·Pedalling Trunkal movement Complex movements involving head, trunk, and limbs Breathing movements	Mouthing movement ·Jaw Opening ·Swallowing ·Sucking ·Yawning ·Munching Hand movements ·Hand to head ·Hand to ear ·Hand to face ·Hand to eye ·Hand to mouth ·Hand to hand ·Hand to hand Trunkal movement	Breathing movements Hand to hand

Different ML algorithms have been used to predict foetal outcomes. Figure 3 represents the integration of different ML techniques with fetal movement attributes to develop an early foetal restriction model.

FIGURE 3.

Integration of different ML techniques with respect to foetal movement parameters.

2.2. Statistical Analysis

We used open access free software such as JASP 0.18.3 and Jamovi 2.3.21 for statistics and Google Collaboratory for ML. We used the Kolmogorov–Smirnov test to test hypotheses and check assumptions. The normally distributed variables are reported as mean ± SD, and the not normally distributed variables are reported as median (IQR). The Mann–Whitney U test was used to compare the normal foetus and foetus with FGR with FMs. DT and LR were performed to find the prediction of FGR with FMs.

The median IQR of the CRL was 61 [55.4, 65.03] mm, and that of the NT was 1.3 [1.1, 1.5] mm. Figure 4 represents the density of distribution of crown‐rump length (CRL) and nuchal translucency in mm. The CRL was considered for determining the age of the foetus. Table 4 presents the distributions of the movements of the 356 foetuses. Startles, trunkal rotation, and movement of the head, trunk, and limb in a complex pattern were 2 [2, 2]. Swallowing, sucking, yawning, munching, hand to head, hand to ear, hand to eye, hand to mouth, upper limb ‐right and left, breathing 0 [0, 0], head anteflexion, head retroflexion, clenching and unclenching of fingers, simultaneous flexion of upper limbs were 2 [1, 2], hand to face, simultaneous flexion and extension of lower limbs 2 [0, 2], head rotation was 1 [0, 1], jaw opening 0 [0, 2], and hand to hand 1 [0, 2]. Kappa analysis was performed for interrater reliability and showed good agreement for different FMs [39]. Figure 5 represents the quantitative aspects of FMs during the FTAS.

FIGURE 4.

Shows the density distributions of crown‐rump length (CRL) and nuchal translucency in mm. The median IQR of the CRL was 61 [55.4, 65.03] mm, and that of the NT was 1.3 [1.1, 1.5] mm.

FIGURE 5.

Depicts the quantitative aspects of foetal movements during the FTAS. HAF, head anteflexion; HRF, HR, head rotation; head retroflexion; LL, lower limb; No. of rotations, number of rotations; UL, upper limb. The figure indicates the number of movements observed during the routine examination, and the values are represented as the median IQR in Table 4.

Table 5 presents the group descriptions of different FMs with early FGR and normal foetuses. The Mann‒Whitney U test table has been added, along with a table description. The test was used because the outcome variables were not normally distributed across the groups.

TABLE 5.

presents group descriptions of different foetal movements.

Types of movements	Group	Mean	SD	p	Effect Size
Startles	FGR	1.85	0.54	0.11	0.15
	Normal	2.02	0.60
Head anteflexion	FGR	1.31	0.74	0.07	0.19
	Normal	1.54	0.80
Head retroflexion	FGR	1.27	0.67	0.03	0.24
	Normal	1.54	0.81
Head Rotation	FGR	0.85	0.83	0.96	0.00
	Normal	0.81	0.74
Jaw opening	FGR	0.85	0.97	0.84	0.02
	Normal	0.82	1.01
Swallowing	FGR	0.08	0.39	0.91	0.00
	Normal	0.06	0.36
Sucking	FGR	0	0	0.38	0.02
	Normal	0.04	0.27
Yawning	FGR	0	0	0.38	0.02
	Normal	0.05	0.29
Munching	FGR	0.04	0.20	0.79	0.01
	Normal	0.10	0.48
Hand to head	FGR	0.15	0.46	0.30	0.09
	Normal	0.30	0.65
Hand to ear	FGR	0.04	0.20	0.08	0.13
	Normal	0.25	0.62
Hand to face	FGR	1.23	0.91	0.91	0.01
	Normal	1.24	0.92
Hand to eye	FGR	0.08	0.39	0.53	0.02
	Normal	0.03	0.22
Hand to mouth	FGR	0.23	0.65	0.52	0.05
	Normal	0.24	0.55
Clenching	FGR	1.12	0.95	0.11	0.17
	Normal	1.44	0.85
Unclenching	FGR	1.12	0.95	0.11	0.17
	Normal	1.44	0.85
Upper limb‐both	FGR	1.38	0.85	0.87	0.02
	Normal	1.40	0.88
Upper limb right	FGR	0.15	0.46	0.48	0.04
	Normal	0.12	0.45
Upper limb left	FGR	0.12	0.43	0.81	0.01
	Normal	0.11	0.45
Hand to hand	FGR	0.62	0.75	0.32	0.11
	Normal	0.82	0.88
Lower limb‐ flexion & extension	FGR	1.19	0.94	0.82	0.03
	Normal	1.23	0.90
Lower limb‐right	FGR	0.08	0.39	0.58	0.03
	Normal	0.11	0.43
Lower limb‐left	FGR	0.08	0.39	0.37	0.05
	Normal	0.15	0.51
Pedalling	FGR	1.50	0.91	0.69	0.04
	Normal	1.56	0.95
Trunkal	FGR	1.92	0.80	0.43	0.08
	Normal	2.09	0.74
Complex rotation	FGR	1.69	0.74	0.23	0.12
	Normal	1.88	0.74
Breathing	FGR	0	0	0.48	0.02
	Normal	0.03	0.25
Ultrasound score	FGR	7.50	1.42	0.09	0.20
	Normal	7.86	1.35

Table 5 provides descriptions of the various foetal movements observed in both early foetal growth restriction (FGR) and normal foetuses. Additionally, a table displaying the results of the Mann–Whitney U test has been included, along with an explanation of the table. This test was utilised due to the outcome variables not being normally distributed across the groups.

Figure 6 presents the scores of the foetuses during foetal movement evaluation. The majority of foetuses were scored as 8 or 9. Seven Foetus scored 10; the difference was potentially due to the development of foetal breathing movements.

FIGURE 6.

Represents the foetal movement scoring pattern. The foetal score is shown on the X‐axis, while the number of foetuses is shown on the Y‐axis.

Figure 7 represents the different FMs correlated with early FGR. Pearson's/Spearman's correlation is used to find the relationship between FMs and the outcome variable. Sucking, yawning, munching, hand‐to‐face, hand‐to‐mouth movements, both upper limbs, lower limb flexion and extension, left lower limb, right lower limb, pedalling, and breathing movements were correlated with early FGR.

FIGURE 7.

Pearson's correlation is used to find the relationship between foetal movements and the outcome variable.

2.3. Data Preprocessing

In ML, the dataset must be pre‐processed. Categorical attributes were encoded, continuous values were scaled, and data balancing was completed. The categorical values are coded into numbers since classifiers cannot accept values in text. Data scaling was performed via the standardisation technique [40, 41]. When the data points are significantly different, the accuracy decreases. Moreover, the classifiers prioritise parameters with larger values, independent of the units taken into consideration. Normalisation and standardisation are the two techniques used in ML to scale datasets. Standardisation was selected for this study because it performs better with outliers. Next, the dataset was split 80:20 into training and testing datasets. There was a noticeable imbalance in the dataset. Compared with normal foetuses, there were fewer early FGR.

Because the models prefer the majority classes, the conclusions derived from the imbalanced data are biased. To balance the training dataset, we employed an oversampling method known as the borderline Synthetic Minority Oversampling Technique (SMOTE) [40]. This method generates new synthetic samples using the K‐nearest algorithm. The previously indicated approach is also effective for borderline cases. Because we did not want to miss any significant trends or patterns, undersampling was discouraged in this study. The testing data were not balanced to guarantee data integrity. Feature selection is crucial because it eliminates noise and redundant data. Additionally, it improves the accuracy, precision, and recall metrics; aids in model training; and makes the model easier to understand. Avoiding dimensionality, enhancing generality to lessen overfitting, and reducing the training time are the next steps [41, 42].

A mutual information curve was utilised to identify the features that contributed most significantly to the model, as illustrated in Figure 8. To determine which features contribute the least to the models, a mutual information curve was constructed. The highest importance was observed at the beginning of the curve [43].

FIGURE 8.

A mutual information curve representing different features that contributed most significantly to the model. The highest importance was observed at the beginning of the curve.

2.4. Machine Learning Optimisation

ML uses algorithms that methodically analyse and understand complex relationships in data. By using previous data as input, ML enables algorithms to predict outcomes precisely. In this work, many ML classifiers, such as the ‘RF, LR, DT, KNN, and CatBoost classifiers,’ are used.

A stacked model was used, and a combination of five models was ensembled to improve the accuracy of the individual classifiers. Furthermore, the stacking strategy employed helps to aggregate predictions from various baseline models, resulting in better overall predictive performance [42]. The selection of hyperparameters impacts a classifier's performance. This tuning procedure is automated by grid search, which produces the optimal parameter values as outputs. To assess the models in this study, we chose a range of classification and loss criteria. The F1 score, average precision (AP), precision (correctness of positive prediction), recall (sensitivity), accuracy (overall correctness), Matthew's correlation coefficient (MCC), Jaccard score, and Hamming loss are some of these measurements. Cross‐validation was performed in the initial stage. A data resampling technique called cross‐validation is used to evaluate how well prediction models can be generalised and to avoid overfitting [44]. It involves splitting the data into several folds training on the remaining folds while using one as a validation set. For a more accurate performance estimate, this procedure is carried out several times, using a different fold as the validation set each time. Cross‐validation ensures that the chosen model performs effectively when applied to fresh data by preventing overfitting [45]. Figure 9 represents the ML methodology used in the research.

FIGURE 9.

Shows the different ML algorithms for generating the stacked model.

We particularly emphasise precision and recall, as they prioritise minimising false positives and false negatives. Tools such as SHAP, LIME, the Qlattice, and anchors were used. An elaborate ML method called SHAP explains how each feature affects the result by giving each feature a relevance value for a particular prediction. This tool offers significant versatility because it is model‐agnostic and may be used with any machine‐learning model. SHAP offers local and global explanations, offering comprehensive insights into individual forecasts and the behaviour of the entire model. By precisely expressing variations in the model's predictions through the SHAP values, consistency and dependability are guaranteed. Additionally, it efficiently sets the SHAP values of missing features to zero. SHAP values help to determine the most important features in a model and understand how each feature influences the prediction outcome. LIME works especially well with smaller datasets and offers insights into individual predictions that are easier to understand. It is versatile across diverse forms of data because of its model‐agnostic nature, which enables it to be used in different models. Anchors explain significant features through a set of ‘rules’ and ‘conditions.’ Two measures are used to measure each anchor (condition): precision and coverage.

The accuracy of the explanations depends on their precision. Coverage defines the number of instances that are predicted via the same conditions, which aids in understanding the model's prediction. This adaptability improves trust in AI‐driven diagnostic procedures by helping patients and clinicians acquire confidence in the diagnostic outcomes generated by AI [46]. The study's results were explained through four XAI approaches (SHAP, LIME, Qlattice, and Anchor) after training and testing the ML models. These XAI algorithms offer insights in the form of graphs and tables, making the results simple for ML users to understand. Figure 10 illustrates the flow of the research from ethical clearance to the development of the prediction model.

FIGURE 10.

Illustrates the flow of the research in the prediction model's development. (Prepared using Biorender‐Institutional access)

3. Results

The performances of the classifiers utilised are discussed in this section. Table 6 summarises the results obtained by the various models. CatBoost yielded the highest accuracy of 87 and a precision of 79 in most of the models except LR, which is 77. CatBoost had a recall of 87 and an F1 score of 83. The least Hamming loss of 0.13 was also achieved by CatBoost. CatBoost also obtained the highest Jaccard score. The MCC obtained was negative. The stacked model has an accuracy of 89, a precision of 79, and a recall of 83.

TABLE 6.

Summarises the results obtained by the various models.

Classifiers	Accuracy	Precision	Recall	F1 score	Hamming loss	Jaccord score	Log loss	Mathews correlation coefficient
1. RF	85	79	85	82	0.15	0.85	5.48	−0.07
2. LR	70	77	70	73	0.30	0.69	10.97	−0.17
3. DT	83	79	83	81	0.17	0.83	6.27	−0.09
4. KNN	65	79	65	71	0.35	0.64	12.54	−0.06
5. CatBoost	87	79	87	83	0.13	0.87	4.70	−0.05
6. Stacked model	89	79	83	81	0.17	0.83	6.27	−0.09

Table 6 CatBoost yielded the highest accuracy of 87 and a precision of 79 in most of the models except LR, which is 77. CatBoost had a recall of 87 and an F1 score of 83. The least Hamming loss of 0.13 was also achieved by CatBoost. CatBoost also obtained the highest Jaccard score. The Mathews correlation coefficient (MCC) obtained was negative. The stacked model has an accuracy of 89, a precision of 79 and a recall of 83.

The table in Figure 11 shows how well a classification algorithm performs, which is called a confusion matrix. An overview of a classifier's classification performance with respect to a set of test data is given by a confusion matrix. The outputs are divided into true positives, true negatives, false positives, and false negatives. It can be used to calculate multiple measures of performance for models [47]. Figure 11 shows the confusion matrix obtained for the stacked model used to run the dataset.

FIGURE 11.

Confusion matrix from the Stack model. The stacked model works well in identifying the true negative.

SHAP is an XAI tool that specifies the weights assigned to each characteristic and determines whether a low or high value best characterises it. SHAP analysis is a feature‐based interpretability method. A fair and impartial indicator of each feature's contribution to each sample's anticipated value is provided by the SHAP values [48]. Figure 12 represents the important features.

FIGURE 12.

SHAP: The importance of different attributes in the prediction model is represented in descending order. Ultrasound score, head retroflexion, and anteflexion are placed as the most important attributes, followed by others. AF, anteflexion; LL‐F & E, lower limb flexion and extension; RF, retroflexion; UL, Upper limb.

A global summary of SHAP values for specific features is shown in the beeswarm plot, where each feature is represented by a row that is sorted according to its mean absolute SHAP value. The SHAP values for that specific feature in each dataset sample are represented by the individual dots in each row [48]. The FGR (left) and normal (right) classes are divided by a hyperplane. A larger value is denoted by red, and a lower value is denoted by blue. Additionally, the markers are sorted according to their significance, with the best characteristic remaining at the top. Hand‐to‐hand movements, ultrasound scores, and head movements are considered important features in FGR prediction, as shown in Figure 13.

FIGURE 13.

Global SHAP interpretation via beeswarm plots. The FGR (left) and normal (right) classes are divided by a hyperplane. A larger value is denoted by red, and a lower value is denoted by blue Hand‐to‐hand movements, ultrasound scores, and head movements are considered important features in FGR prediction. AF, anteflexion; LL‐F & E, lower limb flexion and extension; RF, retroflexion; UL, Upper limb.

By determining which input (i.e., feature) most influences the prediction outcome, local interpretable model‐agnostic explanations (LIME) can construct a fairly simple local surrogate prediction model around the query point that has to be interpreted [49]. Figure 14 shows the LIME XAI technique.

FIGURE 14.

The LIME XAI technique. Ultrasound score more than 8 has good prediction for normal fetus. AF, anteflexion; LL‐lft, lower limb left; LL‐rt, Lower limb right.

The Python‐based architecture is provided by a software program called Qlattice, which iterates through all potential models and selects the one with the best qualities to provide the best fit for the data. For this technology to work, supervised learning is used. With the feyn library, a Qlattice model can be made. An XAI approach produces the Qlattice equation, which aids in the mathematical description of the output in relation to a few input attributes. For binary classification, a logistic regression equation in which various qualities are assigned varying weights as inputs is used. The key feature is illustrated by a sample QLattice explanation in Figure 15.

FIGURE 15.

Key features selected by the QLattice. Head RF, head retroflexion.

Anchor is an XAI method that makes use of conditions and rules. Precision and coverage are used to gauge an anchor's strength. The accuracy of the anchor is defined by precision. The number of instances that use the same circumstances is determined by coverage. Table 7 describes the anchors for various FMs and ultrasound scores. While pedalling and ultrasound scores act as favourable indicators of a normal foetus, head rotation suggests early foetal growth restriction. In contrast, hand‐to‐hand movements have a role in both normal and FGR foetuses. Table 8 presents important features of different methods, such as Pearson's correlation, mutual information, and different XAI techniques.

TABLE 7.

Represents the model explainability of Anchor.

Instance	Patient prediction	Anchor condition	Precision	Coverage
1.	Normal	Hand to hand ≤ 0.00 AND Pedalling > 2.00	77	10
2.	Normal	Hand to hand ≤ 0.00 AND Ultrasound score > 8.00	77	9
3.	FGR	Hand to hand ≤ 0.00 AND Head Rotation ≤ 0.00	85	10

Table 7 describes the anchors for various foetal movements and ultrasound scores. Pedalling more than 2.00 and ultrasound scores more than 8.00 act as favourable indicators of a normal foetus. Head rotation ≤ 0.00 suggests early foetal growth restriction. In contrast, hand‐to‐hand movements have a role in both normal and FGR foetuses ≤ 0.00.

TABLE 8.

Important features for predicting FGR with different statistical and ML algorithms.

Pearson's correlation matrix	Mutual information	SHAP	LIME	Anchor	QLattice
Pedalling Lower limb flexion and extension Upper limb flexion and extension Hand to mouth Hand to face Breathing	Head rotation Hand to hand Pedalling Trunkal Hand to face	Hand to hand Ultrasound score Head retroflexion Head anteflexion	Ultrasound score Hand to ear Hand to hand	Hand to hand Pedalling Ultrasound score Head rotation	Head retroflexion

4. Discussion

The performance of each ML model developed for this study is assessed and contrasted in this section. The best‐performing models were identified and given access to XAI tools. Among the five individual models used and the sixth stacked model developed using the other five models, CatBoost yielded the highest accuracy of 87 and a precision of 79 in most of the models except LR, which is 77. CatBoost had a recall of 87 and an F1 score of 83. The least Hamming loss of 0.13 was also achieved by CatBoost. The highest Jaccard score was also obtained by CatBoost. The MCC obtained was negative. The stacked model has an accuracy of 89, a precision of 79, and a recall of 83.

Furthermore, various XAI techniques have been introduced and used to better understand the importance of each feature and make it more interpretable. The various XAI techniques used, including the SHAP, LIME, QLattice, and ANCHOR methods, were used to show the relative importance of each attribute and describe it graphically.

According to SHAP and its generated plots, as shown in Figures 11 and 12, hand‐to‐hand movements, the ultrasound score, and head retroflexion and anteflexion had the highest weight towards best predicting the output.

The LIME agreed with the SHAP and had the greatest importance for the ultrasound score and hand movements in the prediction of FGR. The QLattice assumes head retroflexion to generate a logistic regression equation to predict the output classification.

Owing to the similarity in all five models, the following can be assumed:

• Ultrasound score has the highest weight in describing the model.

• Hand movements and head rotation are the next most important features.

Our observations show different types of FM evaluation in early pregnancy. A score of 10 was achieved by only seven foetuses, while the majority scored either 8 or 9. This notable difference can be attributed to the development of breathing movements and the observation of various types of mouth movements. We observed early hand movements that facilitate exploration and interaction with the uterine wall. These purposeful movements indicate a connection between sensorimotor development and environmental factors, known as ‘affordances.’ Like the ‘contact paths’ observed in the facial grooming of mice, these movements demonstrate how functional activities are influenced by environmental conditions [50]. Finger clenching and unclenching were also observed during ultrasound. Our findings are consistent with those reported by Pooh. R and Ogura,T. noted a significant change in finger position at 11 weeks [51]. Clinically, a clenched hand with overlapping fingers is commonly associated with aneuploidy syndrome but has been underreported in patients with MED12 deficiency [52].

The ultrasound score had a significant weight in prediction; it is a composite value of 27 different types of FMs. A total of 230 pregnant women were delivered, and the foetal outcomes were collected and correlated with different FMs and ultrasound scores. Babies that develop early fetal growth restriction have ultrasound scores less than or equal to 8. Most movements, with the exception of breathing movements, develop by the gestational age of 11–13 weeks and 6 days.

We did not observe significant breathing during the scan, and the median IQR was 0 [0, 0]. By 10 weeks of pregnancy, the diaphragm formation is completed, providing the anatomical basis for the onset of breathing and hiccups. While the former may occasionally be seen as early as 8 or 9 weeks, by 12 weeks, the phrenic motor neurones appear to have enough synapse connections [53]. Studies suggest that the incidence of foetal breathing develops at approximately 10 weeks of pregnancy and increases with gestational age. Hiccups have been proposed as a form of programmed isometric exercise for the inspiratory muscles, which helps refine the diaphragmatic motion required for effective breathing and lung maturation, diminishes by 10 weeks of gestation, and is only compromised in near‐term foetuses continue to exhibit frequent hiccups after all other movements have ceased [54, 55]. On the basis of the above observations, we suggest that monitoring the movements outlined above can serve as an effective method for the indirect assessment of foetal central nervous system (CNS) function.

Studies related to early foetal head movements are limited. Therefore, we followed three different patterns of foetal head movements: head anteflexion and retroflexion, with a median IQR of 2 [1, 2]o and head rotation of 1 [0, 1].

Our data support the findings of the study by Lüchinger et al., who suggested that by 12 weeks, GMs become more inconsistent in both speed and amplitude, lasting between 1 and 4 min, with fluctuations throughout. This study revealed that the sudden onset of GMs between 9 and 13 weeks is a normal part of development. The presence of abrupt‐onset GMs in early foetal life may suggest that the influence of the supraspinal network and a lack of variation and complexity in movement can be strong indicators of dysfunction in supraspinal components [56]. On the basis of the above observations, we suggest that monitoring the movements outlined above can serve as an effective method for the indirect assessment of foetal CNS function and early foetal growth restriction.

We acknowledge certain limitations in our study. Owing to the strict ethical and regulatory guidelines in our country, conducting research involving pregnant women and foetal data is particularly challenging. However, we obtained the necessary approvals to create a tool that can be used without requiring additional infrastructure. While this is a multicentric study, it was conducted at only two centres, and future research with a larger, more diverse sample could help propel this field forward. Although 356 study participants were initially recruited, only 230 were delivered at our institutes because of prevalent sociocultural practices in our country. Consequently, data could only be gathered from the records of these 230 participants. The study's lack of significant findings may be caused by several types of variables. Minimal effect sizes imply insignificant variations, which may require more substantial samples for analysis. It is difficult to identify significant differences between 2 groups because of the wide range of sample differences between the two groups. The early FGR babies were only 26, whereas the normal foetuses were 204 in the follow‐up. The movements of the subjects being measured could be fundamentally changeable, limiting specific analysis. Furthermore, the large number of comparisons improves the chance of Type I errors, indicating that further modifications could improve analytical reliability. In a longitudinal study with a larger sample size, various maternal, demographic, and genetic factors can be included in the study and correlated with the ultrasound score.

5. Conclusion

Sustainable Development Goal 3 (SDG) aims to ensure healthy lives and promote the well‐being of all of all ages [57]. Our study indicates that elucidating different types of FM during routine first‐trimester 11–13 week 6‐day ultrasounds is not only feasible but also valuable for the functional assessment of the developing CNS and early foetal growth restriction even in a limited resource setting. Moreover, applying different ML techniques and XAI gave good indicators about ultrasound scores and early FGR. Several supervised learning algorithms and XAI approaches were used in this work to predict early foetal growth restriction with foetal movements and developed ultrasound scores. By identifying pregnant women with low scores, appropriate follow‐up can be planned with an interdisciplinary team. Furthermore, the integration of a cloud database that incorporates sociodemographic, genetic, clinical, antenatal, and postnatal data can support advancements in the field of foetal precision medicine.

Author Contributions

Manohar Pavanya: data curation, formal analysis, investigation, methodology, visualisation and writing – original draft. Krishnaraj Chadaga: data curation, formal analysis, resources, software, visualisation, writing – original draft, writing – review and editing. Vennila J: data curation, formal analysis, software, visualization, writing – review and editing. Akhila Vasudeva: investigation, methodology, resources, writing – review and editing. Bhamini Krishna Rao: validation, visualisation, writing – review and editing. Shashikala K. Bhat: conceptualisation, data curation, investigation, methodology, project administration, resources, supervision, validation, visualisation, writing – original draft, writing – review and editing.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

Due to ethical restrictions, the data underlying the study's conclusions are not publicly available. However, the corresponding author may provide them upon request.