Trimester-aware yoga video recommendation using hybrid deep learning for pregnant woman

Abstract

Practicing yoga during pregnancy offers significant physical and mental health benefits for expectant mothers. When incorporated into daily routines, yoga can help mitigate common pregnancy-related discomforts and promote overall well-being. Importantly, trimester-aware yoga recommendations enable women to practice safely according to their stage of pregnancy, preparing both body and mind for a healthier gestational period and childbirth.This study proposes a novel trimester-aware hybrid deep learning framework that integrates multimodal text–video analysis with physiological safety reasoning. The model introduces two key innovations: (i) a Trimester-Weighted Wasserstein Similarity (TW-WD) mechanism for adaptive alignment of user and video embeddings under trimester-specific safety constraints, and (ii) a Safety-Aware Directed Graph Convolutional Relational Neural Network (GCRNN) that propagates health-condition and trimester relationships across the recommendation graph. Experimental results demonstrate an accuracy of 98.3% under fivefold cross-validation and over 97.5% trimester-specific safety compliance, highlighting the framework’s clinical reliability and effectiveness in delivering personalized prenatal yoga recommendations.

Introduction

Yoga is an ancient mind-body practice that has gained increasing popularity in industrialized countries as a therapeutic intervention for a variety of physical, psychiatric, neurological, and immunological disorders^1–3. The term “yoga” literally means “to unite” and is derived from the Sanskrit word yug. Broadly, it encompasses practices aimed at improving health and fostering a unified sense of self. Originating in India, yoga is a comprehensive discipline that integrates breathing exercises (pranayama), physical postures (asana), concentration and meditation (dhyana and dharana), and contemplative practices^4,5. It is widely recognized for promoting balance across physical, mental, emotional, and spiritual dimensions. Although contemporary yoga traditions and schools vary, most incorporate at least physical postures and breathing techniques⁶.

Beyond regulating physiological systems—including the immune, endocrine, cardiovascular, and nervous systems—yoga is reported to enhance psychological well-being (e.g., positive mood and optimism) and physical fitness (e.g., flexibility, strength, and endurance)^7–9. Women face unique physiological and psychological challenges during pregnancy, which require management of diverse physical, mental, emotional, and pain-related states throughout gestation and labor¹⁰. Maintaining maternal health and quality of life is essential for optimal pregnancy outcomes, and self-soothing strategies, relaxation, and psychoeducation play critical roles during this transitional period¹¹.

Maternal stress and anxiety have been linked to numerous adverse effects on fetal development. Fetal exposure to maternal stress and stress-related peptides is a risk factor for altered nervous system programming and brain morphology in infants and children^7–13. Early gestational stress has been associated with delayed fetal maturation, impaired neonatal stress regulation, altered behavioral responses, cognitive impairments, emotional and behavioral difficulties in toddlers, and reduced brain volume in regions related to cognition in children^14–16. Moreover, prenatal maternal stress may increase the risk of long-term adverse outcomes for offspring, including attention deficit hyperactivity disorder and impaired executive function¹⁷.

Given these significant risks, it is crucial to consider safe and effective interventions to support the physical and mental well-being of expectant mothers. A personalized yoga recommendation system for pregnant women can provide such benefits^18,19. By leveraging YouTube videos, textual metadata, and video frame analysis, instructors can recommend yoga routines tailored to the mother’s current trimester and health needs (e.g., back pain relief, stress reduction, improved sleep). This approach ensures that practices are safe, individualized, and provide guidance to help pregnant women manage stress, enhance quality of life, and promote positive health outcomes for both mother and child.

In this research, we propose a novel yoga video recommendation model specifically designed for pregnant women. The primary contributions of this work within the context of safety-aware prenatal recommendation systems are as follows:

• Trimester-aware multimodal framework: Development of a recommendation system integrating textual metadata, visual pose features, and maternal health context for personalized prenatal yoga guidance.

• Trimester-weighted Wasserstein similarity: Formulation of a multimodal similarity mechanism incorporating pregnancy-stage safety constraints.

• Safety-constrained relational reasoning: Adaptation of a directed graph convolutional relational neural network (GCRNN) to model user–pose–video–health relationships, enabling relational reasoning without introducing unnecessary architectural complexity.

• Unified feature extraction pipeline: Systematic integration of BTF-IDF and ResNet152 into a single recommendation pipeline optimized for limited-data and safety-critical settings.

• Comprehensive evaluation: Assessment using ranking-based metrics, fivefold cross-validation, and expert clinical review to ensure robustness and adherence to safety standards.

The remainder of the paper is organized as follows: Sect. 2 reviews related work, Sect. 3 details the proposed yoga pose recommendation methodology with illustrative figures, Sect. 4 presents results and analysis, and Sect. 5 concludes the study.

Related works

The process of data processing

Maharana et al.²⁰ provided a comprehensive review of data preprocessing techniques in machine learning, highlighting its critical role in improving data quality prior to model training. The authors identified common challenges, including noise, corruption, missing values, and inconsistencies, noting that inadequate preprocessing can significantly reduce model accuracy. To address these issues, they discussed various strategies such as classification, clustering, association analysis, and data augmentation. In particular, image augmentation techniques, including flipping and rotation, were shown to expand limited datasets without altering the original information, thereby enhancing model generalization and robustness. Similarly, Zhao et al.²¹ proposed an energy-water optimization framework that integrates few-shot learning with deep autoregressive models to handle scarce data effectively. Their results demonstrated that intelligent augmentation and model optimization could improve prediction accuracy by up to 33%, underscoring the importance of thorough preprocessing and dataset expansion in achieving reliable machine learning outcomes.

Introduction of deep and graph neural network models

Deep learning has also been applied to capture spatial and temporal dependencies in yoga pose recognition and recommendation²². For instance, Dileep Kumar Kadali et al.²³ proposed a Yoga Pose Corrector using Convolutional Neural Networks (CNNs) to evaluate user poses against ideal postures from a curated dataset. While their model demonstrated strong real-time correction capabilities, it relied on limited feature extraction and would benefit from more robust hybrid feature representations to maximize accuracy. Similarly, Sumeet Saurav²⁴ investigated 3D Convolutional Neural Networks (3DCNNs) and a hybrid CNN–LSTM architecture to achieve efficient pose detection on embedded systems. Although these methods achieved high accuracy, they lacked trimester awareness and were not specifically designed to ensure the safety of pregnant users. Huang et al.²⁵ highlighted the instability of predictive performance in repeated experiments with Graph Neural Networks (GNNs), noting that oscillations in certain nodes were a significant source of variance. To address this, they proposed the Graph Relearn Network (GRN), which operates in two stages—pre-predict and relearn—to progressively improve predictions for unstable nodes.

Traditional vs. Machine learning methods in yoga recommendation

Traditional rule-based and clustering techniques have been extended using machine learning (ML) algorithms to better address personalized health conditions²⁶. Khushi Bawistale and Surendran R²⁷ proposed an ML-based yoga guidance system for third-trimester pregnant women using the Lloyd K-Means algorithm. Their model analyzed users’ medical histories and symptoms (e.g., gestational diabetes, preeclampsia, or fatigue) to recommend safe asanas and care measures. While achieving an accuracy of 86.14%, overfitting was observed, highlighting the need for hybrid, safety-aware learning frameworks. Collectively, these studies demonstrate that although ML-based yoga recommendation systems improve personalization and detection accuracy, existing models often lack contextual safety, multimodal integration, and trimester-specific adaptation—gaps addressed in the present trimester-aware, graph-based multimodal learning framework.

Umrao and Bansal²⁶ highlighted the limitations of traditional, non-individualized online yoga instruction. They developed a Personalized Yoga Assistant using natural language processing (NLP) and multi-label learning with BERT, filtering 140 asanas and ranking recommendations based on health conditions using cosine similarity. Wei and Yi²⁸ proposed an optimized ML framework to analyze yoga effects on Chronic Venous Insufficiency (CVI), employing a hybrid Naïve Bayes Classifier with Jellyfish Search Optimizer and Flying Fox Optimization, achieving 91.5% precision and 90% recall, demonstrating superior predictive performance compared to heuristic methods.

Indhuvadana et al.²⁹ introduced MindMat, an AI-driven personalized yoga system designed to address the lack of feedback in traditional instruction, which often leads to incorrect postures and injuries. MindMat integrates user health profiles, real-time pose estimation, and a pose database with contraindications to provide dynamic feedback and corrective guidance. Using MediaPipe BlazePose for skeletal tracking and reinforcement learning for adaptive recommendations, the system improves pose accuracy, flexibility, and injury prevention. Similarly, Irfan et al.³⁰ focused on maternal mental health, proposing an ML framework for large-scale prediction and monitoring of psychological risks during pregnancy. Their model incorporates ensemble learning and a personalized loss function to handle class imbalance, demonstrating nearly perfect regression performance and supporting intelligent, web-based platforms for early diagnosis and personalized interventions, including yoga recommendations and natural remedies, underscoring the growing role of AI in integrative maternal healthcare.

Research gap: Although numerous studies have explored AI-based recommendation and monitoring systems, existing methods often lack domain-specific adaptation for prenatal care and fail to fully incorporate critical contextual factors, such as safety compliance and trimester-specific risks. Furthermore, most previous systems are evaluated primarily using standard classification metrics, without consideration of recommendation-oriented metrics or empirical validation, which limits their applicability in sensitive clinical settings. This research addresses these gaps by developing a deep learning–based recommendation framework that integrates structured maternal health features and is rigorously evaluated using appropriate recommendation metrics.

Problem statement: Recent yoga recommendation frameworks for pregnant women generally do not provide efficient, personalized, safety-aware, or trimester-specific recommendations. Many rely solely on textual metadata or pose categories and do not leverage advanced modeling techniques. Systems built using BERT, k-Means, or CNNs exhibit limited detection accuracy, tend to overfit, and cannot account for pregnancy-specific constraints. There remains a clear research gap for developing systems capable of delivering safe, personalized yoga recommendations that jointly consider semantic, visual, and maternal health features along with the user’s trimester stage.

Proposed methodology

In this section, we propose a novel trimester-aware yoga video recommendation framework that leverages hybrid deep learning to provide clinically safe and personalized yoga pose suggestions for pregnant women. Unlike conventional yoga or video recommendation models, the framework introduces two key innovations: a Trimester-Weighted Wasserstein Similarity (TW-WD) mechanism and a Safety-Aware Directed Graph Convolutional Relational Neural Network (GCRNN). The TW-WD adaptively aligns user and video embeddings by weighting similarity according to trimester-specific physiological safety, enforcing stricter filtering during early trimesters and progressively relaxed constraints in later stages. Concurrently, the Safety-Aware GCRNN extends conventional GCNs by encoding directed relationships among users, poses, and videos that reflect both trimester and health conditions, enabling propagation of safety information across the relational graph.

The workflow begins with preprocessing of video and text data to ensure consistency and quality. Text metadata are processed using stop word removal, tokenization, and lemmatization, while video frames are extracted and denoised using the Fused Probabilistic Wiener Filter (FProbWF) to preserve fine posture details. Features are extracted from text via Boosted Term Frequency–Inverse Document Frequency (BTF-IDF) and from video frames using ResNet152 for deep visual embeddings. These multimodal features, combined with user profile attributes, are integrated and compared using the proposed trimester-weighted Wasserstein similarity. The fused embeddings are subsequently processed by the safety-aware GCRNN to generate trimester-specific, personalized, and medically safe yoga recommendations. The overall architecture of the proposed framework is illustrated in Fig. 1.

Fig. 1.

Block diagram of the proposed method.

The key innovation of this work lies in its trimester-weighted multimodal similarity learning combined with graph-based relational reasoning, which incorporates pregnancy safety constraints directly into the model’s learning process. To achieve this, we propose a trimester-weighted Wasserstein similarity function that conditions multimodal similarity computations on the pregnancy stage. This mechanism enables stricter filtering of potentially unsafe poses during early trimesters while permitting controlled flexibility in later stages. The similarity function is formally defined as follows:

1

In this formulation, denotes the Wasserstein distance between textual and visual feature distributions, represents the contextual relational correlation learned through the graph network, and is a trimester-dependent safety coefficient. This design ensures that similarity evaluation jointly captures semantic relevance, visual alignment, and trimester-specific physiological safety, allowing the network to prioritize pose–video pairs that comply with trimester-specific constraints.

Furthermore, the Graph Convolutional Relational Neural Network (GCRNN) is extended to accommodate heterogeneous node types—users, poses, and videos—with directed edges representing trimester and health-condition relationships. This enables the propagation of relational safety information across the graph. Together, these mechanisms establish a technically novel multimodal–graph fusion pipeline that performs trimester-aware similarity alignment and relational inference within a unified optimization framework—capabilities that are absent in conventional video or yoga recommendation systems.

Pre-processing

Preprocessing of both text and video data is a crucial step for ensuring consistency and quality in recommending yoga poses from YouTube videos for pregnant women. Text preprocessing reduces variability in video metadata and is performed through three main steps: stop word removal, tokenization, and lemmatization. Video preprocessing focuses on enhancing frame quality by removing noise from the extracted individual frames. In this work, a Fused Probabilistic Wiener Filter (FProbWF) is applied for denoising, preserving fine posture details and improving the clarity of video images for accurate feature extraction and subsequent recommendation.

Data pre-processing

The three major phases like stop word removal, tokenization, and lemmatization^31,32 are discussed below:

Phase 1: Stop word removal

Stop words are common words used to connect or complete sentences in textual data, but they generally carry little semantic significance. Since they do not contribute meaningfully to most natural language processing tasks, these words are typically removed to reduce noise and improve computational efficiency. Eliminating stop words does not compromise the semantic content of the original text; rather, it helps focus on the meaningful terms. In the proposed method, stop words are removed by comparing each token in the text against a predefined stop word list, ensuring that only relevant words are retained for subsequent feature extraction and analysis.

Phase 2: Tokenization

After stop words are removed, the text is tokenized by separating the significant words from the input. Tokenization is a critical step, as it identifies the key terms that contribute most to the semantic meaning of the text. This process can be challenging for languages without explicit word boundaries, as well as for texts containing symbols or special characters. In the proposed method, each sentence is first divided into a list of words, and the meaningful tokens are subsequently extracted from this list to be used in downstream feature extraction and analysis.

Phase 3: Lemmatization

Lemmatization is the process of converting a word into its base or dictionary form while preserving its contextual meaning. After tokenization, each word is analyzed, and complex or inflected forms are replaced with their corresponding lemmas. Although this step may appear minor, it improves overall system performance by reducing redundancy in the vocabulary, enhancing consistency in textual representation, and increasing processing efficiency for downstream feature extraction and similarity computations.

Video-to-frame extraction and preprocessing

Each collected video was converted into a sequence of RGB frames using FFmpeg for frame extraction and OpenCV for resizing and normalization. Frames were uniformly sampled at a fixed rate of one frame per second to capture representative yoga poses while keeping feature computation tractable. All frames extracted from a given video were treated as a single unit during dataset partitioning. Frames were resized to 224 × 224 pixels to match the input requirements of ResNet152, and pixel values were normalized using ImageNet statistics (mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225])³³.

To enhance frame quality, each image was denoised using the Fused Probabilistic Wiener Filter (FProbWF), which incorporates Gaussian mask kernels, local mean filtering, and sliding-window variance adjustment³⁴. This approach divides each frame into smaller pixel groups, processes each group individually, and effectively reduces noise while preserving fine posture details. For videos with fewer than two frames, all available frames were used; for longer videos, uniform sampling at 1 fps was applied to reduce computational costs and eliminate redundancy³³. Denoised and normalized frames were then exported as JPEG or PNG files and forwarded to the feature extraction pipeline for subsequent processing. Let denote an extracted video frame with pixel coordinates . To account for acquisition and compression noise, the observed frame is modeled as the sum of the clean image and additive Gaussian noise:

2

where represents Gaussian noise at pixel . To estimate the corresponding Gaussian standard deviation, convolution with a Gaussian mask is applied to the frame image, enabling noise-aware preprocessing prior to feature extraction. A predetermined noise threshold is used to determine the mask size, allowing selection among different kernel options. Subsequently, mean filtering is applied using a local mask via convolution with the selected kernels. A sliding-window subtraction technique is then employed to generate a variance-adjusted image, ensuring uniform distribution by leveraging local variance features to produce balanced outcomes. Finally, a denoised image is obtained by averaging the results from the uniform distribution, effectively reducing noise while preserving important posture details.

Before preprocessing, each video was reviewed to ensure that only pregnancy-safe yoga poses were included in the dataset. This screening process followed established prenatal yoga safety guidelines and expert recommendations, ensuring that all retained videos were suitable for pregnant participants.

Feature extraction

From the preprocessed text and extracted keyframes, relevant features for yoga pose recommendation are obtained by integrating Boosted Term Frequency–Inverse Document Frequency (BTF-IDF) for text and ResNet152 for visual data. These complementary models are employed to capture both semantic and visual information critical for accurate, personalized recommendations. The following sections provide a detailed description of each feature extraction model and its role in the proposed framework.

Feature extraction from text using BTF-IDF

The BTF-IDF approach is employed to enhance the representation of textual data for feature extraction. In this process, raw input text is transformed into a set of features, each capturing distinct semantic information. Traditional feature extraction methods often require significant computation time, have limited capacity to learn high-level features, and may lose critical information during processing. To address these limitations, the proposed study utilizes efficient techniques to derive meaningful features from preprocessed text. Boosted Term Frequency–Inverse Document Frequency (BTF-IDF)³⁵ enables semantic feature extraction by quantifying the importance of each word in the dataset. High-quality feature representation is achieved by combining unigram and bigram models, which capture both individual words and word-pair relationships. Additionally, advanced dimensionality reduction techniques are applied to mitigate information loss, ensuring compact and informative feature vectors suitable for downstream recommendation tasks.

Boosted term frequency-inverse document frequency (BTF-IDF)

Textual features are extracted using Term Frequency–Inverse Document Frequency (TF-IDF), also referred to as the inverse document frequency weight feature. The term frequency (TF) component measures how often a specific term occurs within a text, helping to normalize for longer documents³⁶. The inverse document frequency (IDF) component provides a global evaluation by quantifying the importance of a term across the entire dataset. By combining TF and IDF, the method assigns weighted importance to terms, enhancing the representation of key textual features for downstream recommendation tasks. To improve computational efficiency while retaining discriminative textual information, a normalized weight is assigned to each textual feature, as defined in Eq. (3)

3

where denotes the normalized weight of the -th textual feature, represents the average term frequency, is the total number of documents, and denotes the average document frequency of the corresponding term. This formulation emphasizes informative terms while suppressing redundant features, thereby reducing computational overhead during multimodal fusion. Furthermore, Q and q_i indicate the quantity text associated with a certain feature item. Additionally, the featured item’s frequency of occurrences is designated. The TF-IDF approach can successfully differentiate the categories if a term occurs frequently. However, as TF-IDF ignores the semantic meanings and spatial context of words, its classification ability is limited. In the TF-IDF paradigm, this results in a drop in both recall and precision. The proposed work extends the conventional BTF–IDF formulation to more effectively evaluate the importance of textual feature items by incorporating additional contextual and linguistic factors. This enhanced weighting scheme is defined in Eq. (4):

4

where denotes the significance of the -th textual feature, represents the term frequency of the feature, is the total number of documents, and denotes the average document frequency. The parameter accounts for feature length normalization, captures the frequency of occurrence at the -th position, and represents the part-of-speech weighting factor, enabling linguistically informed feature emphasis.

ResNet152 for image feature extraction

After preprocessing the video frames, image features are extracted to represent visual information for yoga pose recommendation. While numerous traditional feature extraction methods exist, they are often time-consuming and require expert intervention. To address these limitations, this work employs a deep learning–based ResNet152 model^37,38. ResNet152 leverages residual connections and identity mapping to enable training of very deep networks without increasing training error or encountering gradient vanishing problems, improving classification accuracy without adding model complexity. The architecture of ResNet152 is illustrated in Fig. 2.

Fig. 2.

Structure of ResNet152 model.

ResNet152, part of the ResNet family of deep neural networks, consists of 1 convolutional layer, 1 max pooling layer (MPL), 1 average pooling layer (APL), and 1 fully connected layer (FCL)³⁹. The convolutional layer extracts hierarchical features using multiple kernel sizes, while the MPL removes redundant information. The APL smooths the feature maps, and the FCL performs final classification. Due to its superior accuracy compared to other ResNet variants, ResNet152 is employed in this study to extract robust visual embeddings from yoga video frames.

Even with extremely DNNs, the vanishing gradient issues can be resolved using ResNet’s residual learning unit structure, which reduces the deprivation DNN. With a shortcut link that adds new inputs and generates new outputs, this unit’s architecture is a feed-forward network⁴⁰. The neural network block employed in this work follows the residual learning principle of ResNet, where the input feature map is propagated through the network to learn a residual mapping rather than a direct transformation. Let denote the desired underlying mapping and represent the residual function learned by the network. The residual is defined as the difference between the target output and the input feature:

5

Rearranging Eq. (5), the original mapping can be expressed as:

6

This residual formulation enables the network to focus on learning the precise corrective component required to refine the input features, thereby facilitating stable gradient propagation in deep architectures. Multiple stacked residual blocks, each composed of three convolutional layers, are employed to construct the ResNet-152 backbone used in the proposed framework. By increasing the network’s capacity, it enhances model creation and maintains the knowledge accumulated during training. The layers learning the residual P(y), and the identity connection is assigned because of y. While the residual network levels are learning the residual Q(y), the layers that correspond to the conventional network are learning the true output Q(y). It is significantly simpler to learn input and output residual than to analyze only input. Since previous layers are skipped, the identity residual model allows the activation function to be reused without adding complexity to the framework⁴¹.

Loss function in ResNet152 network model

Accurate recommendation of yoga poses for pregnant women can be adversely affected by loss of function during training. The ResNet152 model mitigates this issue by effectively minimizing the loss function, enabling the multimodal feature system to perform reliably. However, continuous emphasis of the loss function on negative frames can reduce the true positive rate, causing slight imbalance during network optimization. This imbalance may result in newly updated parameters surpassing the initially degraded parameters. Addressing such challenges ensures stable training and improves the overall robustness and accuracy of the pose recommendation system. Consequently, the impact of class imbalance across frames is mitigated through a weighted loss formulation, enabling stable optimization even in the presence of uneven foreground–background distributions. The loss function employed is defined in Eq. (7)

7

where denotes the predicted pixel probability, represents the corresponding binary ground truth mask, is a balancing coefficient controlling class importance, and and are hyperparameters that regulate the contribution of hard-to-classify samples. This formulation effectively suppresses the dominance of majority classes while emphasizing informative regions during training.

Frame-level feature extraction and temporal aggregation

Each frame was denoised using the Fused Probabilistic Wiener Filter (FProbWF). The last classification layer of ResNet152 was discarded, and the output of the global average pooling layer was used as a 2048-dimensional feature vector. For each video, a video-level embedding was obtained by averaging the frame-level feature vectors across all sampled frames, providing a compact and robust representation. These embeddings were then fused with textual features and passed to the GCRNN via the Wasserstein distance-based similarity computation module. This approach produces stable pretrained spatial representations even on a relatively small dataset while maintaining moderate training time and memory requirements.

Compared to 3D convolutional networks such as C3D, I3D, 3D-ResNets, or transformer-based video encoders like ViViT, ResNet152 is more efficient for the current task. Although these models can capture spatial and temporal dynamics, they require large annotated datasets, extensive training, and high GPU memory. Given that the dataset comprises 200 videos, which primarily rely on static spatial arrangements of individual frames to represent basic yoga poses, ResNet152 with transfer learning provides stable spatial embeddings, reduces overfitting, and lowers computational demands. Temporal and relational patterns are subsequently encoded using GCRNN-based sequence learning, frame sampling, embedding aggregation, and Wasserstein similarity, eliminating the need for complex spatio-temporal models while effectively capturing both spatial and temporal features.

By computing TF-IDF and Wasserstein distance across modalities, the framework enables cross-modal similarity learning. Textual data, including user search queries and yoga pose descriptions, are represented via TF-IDF, which transforms inputs into sparse vectors that highlight highly informative words. ResNet152 converts video frames into 2048-dimensional embeddings, which are L2-normalized and treated as visual probability distributions. The Wasserstein distance (also called Earth Mover’s Distance) measures similarity between video embeddings and user query representations by computing the minimum cost of transforming one distribution into another. This framework allows TF-IDF to capture semantic similarity in text space and Wasserstein distance to capture visual similarity, providing an effective method to map user queries to corresponding yoga video frames.

Multimodal feature fusion

The textual and visual feature representations extracted through BTF-IDF and ResNet152, respectively, are fused to obtain a unified multimodal embedding that can be effectively propagated in the GCRNN layer. Let denote the semantic feature vector obtained from the BTF-IDF module and represent the visual embedding extracted from the ResNet152 network (output of the global average pooling layer).

To ensure feature comparability across modalities, the extracted textual and visual features are first normalized using normalization:

8

The normalized features are subsequently projected into a shared latent space of dimension through learned linear transformations and , yielding:

9

The multimodal representation is formed using a weighted concatenation strategy that balances textual and visual contributions

10

where denotes vector concatenation, and and are learnable modality-balancing coefficients satisfying . This fusion strategy enables adaptive alignment between semantic and visual information during training. The resulting joint embedding forms the input node representation for the GCRNN layer. This fused embedding preserves both the semantic context from textual metadata and the spatial–visual context from yoga pose frames, enabling the relational graph to capture trimester-specific relationships among users, poses, and videos. By explicitly defining the projection and fusion processes, the integration across modules becomes mathematically well-grounded, and the information flow into the GCRNN is clearly established, ensuring effective downstream relational reasoning and safety-aware recommendation.

Wasserstein distance similarity

After feature extraction, Wasserstein distance similarity is employed to compare feature distributions in the proposed framework. Each user —characterized by profile attributes such as trimester, difficulty level, and health condition—and the extracted features from text metadata and yoga pose frames are represented as probability distribution vectors. Similarity is then assessed by computing the Wasserstein distance between these distributions. A strong equivalence between feature distributions is indicated by lower Wasserstein distance values, effectively linking entity representations based on distributional similarity^39,40. This approach quantifies the minimum cost required to transform one probability distribution into another, providing a principled measure for matching user profiles with appropriate yoga poses in a safety-aware recommendation system.

Let q∈R^N and r∈R^N signify two entities’ distribution (either users or different preferences. The cost of aligning features between q and r is encoded in the transportation cost matrix Co ∈R^KxN. To compare user preferences across modalities, we first compute the Frobenius inner product between matrices representing the distributions. Given matrices and , the Frobenius inner product under the ideal transport strategy is defined as:

11

where denotes the Frobenius dot-product. This provides a measure of alignment between two distributions. Building on this, we compute the Wasserstein distance between distributions as a measure of dissimilarity. Formally, the distance between distributions and is given by:

12

where represents the set of admissible couplings between and . Intuitively, this captures the minimum “cost” of transporting one distribution to match the other. Finally, to convert the Wasserstein distance into a similarity score suitable for fusion, we normalize the distance matrix. Specifically, the similarity matrix is computed as:

13

where is the computed Wasserstein distance matrix between each pair of users or preferences, is a matrix of ones of appropriate size, and denotes the maximum distance value. This normalization ensures that larger distances correspond to lower similarity scores, and vice versa.

Yoga pose recommendation using GCRNN

Following similarity computation, a Graph Convolutional Relational Neural Network (GCRNN) is employed to enhance the accuracy of yoga pose recommendations. The GCRNN improves representation quality by capturing structured relationships among users (pregnant women with varying trimesters and health conditions) and items (YouTube videos) through graph-based connectivity. By modeling these relational dependencies, the network can propagate information about trimester-specific safety and personalized preferences across the graph. The architecture of the GCRNN is illustrated in Fig. 3.

Fig. 3.

Structure of GCRNN for yoga recommendation.

In this framework, the GCRNN learns node embeddings by aggregating multi-relational information from adjacent nodes while accounting for different types of relationships. The resulting embeddings capture complex dependencies among users, poses, and videos, enabling the network to propagate safety- and trimester-aware information throughout the graph. A fully connected layer then generates personalized yoga pose recommendations based on these learned node embeddings. The following section provides a concise mathematical formulation of the model .

GCRNN: In this framework, videos, poses, users, and health conditions are modeled as graph nodes, while edges represent relational information (e.g., “video contains pose,” “pose alleviates back pain,” “user has back pain”). The GCRNN learns these relational structures and propagates the importance of information across the graph, enabling the identification of videos most relevant to a given user. By integrating content features (textual and visual) with relational knowledge (health requirements, poses, and videos), the proposed model delivers highly personalized and safe yoga recommendations tailored to the specific trimester and health needs of pregnant women.

Following graph convolution theory⁴¹, for a sample node with a convolutional kernel parameterized in the Fourier domain as , the spectral graph convolution is expressed as:

14

where contains the eigenvectors and is the diagonal eigenvalue matrix of the graph Laplacian . Intuitively, this convolution aggregates information from adjacent nodes, since the Laplacian encodes the global structure of the graph.

To reduce computational complexity, can be approximated using Chebyshev polynomials, which allows efficient calculation of spectral convolutions. By constraining the number of polynomial coefficients, we define the propagation of a single layer.

Consequently, the layer-wise update of node features can be formulated as:

15

where represents the hidden layer activations of node at layer , denotes the neighbors of node , and are learnable weight matrices, and is the activation function (here chosen as the ReLU function). This formulation intuitively propagates and updates node features by combining information from the node itself and its neighbors.

The GCRNN, based on the Graph Convolutional Network (GCN) and illustrated in Fig. 4, is enhanced by incorporating edge relationship types, transforming the model from an undirected graph into a practical directed graph^20,21. Unlike the original GCN, the relations in GCRNN are directed, but both each relation and its inverse are included in the adjacency matrix for representation. To effectively capture heterogeneous relational information, the model aggregates neighboring nodes using a weighted sum across different relation types, allowing the network to propagate information according to both the direction and nature of each edge^25,26. To prevent overfitting, the model regularizes the weight matrices during training. In the context of knowledge graphs, GCRNN operates on a graph defined as , where denotes the set of nodes (entities), represents the set of relation types, and contains edges labeled with relation types .

Fig. 4.

Architecture and relational structure of the proposed GCRNN. Blue nodes denote users, green nodes denote yoga poses, and edges represent trimester- and health-condition-specific relations link in users, poses, and videos.

The GCRNN propagation rule updates the hidden state of each node (entity) by aggregating information from its neighbors across different relation types:

16

Here: is the hidden state of node at layer , is the weight matrix corresponding to relation type , is the set of neighbors of node , is a nonlinear activation function (e.g., ReLU).

To reduce the number of parameters, GCRNN represents the relation-specific weights as a linear combination of basis transformations:

17

where are trainable basis matrices shared across relations, and are learnable coefficients. This approach allows the model to efficiently handle multiple relation types without introducing excessive parameters. Building on this, an improved propagation rule further enhances the aggregation by incorporating additional relational and virtual node information:

18

where are trainable basis matrices shared across relations, and are learnable coefficients. This approach allows the model to efficiently handle multiple relation types without introducing excessive parameters. Building on this, an improved propagation rule further enhances the aggregation by incorporating additional relational and virtual node information:

Results and discussion

To evaluate the performance of the proposed model in recommending yoga poses for pregnant women, experimental assessments were conducted. Multiple metrics—including accuracy, precision, recall, F1-score, specificity, mean squared error (MSE), mean absolute error (MAE), and root mean squared error (RMSE)—were used to comprehensively assess each model. The recommendation performance of the proposed framework was compared against established models, including Deep Belief Network (DBN), Deep Neural Network (DNN), Gaussian Support ResNet (GS-ResNet), Restricted Boltzmann Machine with Autoencoder (ResBM-AE), and Relation-Aware Attentional Gated Convolutional Neural Network (ReAAtt-GCNN). A detailed comparison of these metrics across models is provided in this section. System design specifications for the proposed model are summarized in Table 1, and hyperparameter values are listed in Table 2. Additionally, Fig. 5 illustrates various yoga poses recommended for pregnant women in the first, second, and third trimesters.

Table 1.

System configuration used for model implementation and experimental evaluation.

System parameter	Configuration
Processor	Intel(R) Core (TM) i5-7500 CPU @ 3.00 GHz 3.00 GHz
System type	64-bit operating system, x64-based processor
Installed RAM	16.00 GB (15.8 GB usable)
Pen and touch	No pen or touch input is available for this display

Table 2.

Hyperparameter settings used for training and evaluation of the proposed model.

Hyperparameter	Value
Learning rate	0.01
Dropout	0.2
Training data	80%
Testing data	20%
Epochs	100
Normalization	Batch Normalization
Optimizer	Adam (β₁ = 0.9, β₂ = 0.999)

Fig. 5.

Different yoga poses for pregnant women.

The architectural and experimental configuration details are consolidated in Table 3.

Table 3.

Experimental configuration.

Parameter	Value / Setting
Node types	User (pregnant profile), Yoga Pose, Video – 3 types
Relation types	user–pose, pose–video, user–video, health-condition (with inverse edges) – 4 relations
Hidden-layer dimensions	128 → 64 → 32
Total trainable parameters	≈ 2.4 million
Average training time	≈ 38 min per epoch

Dataset description

The Pregnancy Yoga Pose Recommendation Dataset comprises 200 carefully curated yoga videos sourced from YouTube and the Yoga Download platform. YouTube provides diverse, real-world instructional content that enhances the ecological validity of the proposed framework, while Yoga Download offers professionally structured prenatal yoga sessions, ensuring content reliability and instructional consistency. Each video in the dataset is annotated with attributes such as recommended trimester, difficulty level (easy, medium, hard), and primary health benefit (e.g., back pain relief, stress reduction, flexibility, or strength building). Although the dataset contains only 200 videos, temporal segmentation produced over 35,000 valid frames, substantially increasing the effective data volume for robust model training. The dataset supports personalized recommendations by integrating content-based textual features with user-specific factors, including pregnancy stage and health goals. Unsafe poses—such as prolonged supine positions, deep twists, and extreme abdominal compression—were excluded based on guidance from a certified prenatal yoga instructor and review by a maternal health specialist. All videos are publicly available, contain no personally identifiable information, comply with ethical data usage standards, and include source URLs for transparency. Future work will expand the dataset across additional platforms to improve coverage and generalizability.

Expert validation involved a panel of five domain experts, three obstetricians and two certified prenatal yoga instructors, each with over five years of experience who independently reviewed a representative subset of system-generated recommendations across all trimesters. Recommendations were assessed for correctness (alignment with user profile), clinical relevance (adherence to prenatal yoga safety guidelines), and safety (absence of contraindicated postures). This comprehensive dataset construction and expert evaluation ensure that the system provides reliable, safe, and clinically relevant yoga recommendations for pregnant women.

The sample data collected from YouTube for recommending yoga poses for pregnant women is given in Fig. 6.

Fig. 6.

Sample data collected from YouTube for recommending yoga poses for pregnant women.

Performance indicators and their formulation

Each performance metric and its predicting methodology are covered in detail in this section. Its mathematical formulation is exposed in Table 4. Here, true positive (TP_pos), true negative (TN_neg, false positive FP_pos, false negative FP_neg, true value , prediction y_i, and total amount of data points n^28–30.

Table 4.

Performance indicators and their formulation.

Metrics	Description	Formulation
Accuracy	It defines how many correctly identified samples in each class.
Recall	It uses all of its initial samples to calculate the number of correctly recognized samples in a class.
Precision	It calculates the correctly classified samples from the detecting class’s true positive samples.
Specificity	It determines the proportion of genuine negatives that the model correctly identified.
F-Measure	It is determined using the precision and recall of the model to determine its overall efficiency.
MAE	It determines the average absolute variance between the expected and actual results.
RMSE	The average squared difference between the expected and actuaøl values is computed.
PSNR	PSNR quantifies overall fidelity between the denoised image and the original clean image
SSIM	SSIM evaluates perceptual similarity between two images based on luminance, contrast, and structural information.

The Normalized Discounted Cumulative Gain (NDCG@K) is employed to evaluate ranking quality by assigning higher importance to relevant items that appear earlier in the recommendation list. The Discounted Cumulative Gain (DCG@K) is defined as the sum of graded relevance values, discounted logarithmically based on the position of each item in the ranked list. Formally,

19

where denotes the relevance of the item at position .

Assessment in relation to various evaluation metrics

To quantitatively justify the use of the Fused Probabilistic Wiener Filter (FProbWF) for video pre-processing, a comparative experiment was conducted against three conventional denoising algorithms: the standard Wiener Filter, Gaussian Blur, and Median Filter. As summarized in Table 5, each method was applied to a representative subset of extracted yoga video frames (), and denoising performance was evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). Higher values of PSNR and SSIM indicate superior noise suppression and better preservation of structural details, demonstrating the effectiveness of FProbWF for maintaining the quality of yoga pose frames.

Table 5.

Denoising comparison.

Filter type	PSNR (dB)	SSIM
Standard Wiener tilter	26.8	0.87
Gaussian blur	25.9	0.84
Median filter	27.4	0.88
FProbWF (Proposed)	31.2	0.93

The proposed FProbWF achieved a PSNR of 31.2 dB and an SSIM of 0.93, outperforming the standard Wiener Filter (26.8 dB / 0.87), Gaussian Blur (25.9 dB / 0.84), and Median Filter (27.4 dB / 0.88). This represents a 16–20% improvement in PSNR and a 5–9% improvement in SSIM, demonstrating FProbWF’s superior ability to suppress noise while preserving fine pose structures for more reliable ResNet152 feature extraction. To ensure methodological rigor and reproducibility, all experiments were conducted using a fivefold cross-validation strategy, with 80% of the data used for training and 20% for testing in each fold. Frame-level modeling captured intra-video pose variability, while cross-validation and evaluation were performed strictly at the video level to prevent data leakage.

Table 6 presents the fivefold cross-validation results for accuracy, precision, recall, F1-score, and specificity. Performance was consistent, with a mean accuracy of 98.30% (range 98.10–98.50%), confirming the stability and reliability of the proposed model.

Table 6.

Performance of the proposed model under 5-fold cross-validation.

Fold	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Specificity (%)
1	98.2	97.4	98.1	97.7	97.7
2	98.4	97.6	98.3	97.8	97.9
3	98.3	97.5	98.2	97.7	97.8
4	98.1	97.3	98	97.6	97.7
5	98.5	97.7	98.4	97.9	97.9
Mean ± SD	98.3 ± 0.14	97.5 ± 0.15	98.2 ± 0.14	97.7 ± 0.11	97.8 ± 0.09

Precision and recall remain stable, with mean values of 97.50% and 98.20%, respectively, yielding an F1-score of 97.70%. The mean specificity of 97.80% further confirms the reliability of the model. These results demonstrate the validity and generalizability of the proposed methodology across different data partitions.

Statistical reliability and confidence analysis

To enhance the statistical rigor of the evaluation, the reliability of the proposed trimester-aware yoga recommendation framework was assessed using fivefold cross-validation. All primary performance metrics are reported as mean ± standard deviation, reflecting variability across data partitions. To further quantify the stability of the results, 95% confidence intervals (CI) were computed for each metric, providing a robust measure of statistical reliability:

The proposed model achieved a mean accuracy of 98.30%, with a narrow 95% confidence interval (CI) of [98.18%, 98.42%], indicating strong consistency and minimal variance across folds. Similarly tight confidence intervals were observed for precision, recall, F1-score, and specificity, demonstrating the model’s robust generalization performance and reliability across different data partitions.

Baseline methods were evaluated using reported aggregate performance values, whereas the proposed framework was assessed via repeated fivefold cross-validation. Even under this conservative evaluation setting, the proposed model demonstrates consistent and substantial performance gains across both classification and recommendation metrics. The low variance observed across folds confirms that these improvements are stable and not due to random variation. Furthermore, ranking-based metrics—including Precision@K, NDCG@K, and GAP@K—show uniform improvements across multiple cut-off levels, highlighting reliable recommendation behavior. Expert validation further supports the clinical relevance of the system, with an agreement rate of 94.3%. Collectively, these results establish the robustness, reliability, and practical effectiveness of the proposed framework.

Comparative analysis against existing models, DNN, DBN, GS-ResNet, ResBM-AE, and ReAAtt-GCNN—illustrated in Figs. 7a–d, further demonstrates the approach’s effectiveness. Across all evaluation metrics, the proposed model consistently achieves superior performance, capturing complex user–pose–video relationships more effectively than existing methods. Specifically, it attains 98.3% accuracy, outperforming all baselines, along with 98.2% precision, 97.5% recall, and 97.7% F1-score for yoga pose recommendation in pregnant women. By integrating textual and visual features through advanced feature extraction and a GCRNN-based recommendation model, the framework effectively models multi-hop user–item interactions, reduces data ambiguity and redundancy, and surpasses existing methods in both accuracy and overall performance.

Fig. 7.

Assessment of (a) Accuracy, (b) Precision, (c) Recall, and (d) F1-score.

Prevailing deep models, such as DNN (AUC = 0.872) and DBN (AUC = 0.911), performed reasonably well but exhibited limitations in balancing precision and recall. Hybrid approaches, including GS-ResNet (AUC = 0.936), ResBM-AE (AUC = 0.922), and ReCAtt-GCNN (AUC = 0.925), achieved improved performance by capturing deeper representations and contextual information. The proposed hybrid model (AUC = 0.942) outperformed all existing methods, demonstrating higher precision across all recall levels. By integrating textual features (TF-IDF), visual pose embeddings (ResNet152), Wasserstein similarity, and GCRNN, the framework effectively captures structured user–pose–video relationships, improving recommendation accuracy. Consequently, it delivers precise, pregnancy-safe yoga video recommendations, minimizing irrelevant suggestions while ensuring that all recommended videos are suitable for the user’s trimester and health condition Fig 8.

Fig. 8.

Assessment of specificity.

Prevailing deep models, such as DNN (AUC = 0.872) and DBN (AUC = 0.911), performed reasonably well but were limited in balancing precision and recall. Hybrid deep models, including GS-ResNet (AUC = 0.936), ResBM-AE (AUC = 0.922), and ReCAtt-GCNN (AUC = 0.925), achieved improved performance by capturing deeper representations and contextual dependencies. The proposed hybrid model (AUC = 0.942) outperformed all existing methods, demonstrating higher precision across all recall levels. By integrating self-contained features—textual features (TF-IDF), visual pose embeddings (ResNet152), Wasserstein similarity, and GCRNN—the framework effectively captures structured user–pose–video relationships, enhancing recommendation accuracy. Consequently, the system provides precise, pregnancy-safe yoga video suggestions, minimizing irrelevant recommendations while ensuring that all recommended videos are appropriate for the user’s trimester and health conditions.

Table 7 presents a comparison of recommendation quality between baseline models and the proposed framework. Across all evaluation metrics, Precision@K, NDCG@K, and GAP@K—the proposed method consistently outperforms the baselines. Specifically, the model achieves a Precision@5 of 0.980, surpassing baseline values of 0.975, 0.972, and 0.970. Similarly, the NDCG@10 of the proposed approach is 0.978, exceeding the nearest baseline of 0.975, with other baselines at 0.969 and 0.968. The GAP@20 reaches 0.972, demonstrating improved ranking performance and recommendation relevance, as illustrated in Fig. 9. These gains highlight the model’s effectiveness in capturing both retrieval quality and user-centered relevance. Performance stability is achieved through decoupled similarity measurement across modalities: TF-IDF captures textual features, while Wasserstein Distance provides a principled comparison of video embeddings in distributional space. This separation ensures robust representation of user queries and yoga pose descriptions, maintains conceptual consistency, prevents methodological crosstalk, and enables efficient handling of heterogeneous data.

Table 7.

Recommendation quality of the baselines with the proposed model.

Method	P @5	NDCG @5	GAP @5	P @10	NDCG @10	GAP @10	P @20	NDCG @20	GAP @20
GS-ResNet	0.975	0.978	0.976	0.97	0.975	0.974	0.965	0.973	0.972
ResBM-AE	0.972	0.975	0.973	0.968	0.972	0.971	0.962	0.97	0.969
ReAAtt-GCNN	0.97	0.973	0.971	0.965	0.97	0.97	0.961	0.969	0.968
Proposed	0.98	0.981	0.979	0.975	0.978	0.976	0.97	0.975	0.972

Fig. 9.

Precision-recall analysis.

Figure 10 illustrates the Precision@K scores of the proposed yoga pose recommendation system for pregnant women. Precision@K reflects the proportion of correctly recommended videos within the top-K suggestions. At K = 1, the system achieves very high precision (0.98), indicating that the first recommended video closely aligns with the user’s trimester and health requirements. As K increases (K = 3, 5, 10), precision gradually declines, reflecting the inclusion of less closely aligned results. Nevertheless, even at K = 10, precision remains above 0.88, demonstrating the robustness of the proposed framework in consistently recommending clinically safe and trimester-appropriate yoga videos.

Fig. 10.

Analysis of Precision@K scores.

Table 9 presents a comparison of RMSE and MSE across baseline models and the proposed framework. The proposed method achieves a mean squared error (MSE) of 0.0289, outperforming DNN, RBM-AE, MBO-DBN, and RAAGCN. Similarly, the root mean square error (RMSE) for the proposed yoga pose recommendation system is 0.17, indicating lower prediction errors relative to existing approaches. By integrating textual and visual features, the model effectively captures complex user–pose–video interactions while mitigating noise and redundancy. The deep architecture further identifies non-linear correlations in the data, reducing errors and improving recommendation accuracy. These results demonstrate the model’s ability to provide precise, trimester-appropriate, and clinically safe yoga pose recommendations Fig 11.

Table 9.

MSE and RMSE analysis.

Model	MSE	RMSE
DNN	0.7225	0.85
DBN	0.3844	0.62
GS-ResNet	0.2601	0.51
ResBM-AE	0.1521	0.39
ReAAtt-GCNN	0.09	0.3
Proposed	0.0289	0.17

Fig. 11.

Analysis of safety compliance percentage.

The high safety compliance observed across trimesters—including 99.18% compliance in the first trimester—stems from the integration of the Trimester-Weighted Wasserstein Distance (TW-WD), which enforces stricter similarity constraints during early pregnancy, and the Safety-Aware GCRNN, which propagates trimester- and health-specific safety relationships across the recommendation graph. This combination systematically suppresses unsafe pose recommendations. The training and testing procedures, illustrated in Fig. 12, employed an 80/20 split of the dataset. The model achieved low training loss and high accuracy, demonstrating effective learning and error reduction. Consistently low testing loss with strong testing accuracy indicates robust generalization to unseen data, while the alignment between training and testing results confirms stability against noise and overfitting, ensuring reliable predictions.

Fig. 12.

Training and testing analysis.

To complement the quantitative evaluation, a panel of gynecologists and obstetricians reviewed a representative subset of system recommendations. Expert assessments of correctness, clinical relevance, and safety, summarized in Table 8, indicate a high level of agreement, with an overall validation score of 94.3%, confirming the practical reliability and safety of the proposed framework.

Table 8.

Expert validation scores.

Evaluation aspect	Expert agreement (%)
Correctness of recommendations	94.2
Safety of recommendations	95.1
Clinical relevance	93.7
Overall expert validation score	94.3

The proposed framework demonstrates strong performance across ranking metrics, including Precision@K, Recall@K, NDCG@K, and GAP@K. To evaluate clinical relevance and safety, a representative subset of recommendations was independently reviewed by a panel of gynecologists and obstetricians for correctness, adherence to prenatal safety guidelines, and clinical appropriateness, achieving a high level of agreement with an overall expert validation score of 94.3%. These results confirm that the system provides recommendations consistent with established clinical guidelines, reinforcing both its technical reliability and practical credibility. Additionally, Fig. 10 reports a 99.19% safety compliance rate during the first trimester, further demonstrating the framework’s effectiveness in delivering safe, trimester-appropriate, and clinically sound yoga recommendations.

The comparison of RMSE and MSE is presented in Table 9. The proposed method’s MSE of 0.0289 is lower than that of DNN, RBM-AE, MBO-DBN, and RAAGCN. Additionally, the proposed method for yoga pose recommendation achieves an RMSE of 0.17. By combining text and visual features, the model outperformed DNN, RBM-AE, MBO-DBN, and RAAGCN in terms of error rates. Despite noise and redundancy reduction, this enhanced representation still captures intricate interactions. The model’s deep architecture reduces error metrics across evaluation data by successfully identifying non-linear correlations, improving accuracy and minimizing errors.

Limitation and future work

Although frame extraction increased the effective training set to approximately 35,000 frames sampled at 1 frame per second, the dataset comprises only 200 unique source videos. Consequently, the model primarily learns static posture representations rather than long-term motion dynamics, and the reported results should be interpreted as application-specific rather than universally generalizable. Future work will address this limitation by expanding the dataset to include a larger and more diverse collection of standardized prenatal yoga recordings, incorporating controlled clinical video data, higher frame-rate temporal sampling, and cross-dataset validation to evaluate robustness across different recording conditions and instructor styles.

Conclusion

This research presents a hybrid, safety-aware recommendation framework that integrates data preprocessing, advanced feature extraction, and relational graph learning to deliver trimester-specific yoga pose recommendations for pregnant women. The proposed system balances accuracy and safety by employing FProbWF for denoising, BTF-IDF and ResNet152 for multimodal feature extraction, and Wasserstein Distance for similarity measurement, while the GCRNN model captures complex relationships among users, videos, and yoga poses to enable personalized recommendations. Experimental results demonstrate strong performance, achieving 98.3% accuracy, 97.5% recall, 98.2% precision, and a 97.7% F1-score, alongside safety compliance exceeding 97.5% across all trimesters. Beyond quantitative performance, this work underscores the broader significance of embedding explicit safety constraints and expert-validated rules within AI-driven healthcare recommendation systems. The proposed framework provides a scalable and trustworthy blueprint for deploying responsible AI in sensitive domains such as prenatal wellness, where ethical compliance, user safety, and contextual awareness are critical. While this study focuses on pregnancy yoga recommendations, the underlying methodology is readily extensible to other preventive and lifestyle-oriented healthcare applications. Future work will explore transformer-based classifiers incorporating richer user inputs and expand the dataset to include more diverse populations, further enhancing generalizability and real-world applicability.

Author contributions

Conceptualization, K.B. and S.R.; methodology, M.K.; software, M.K.; validation, K.B., S.R. and M.K.; formal analysis, K.B.; investigation, S.R.; resources, K.B.; data curation, M.K.; writing—original draft preparation, S.R.; writing—review and editing, S.R.; visualization, M.K.; supervision, M.K.; project administration, S.R.; funding acquisition, K.B. All authors have read and agreed to the published version of the manuscript.

Funding

 No external Funding for this work.

Data availability

The datasets generated and analyzed during the current study are available in the Rajendran, S. (2026). Yoga Video Recommendation using Hybrid Deep Learning for Pregnant Woman [Data set]. Zenodo. [https://doi.org/10.5281/zenodo.18212874](https:/doi.org/10.5281/zenodo.18212874).

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

Ethical approval was obtained from the Institutional Ethics Committee prior to commencing this study. All procedures were conducted in compliance with relevant guidelines and regulations, ensuring the ethical handling of data and adherence to the sensitivity of pregnancy-related health information.