Prediction of perinatal depression among women in Pakistan using Hybrid RNN-LSTM model

Traditional statistical methods

A study in rural Bangladesh found 56% prenatal depression rate (Insan et al., 2023), strongly associated with intimate partner violence and husband’s preference for male children. The research involved 235 pregnant women from March to November 2021 and employed statistical analysis, which showed that increased family support reduced depression risk. Likewise, Farooq et al. (2022) conducted a study in Lahore, Pakistan, and found that 82% of pregnant women were depressed—significantly higher than the global average of 10–15%. They used the Hamilton Rating Scale for Depression (HRSD) to diagnose depression, highlighting the severity of the mental health concern among pregnant women in Pakistan. A comparison of depression, stress, and anxiety levels in 180 pregnant women using a cross-sectional design and validated instruments (Doty et al., 2022), found significantly higher abnormal anxiety rates among inpatient (IP) compared to low-risk outpatient (LRO) women.

Machine learning-based models

Jigeer et al. (2022) observed a connection between residential noise exposure during pregnancy and an increased risk of prenatal anxiety and depression in the samples collected from Shanghai, China. The researchers found that depression was more common in 46% of women who worked but quit their jobs due to pregnancy. To analyze the data, the multivariate logistic regression (MLR) model was applied, achieving an accuracy of 71%. This suggested that noise exposure could be a significant factor affecting mental health during pregnancy. Ayyub et al. (2018) found that 39.5% of pregnant women in slum settlements of Lahore suffered from prenatal depression, which was associated with food insecurity. Prenatal depression became a common problem in Pakistan, with a prevalence rate of up to 56% in rural areas. In most cases, intimate partner violence during pregnancy and perceived preference of the husband for male gender of a child, were associated with an increased risk of PND symptoms in Bangladesh (Insan et al., 2023).

Rabinowitz et al. (2023) found that depression was higher in women who were working after childbirth, possibly due to the added stress of balancing a career with caring for a new baby. These multiple roles can lead to role overload, which can hurt a postnatal mother’s psychological well-being. Javed et al. (2021) constructed a multi-layer perceptron neural network (MLP-NN) classifier to identify depression in expected women. They used Relief algorithm to identify features from a dataset of 500 women from Pakistan during prenatal period before training the classifier. The results indicated that MLP and support vector classifiers obtained promising results in identifying prenatal depression with 88.0% and 80.0% and prenatal anxiety with 85.0% and 77.7% accuracy, respectively.

Liu et al. (2023) obtained data from a sample of women who underwent cesarean births where the control population was partitioned into cohorts for training and testing. A total of six distinct machine learning models were used to evaluate their predictive efficacy on the testing population. The SHAP approach was employed to interpret the models. The XGBoost model showed an AUROC of 0.789 in the training and 0.744 in the testing with a threshold of 21.5% for postnatal depression risk probability.

Preis et al. (2022) used machine learning algorithms to analyze data from PROMOTE, a novel psychosocial screening tool for prenatal depression. They found that the model could accurately predict depression with an accuracy of 80%, sensitivity of 75%, specificity of 81%, positive predictive value of 79%, and negative predictive value of 97%. Prabhashwaree & Wagarachchi (2022) used four ML models, Feed-Forward Neural Network (FFANN), Adaptive Neuro-Fuzzy Inference System, Genetic Algorithm (ANFIS-GA), random forest (RF), and support vector machine (SVM) on the dataset containing 686 postnatal data of Sri Lankan women. FFANN model showed the best results with 97.08% accuracy.

Garbazza et al. (2023) carried out a prospective cohort investigation into sleep patterns and mood fluctuations during the perinatal phase. Data was collected from 439 pregnant women during their first trimester, encompassing 47 variables related to socio-demographics, psychology, blood markers, medical/gynecological aspects, and subjective sleep experiences. During the second trimester, data was collected from 353 women, comprising 33 parameters. These two distinct datasets were subsequently subjected to a bivariate analysis to assess their correlations with postnatal depression. SVM was used for predicting early pregnancy PND chances with the sensitivity and specificity of 54.3% and 82.6% respectively. However, polysomnographic (PSG) variables did not enhance the model’s predictive accuracy for PND, instead it relied on subjective sleep issues.

Shin et al. (2020) conducted a study using 28,755 records to assess the effectiveness of nine ML algorithms in predicting postnatal depression. The techniques used in the study were stochastic gradient boosting (GB), recursive partitioning and regression trees, naive Bayes (NB), k-nearest neighbors (kNN), logistic regression (LR), RF, SVM, and neural networks. The RF algorithm attained the highest overall classification accuracy, with a value of 79.1%, and the largest area under the receiver-operating-characteristic curve (AUC) value, which was 0.884. The SVM gained the second-highest performance, with an AUC score of 0.864. Qasrawi et al. (2022) used different ML algorithms such as decision tree (DT), SVM, KNN, RF, NB, and GB. They used a dataset of 3,569 women with the highest accuracy of 83.3% on GB (Goodfellow et al., 2020).

Deep learning-based models

Sharma et al. (2022) conducted a study to figure out stress in pregnant women. They used a deep-learning model based on deep recurrent neural networks (DRNNs) to accurately predict whether working pregnant women were stressed. The proposed DRNN model achieved an accuracy of over 96%. This was a new and improved framework for estimating the stress levels of working pregnant women.

Oğur et al. (2023) used different ML and DL models, including NB, DT, RF, gradient boosting tree (GBT), and deep feedforward neural network (DFFNN). The results showed that the DFFNN model achieved 89.60% higher accuracy with 83.26% precision, 80% recall, and 93.33% F1 score and ML model NB achieved 92.45% accuracy. Table 1 presents an overview of the studies related to perinatal depression.

The existing work utilized datasets having class imbalance problems, which generated biased results affecting the performance of the model (Gopalakrishnan et al., 2022; Andersson et al., 2021). Our study aims to develop a novel dataset named PERI_DEP. In contrast to the previous studies, we utilize SMOTE and generative adversarial networks (GAN) for data augmentation and propose an innovative hybrid RNN-LSTM-based deep learning framework. SMOTE generates new samples by interpolating between existing minority samples, while GAN is a powerful neural network that utilizes a generator and a discriminator in an adversarial setting to learn the data distribution and generate high-quality synthetic data (Goodfellow et al., 2014). We have specifically designed this framework for predicting PND, leveraging the rich dataset we have developed.

Dataset development

We gathered data from June 2023 to November 2023 from hospitals in Lahore and Gujranwala regions, employing rigorous questionnaires that included demographic information and mental health assessments (PHQ-9, EPDS). The institutional review board at the Services Institute of Medical Sciences, Lahore Pakistan, approved this study and the approval number was IRB/2023/1155/SIMS. Our PERI_DEP dataset contained 14,008 records from pregnant women and new mothers.

All data was extensively verified and corrected for discrepancies through direct participant interaction to maintain high data integrity. Eligible outpatient department participants were recruited, informed of the study’s purpose, and assured of confidentiality before obtaining their written consent. They then completed questionnaires under the premise of optional participation and data security.

Dataset preprocessing

Current ML and DL algorithms often exhibit a significant decline in performance when confronted with raw data due to their inherent lack of structure and machine-interpretable features. Therefore, preprocessing becomes an essential step in data preparation. In our experiment, data preprocessing includes data cleaning, normalization, one hot encoding, and data labeling. During preprocessing, we identified, 12 samples for which respondents had not provided any response. We removed these samples, leaving 13,996 instances where positive samples are 7,293 and negative samples are 3,903 as shown in Fig. 2A. We used a standard split of 80:20 for training and testing; the training data contains 11,196 instances and the testing split is 2,800, where 1,019 are positive and 1,781 are negative samples as shown in Fig. 2B. As the dataset was imbalanced, SMOTE and GAN oversampling was applied as discussed in the next section.

Dealing with class imbalance

The dataset developed in our study exhibits a notably greater quantity of instances labeled as Depressed compared to those labeled as Non-depressed, leading to an issue of class imbalance. Within the context of this research, a preference for oversampling over undersampling has been made to avoid loss of information and preserve the model’s performance. Data augmentation encompasses the application of SMOTE (Chawla et al., 2002) and GAN (Goodfellow et al., 2014) to create synthetic samples derived from the original dataset. Figure 3 shows the data augmentation using both SMOTE and GAN.

SMOTE generates new synthetic samples for the imbalance class by interpolation between existing samples and their nearest neighbors. While GANs are powerful machine learning models that use two neural networks competing with each other to generate synthetic data that looks real. GAN is better than SMOTE in generating realistic data, handling complex values and diversity, and being capable of generating unlabelled data. The original training data split contains 3,903 non-depressed (negative) cases and 7,293 depressed cases (positive). To balance our dataset, we used both the SMOTE and GAN models, respectively to generate 3,390 samples of non-depressed class. After applying SMOTE and GAN, our training dataset size increased to 14,586. Figure 4 shows that both negative and positive samples have the same number of samples (7,293) in the training data.

Hybrid RNN-LSTM-based model training

We proposed a hybrid model combining RNN and LSTM, for higher classification performance to predict PND in women. After performing multiple experiments as discussed in “Results and Discussion”, we found that RNN performed well at capturing short-term dependency in the data. Whereas, LSTM is good at capturing long-term dependencies in complex data by mitigating vanishing gradient problem through its gating mechanism (Sam et al., 2023). In our hybrid RNN-LSTM model, the inputs to the RNN model consist of PERI_DEP preprocessed data. The architecture of RNN includes following layers and components: ‘x’ denotes the input layer, ‘h’ represents the hidden layers, ‘y’ indicates the output layer, and ‘w’ signifies the weights that connect these layers. Each layer functions at a specific time-step ‘t’, reflecting the sequential nature of the data processing in the RNN. The output ‘y’ at any timestamp is computed based on these interactions and the temporal sequence of data, according to Eqs. (1) to (3).

(1) $$a^t=b_1+W{h}^{(t-1)}+U{x}^{(t)}$$

(2) $$h^t=A(a^{(t)})$$

(3) $$y^t=b_2+V{h}^{(t)}$$where $b_1$ and $b_2$ are bias vectors and ‘W’, ‘U’, and ‘V’ are the weights of layer connections, namely hidden–hidden, input–hidden, and hidden–output, respectively, and ‘a’ is the ReLU Activation function (Gautam et al., 2022). RNNs struggle with issues like vanishing gradients when they need to learn dependencies over long periods. Long short-term memory (LSTM) networks are specifically engineered to overcome these challenges. Each LSTM memory unit comprises a cell state and three distinct gates: the input gate, the forget gate, and the output gate. This structure is crucial for effectively managing information flow, allowing the network to retain or forget data as needed (Uddin et al., 2022).

The input gate $I_t$ can be as expressed in Eq. (4):

(4) $$I_t=\beta (W_{LI}{L}_t+W_{HL}{H}_{t-1}+b_1)$$where W is the weight matrix, b is bias vectors, and $\beta$ is the logistic function. The forget gate $F_t$, presented in Eq. (5), can be represented as:

(5) $$F_t=\beta (W_{LF}{L}_t+W_{HF}{H}_{t-1}+b_F).$$

The output gate $V_t$ generates the unit’s output and can be described as depicted in Eq. (6):

(6) $$V_t=\beta (W_{LV}{L}_t+W_{HV}{H}_{t-1}+b_V).$$

In this hybrid model, the input layer processes the PERI_DEP dataset, representing numerical records. The initial hidden layer of RNN is designed to capture short-term dependencies, while the subsequent LSTM hidden layer focuses on identifying long-term dependencies. Following these, the network incorporates hidden layers, each with 32 units and employing ReLU activation functions, to decipher complex patterns within the sequential data. The final layer, called the output layer, takes the output from the last hidden layer and maps it to a binary classification task: depressed or non-depressed. To achieve this classification, the output layer employs a sigmoid activation function. Figure 5 illustrates the architecture of the hybrid RNN-LSTM-based deep learning framework used in this study.

Baseline models

To evaluate our hybrid RNN-LSTM model on the PERI_DEP dataset, we compare it with several baseline models. As deep learning baselines, we use CNN-BiLSTM, Thekkekara, Yongchareon & Liesaputra (2024) and LSTM-CNN (Srivatsav & Nanthini, 2024) models. For machine learning baselines, we include well-established algorithms such as random forest (Huang et al., 2023), decision tree (Qasrawi et al., 2022), and gradient boosting tree (Qasrawi et al., 2022). The baselines provide a comprehensive comparison for evaluating the effectiveness of our model.

Experimental steps

Ablation study

We conduct an ablation study that evaluates the performance of various models and training strategies under three scenarios: without augmentation, with SMOTE, and with GAN augmentation as demonstrated in Table 3.

Without augmentation: The models are trained directly on the original dataset without any augmentation. Our Hybrid RNN-LSTM model achieves the highest performance across all metrics, showcasing the advantages of combining RNN and LSTM features with an F1-score of 95%. The simple LSTM also performs well, demonstrating its ability to effectively handle sequential data by achieving an F1 score of 94%. In contrast, the performance of RNN models indicates that they struggle with complex patterns. As our dataset is imbalanced, we apply two oversampling techniques for data augmentation.

Data augmentation with SMOTE: First we applied SMOTE to the training dataset. SMOTE generates new instances by interpolating between existing minority samples. After applying SMOTE the results enhance, RNN achieves an accuracy of 90%, while LSTM and our proposed Hybrid RNN-LSTM model archive an accuracy of 93% and 95% respectively. This demonstrates that LSTM maintains consistency while RNN helps mitigate class imbalance, enhancing overall generalization. Our model achieves the best results with SMOTE oversampling.

Data augmentation with GAN: Secondly we applied the GAN oversampling technique on the training dataset, GAN generates realistic synthetic. The hybrid RNN-LSTM consistently performs well, with an accuracy of 94%. This shows our Hybrid RNN-LSTM model performs better on complex time-series and sequential data.

Hybrid RNN-LSTM results

The hybrid RNN-LSTM model achieved an overall accuracy of 95% in predicting depression among women. The model performed well in predicting depression. Figure 6A shows the training and validation accuracy graph, and Fig. 6B depicts the training and validation loss graph. The hybrid RNN-LSTM model shows the best results with the highest accuracy of 95%, 95% precision, 97% recall, 96% F1-score, and 99% AUC.

We compare the performance of existing ML and DL models, including RF (Huang et al., 2023), DT (Qasrawi et al., 2022), gradient boosting tree (Qasrawi et al., 2022), CNN-BiLSTM, (Thekkekara, Yongchareon & Liesaputra, 2024) and LSTM-CNN (Srivatsav & Nanthini, 2024) models with our proposed hybrid RNN-LSTM model, utilizing our novel dataset. Although these established ML and DL models showed enhanced accuracy, their performance did not match the superior results achieved by our newly developed model. Table 4 summarizes the performance metrics for all models with the best results highlighted in bold. Figure 7 represents model results before and after data augmentation using SMOTE on the base of the F1-score. The results demonstrate that augmentation positively impacts the performance of all the models. In addition, the proposed model overshadows the performance of all competing models in terms of overall evaluation metrics.

According to these results, a large proportion of Pakistani women included in the study sample were facing mental health issues. Several factors were found to be associated with their psychological stress, including young maternal age, low husband family support, low monthly income, having multigravida, and child male gender preference. The findings of this study are concerning, and highlight the need for intervention programs to support women with perinatal psychological distress. This research could help to reduce adverse birth outcomes. Maternity clinics should screen women properly to identify those with symptoms of perinatal psychological distress, allowing for the provision of appropriate counseling and treatment.