Detecting Postpartum Depression Stages in New Mothers: A

Comparative Study of Novel LSTM-CNN vs. Random Forest

P. Srivatsav

and S. Nanthini

Department of Computer Science and Engineering, Saveetha School of Engineering,

Saveetha Institute of Medical and Technical Sciences, Saveetha University, Chennai, Tamil Nadu, 602105, India

Keywords: Women, Neural Networks, Depression, Prediction, Novel Long-Short Term Memory with Convolutional

Neural Networks, Random Forest Algorithm, Deep Learning, Machine Learning.

Abstract: A Novel long-short term memory with convolutional neural networks (LSTM-CNN) is used to predict

postpartum depression and compared it with Random Forest (RF) Algorithm. Materials and Methods: For this

research two groups were taken: The Novel Long-Short Term Memory with Convolutional Neural Networks

(LSTM-CNN) and for comparison the Random Forest (RF) Algorithm was considered. After careful

consideration each with a sample size of 20 to help in this research. Results: The outcomes of the study are

shown in the following table (LSTM-CNN). The mean accuracy of the LSTM -CNN is 77.75% and the

Random Forest (RF) Algorithm model is 72.12%, respectively. The significance of the Independent sample

t-test is evident with a p-value of 0.04 (p < 0.05), underscoring the statistical significance of the comparison

between the LSTM-CNN model and the Random Forest algorithm in the study. Conclusion: The LSTM-CNN

technique outperformed the Random Forest (RF) Algorithm and other machine learning algorithms in terms

of accuracy, and deep learning algorithms have generally showed promise in the prediction of Postpartum

depression.

1 INTRODUCTION

Artificial intelligence (AI) has the potential to help in

various ways when it comes to postpartum depression

(PPD). AI can help predict which women are most

likely to develop PPD and provide targeted

interventions to prevent or minimize the severity of

the condition. By analyzing data from various

sources, including demographics, medical history,

and social determinants of health, In order to give

women who are at high risk of PPD preventive

measures, healthcare professionals can identify them

using AI algorithms (Liu et al., 2023). There are

several apps available that can help new mothers

assess their risk of PPD and provide information on

symptoms to look out for. Some apps can even

provide a diagnosis based on self-reported symptoms,

although it's important to note that a professional

diagnosis is typically required for treatment (Wisner,

Parry, and Piontek 2002) Deep learning is extensively

used in computer vision applications such as object

recognition, image segmentation, face recognition,

and object detection and is used in NLP tasks such as

Research Scholar, Research Guide, Corresponding Author

language translation, sentiment analysis, speech

recognition, and chatbot development (Xu and

Sampson 2022)(Padma, S et al. 2022). Deep learning

is utilized in the healthcare industry for a variety of

activities including disease diagnosis, drug discovery,

and medical picture analysis and The recognition of

speech and images is made possible through machine

learning. Fraud detection systems can act fast to stop

fraud by using machine learning algorithms that can

be trained to recognize patterns of behavior that point

to fraud (G. Ramkumar et al 2022).

There are 447 publications on work done in

postpartum depression and machine learning and 11

IEEE Xplore publications that discuss postpartum

depression with the advancements of machine

learning. Research shows that the dataset validation

provides more accuracy for early detection than other

individual classifiers already in use. Nowadays data

mining and Machine learning plays a vital role in

detecting depression. From past research various

factors can impact the likelihood of developing PPD,

and analyzing tweets from diverse populations can

help to identify those who may be at higher risk. By

Srivatsav, P. and Nanthini, S.

Detecting Postpartum Depression Stages in New Mothers: A Comparative Study of Novel LSTM-CNN vs. Random Forest.

DOI: 10.5220/0012569700003739

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Artiﬁcial Intelligence for Internet of Things: Accelerating Innovation in Industry and Consumer Electronics (AI4IoT 2023), pages 109-115

ISBN: 978-989-758-661-3

109

reaching the conclusion that machine learning

algorithms possess the capacity to analyze extensive

datasets, effortlessly carry out intricate computations

on these data, and deliver predictive or insightful

outcomes, previous studies (Saqib, Khan, and Butt

2021) have demonstrated the capability of machine

learning algorithms in handling substantial data

volumes and generating meaningful predictions or

informative results.

Research has shown that various factors can

impact the likelihood of developing PPD, and

analyzing tweets from diverse populations can help to

identify those who may be at higher risk (Zhong et al.

2022) Due to the broad purpose nature of most

sentiment analysis techniques and their potential

inefficiency when used in contexts like healthcare or

finance, this research gap was discovered in the

current system. There is a need for tools that can

integrate domain-specific knowledge and language to

provide more accurate sentiment analysis results

(Yonkers et al. 2001). The primary goal of the

research is to create a sentiment analysis model that

would combine and evaluate sentiments from all

domains. The research team focussed on building a

sentiment analysis tool for the society that will

improve the accuracy and effectiveness especially in

more complex contexts (Thurgood, Avery, and

Williamson 2009).

2 MATERIALS AND METHODS

The research took place in the Saveetha School of

Engineering's Data Science lab at the Saveetha

Institute of Medical and Technical Sciences in

Chennai. After the datasets were gathered,

preprocessing and data cleaning methods were

carried out to remove any irrelevant or extraneous

information. The hardware to build this model CPU

was: processor Intel Core i3, 4 GB of RAM, and a 500

GB hard drive. The Sentiment140 dataset is a

collection of tweets that allows you to discover the

sentiment of a brand, product, or topic on Twitter

(Dutta and Deshmukh 2022). Originally, tweets in

this dataset were classified according to their

emoticons; for example, tweets with joyful emoticons

were classified as positive, while tweets with sad

emoticons were classified as negative The research

work was conducted on 2 groups of samples size 20

each (Iqbal, Chowdhury, and Ahsan 2018). This

model was run on Windows 10 and Jupyter notebook.

Table 1. shows a snapshot of the benchmark

Sentiment140 dataset that was used to analyze the

early markers of Postpartum Depression and how

deep learning algorithms have improved the

accuracy.

Novel Long-Short Term Memory with

Convolutional Neural Network Algorithm

When combining LSTM with CNNs, the input data is

first processed by the CNN layers to extract relevant

features from the data. The output of the CNN layers

is then fed into the LSTM layers, which process the

sequence of features and capture temporal patterns in

the data. To grow the classifier precision and decrease

the presence of noise and carry out a characteristic

choice (Dadi et al. 2020). In place of using complex

semantic evaluation, a concept is uniquely identified

via hashtags contained in the emotion tweet,

particularly, the metadata tags that are used in Twitter

to suggest the context or the flow of a tweet (Anokye

et al. 2018). Table 6. shows the Novel Long-Short

Term Memory with Convolutional Neural Networks

(LSTM- CNN)

Random Forest Algorithm

The Random Forest algorithm is widely employed in

machine learning for classification tasks, Algorithm,

and other tasks. It uses an ensemble learning

technique to merge various Random Forests to build

a more reliable and accurate model. Here's how the

Random Forest algorithm works:Data preparation:

First, training and testing sets are created from the

data. The testing set is used to assess the model's

performance once it has been built using the training

set. These Random Forests are constructed using a

process called "recursive partitioning," which

involves repeatedly dividing the data into smaller and

smaller subsets depending on the most crucial

qualities. Selecting the best split: At each split, the

algorithm selects the best feature and threshold value

to split the data. This is done by calculating the

information gain or Gini impurity of each feature and

selecting the one that provides the most useful

information for separating the data into the target

classes. For classification problems, the most

common method is to use a majority vote, where the

class with the most votes across all the trees is

selected as the final prediction. The average of all the

individual tree forecasts is frequently used as the final

prediction in algorithm problems. Considering the

model: Finally, the model's performance in terms of

accuracy and generalization is assessed using the

testing set. In doing so, metrics like accuracy,

precision, recall, and F1-score are computed by

comparing the projected values to the actual values.

Statistical Analysis

Data analysis, data management, and data

visualization were all accomplished using IBM's

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

110

Statistical Package for the Social Sciences (SPSS).

Accuracy values for both the independent and

dependent variables are enhanced by the former's

unique characteristics, which aid in making accurate

predictions of the former's n values.The G-power of

80%, and a maximum allowable error of 0.05. SPSS

has a number of features, including data analysis,

assumptions, and predictive models.

3 RESULTS

The dataset is displayed in Table 1 for different

backgrounds and locations. The simulated accuracy

analysis of the Random Forest (RF) algorithm and

Novel Long-Short Term Memory with Convolutional

Neural Networks (LSTM-CNN) is shown in Table 2.

For the Novel Long-Short Term Memory with

Convolutional Neural Networks (LSTM-CNN) and

the Random Forest technique, respectively, the mean

values are 77.75% and 72.12%, and the standard

deviations are 4.9 and 5.1, respectively.The

comparison between the LSTM-CNN model and the

Random Forest algorithm is statistically significant,

as indicated in Table 4 by the independent t-test

significance value of p=0.04(p0.05). Figure 1 depicts

the analysis of a bar graph based on the effectiveness

of two algorithms. The mean efficacy of the Random

Forest technique is 72.12% and 77.75% for Novel

Long-Short Term Memory with Convolutional

Neural Networks (LSTM-CNN), respectively.

According to the results, the Novel Long-Short

Term Memory with Convolutional Neural Networks

(LSTM-CNN) is a more effective method than the

Random Forest one. The suggested algorithm's mean

accuracy for the graphed representation is 72.12%

and 77.75%, respectively.A comparison of the

Random Forest method and the Novel Long-Short

Term Memory with Convolutional Neural Networks

(LSTM-CNN) algorithm is shown in Figure 1. Table

5 displays the novel LSTM's convolutional neural

network pseudocode.

Table 1: Snapshot of the Sentiment140 Dataset.

ItemID Sentiment SentimentSource SentimentText

1 0 Sentiment140 is so sad for my APL friend.............

2 0 Sentiment140 I missed the New Moon trailer...

3 1 Sentiment140 omg its already 7:30 :O

4 0 Sentiment140

I was suposed 2 just get a crown put

on (30mins)...

Table 2: Accuracy(%) of Novel Long-Short Term memory with Convolutional Neural Networks is compared with

Accuracy(%) of Random Forest Classifier .

S.NO Accuracy(%)of Novel LSTM-CNN

Accuracy(%)of Random Forest

Classifier

1. 82.90 79.87

2. 76.79 76.00

3. 80.99 73.09

4. 79.09 71.09

5. 73.89 65.69

6. 81.00 72.08

7. 70.78 62.80

8. 68.00 69.09

9. 78.09 70.80

10. 81.09 77.09

Accuracy 77.75 72.12

Detecting Postpartum Depression Stages in New Mothers: A Comparative Study of Novel LSTM-CNN vs. Random Forest

111

Table 3: Mean, standard deviation and error mean of Random Forest and Novel LSTM- CNN algorithm.

groups N Mean Std.Deviation Std.Error mean

Accuracy LSTM-CNN 10 77.7500 4.91859 1.55539

RF 10 72.1200 5.16027 1.63182

Table 4: Independent t- test has a significance value p=0.04(p<0.05) indicating the study between the LSTM-CNN model and

the Random Forest algorithm is statistically significant.

Independent T- Test

Levene's test for equality of variances T-TEST for equality of means

F Sig. t df

Significance

(2-tailed)

Mean

Difference

Std.Error

Differences

95% confidence of the

difference

lower upper

Accura

Equally

Variance

Assumed

.002 .969 2.441 18 .04 5.00200 2.25435 .76579 10.25821

Equal

Variance

not

Assume

2.441 17.959 .04 5.00200 2.25435 .76579 10.23899

Table 5: Pseudocode for Novel LSTM with Convolutional Neural Network (CNN).

a: Input data with shape (num_samples, max_seq_length)

: Target labels with shape (num_samples, num_classes)

1.Normalize and perform data augmentation on the input data

-Create two ImageDataGenerator objects to normalize the training and validation data

2. Define the convolutional neural layers:

-Define input shape for the model with shape (max_seq_length,)

-Define an input layer for the model with the defined input shape

3. Apply the embedding layer to the input layer

4. Define a convolutional layer with num_filters filters, kernel size of kernel_size, and ReLU activation

5.Apply the convolutional layer to the output of the embedding layer

6.Define a max pooling layer with pool_size pooling window

7.Apply the max pooling layer to the output of the convolutional layer

8.Define LSTM layer with lstm_units units and dropout rate (0.2)

9.Apply the LSTM layer (with 128 units) to the output of the max pooling layer

10. Define a fully connected output layer with num_classes units and softmax activation

11. Define the LSTM-CNN model with the input layer and output layer

12. Compile the model. Prediction measure is Accuracy.

13.Train the model for num_epochs epochs with batch size of batch_size:

-Randomly shuffle the data

-Split the data into batches of size batch_size

-For each batch, do the following:

-Compute the gradients of the loss with respect to the model parameters

-Update the model parameters using the optimizer

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

112

Table 6: Pseudocode for Random Forest.

Input: Sentiment140.csv

1.Start by selecting a random sample of the training data.

2. For each tree in the forest:

-Randomly select a subset of features to use as input for that tree.

- Using the selected features, build a Random Forest model on the subset of data.

3. Once all the trees have been built, predictions are made by taking the average prediction of all the trees in

the forest.

Figure 1: Bar graph analysis of the Random Forest algorithm and Long-Short Term Memory (LSTM-CNN) algorithm. A

graphical representation shows a mean accuracy of 77.75 % and 72.12% for the proposed algorithm, respectively. Random

forest vs. Long-Short Term Memory (LSTM), Y-axis: mean precision +/- 2 SD.

4 DISCUSSION

LSTM -CNN has helped achieve a better

accuracy(77.75%) when compared to Random forest

Algorithm which achieved a lower accuracy

(72.12%). The comparison of the Novel LSTM-CNN

with the Random forest Algorithm (Xin and Rashid

2021) shows that the Novel Long-Short Term

Memory with Convolutional Neural Networks

(LSTM-CNN) is better than the Random forest

Algorithm. The Novel Long-Short Term Memory

with Convolutional Neural Networks (LSTM-CNN)

has an accuracy of 77.75 % and Random forest

Algorithm is 72.12 % showing that the Novel Long-

Short Term Memory with Convolutional Neural

Networks (LSTM-CNN) is better than the Random

forest Algorithm. The two-tailed test has a

significance value p=0.04 (p<0.05) indicating the

study between the LSTM-CNN model and the

Random Forest algorithm is statistically significant.

As a result of the discusses and conclusions above, by

concluding that the Novel Long-Short Term Memory

with Convolutional Neural Networks (LSTM-CNN)

appears to perform and be more accurate than the

Random Forest Algorithm in all circumstances. The

Bar graph analysis seen in Fig. 1 shows the Novel

Long-Short Term Memory with Convolutional

Neural Networks (LSTM-CNN)

The positive effects of using a combination of

LSTM and CNN can make the model more robust to

noise in the input data. LSTM-CNN can be scaled up

to handle large amounts of data and complex tasks.

The negative effects of (LSTM-CNN) is a

computationally expensive model (Karmiani et al.

2019) due to the high number of parameters, which

can make it difficult to train and deploy on low-

resource devices. In order to prioritize input, the

LSTM-CNN employs a sigmoid neural net layer

(Ifriza and Sam'an, 2021). The gates safeguard and

regulate information flow, resolving the vanishing

gradient issue with typical RNNs (Tan and Lim

2019). The algorithm utilized, which is based on

stochastic gradient descent (Nikmah et al., 2022), aids

in enforcing consistent error propagation (neither

bursting nor disappearing) through its internal units.

Detecting Postpartum Depression Stages in New Mothers: A Comparative Study of Novel LSTM-CNN vs. Random Forest

113

When working with large datasets, the use of LSTM

models might be computationally expensive (Nikmah

et al. 2022).

The LSTM-CNN is a complex architecture that

requires a large amount of computational resources

and training data. This makes it difficult to implement

and train on smaller datasets or less powerful

hardware. LSTM-CNN is built to handle long

sequences, however because of the vanishing gradient

problem, it can still have trouble with extremely long

sequences. This happens when the weights' gradients

during backpropagation shrink too much, making it

challenging to update the weights and gain insight

from the data. LSTM-CNN requires large amounts of

labeled training data to achieve good performance.

This can be a limitation in applications where labeled

data is scarce or difficult/expensive to obtain. The

future scope of the Novel LSTM-CNN is vast and

varied, with potential applications in diverse fields

and can be used to analyze speech patterns and detect

changes in tone (Thurgood, Avery, and Williamson

2009; Stewart and Vigod 2016), pitch, and other

vocal characteristics that may indicate depression.

The Novel LSTM-CNN can also be used to analyze

facial expressions and detect changes in emotion that

may indicate depression. The Novel LSTM-CNN can

be used to analyze EEG signals and detect

abnormalities that may indicate depression. As a

future work other datasets like the Multimodal

Dataset for Mental Health Analysis

(MMDMA),(Thurgood, Avery, and Williamson

2009; Stewart and Vigod 2016) that incorporate a

wider range of modalities such as physiological

signals, audio, and video recordings in addition to text

can be used for depression prediction.

5 CONCLUSION

In this research work, it was possible to develop a

reliable and accurate ML model that can accurately

predict Postpartum depression. The Long Novel

Long-Short Term Memory with Convolutional

Neural Networks (77.75%) is more accurate when

compared with Random Forest Algorithm (72.12%)

for predicting depression of new mothers through

social media.

REFERENCES

Anokye, Reindolf, Enoch Acheampong, Amy Budu-

Ainooson, Edmund Isaac Obeng, and Adjei Gyimah

Akwasi. (2018). “Prevalence of Postpartum Depression

and Interventions Utilized for Its Management.” Annals

of General Psychiatry 17 (May): 18.

Byvatov, Evgeny, Uli Fechner, Jens Sadowski, and Gisbert

Schneider. (2003). “Comparison of Support Vector

Machine and Artificial Neural Network Systems for

Drug/nondrug Classification.” Journal of Chemical

Information and Computer Sciences 43 (6): 1882–89.

Cellini, Paolo, Alessandro Pigoni, Giuseppe Delvecchio,

Chiara Moltrasio, and Paolo Brambilla. (2022).

“Machine Learning in the Prediction of Postpartum

Depression: A Review.” Journal of Affective Disorders

309 (July): 350–57.

Chai, Junyi, Hao Zeng, Anming Li, and Eric W. T. Ngai.

(2021). “Deep Learning in Computer Vision: A Critical

Review of Emerging Techniques and Application

Scenarios.” Machine Learning with Applications 6

(December): 100134.

Dadi, Abel Fekadu, Temesgen Yihunie Akalu, Adhanom

Gebreegziabher Baraki, and Haileab Fekadu Wolde.

(2020). “Epidemiology of Postnatal Depression and Its

Associated Factors in Africa: A Systematic Review and

Meta-Analysis.” PloSOne 15 (4): e0231940.

Dutta, Sushmita, and Prasad Deshmukh. (2022).

“Association of Eating Disorders in Prenatal and

Perinatal Women and Its Complications in Their

Offspring.” Cureus 14 (11): e31429.

G. Ramkumar, G. Anitha, P. Nirmala, S. Ramesh and M.

Tamilselvi, "An Effective Copyright Management

Principle using Intelligent Wavelet Transformation

based Water marking Scheme," 2022 International

Conference on Advances in Computing,

Communication and Applied Informatics (ACCAI),

Chennai, India, 2022, pp. 1-7, doi:

10.1109/ACCAI53970.2022.9752516.

Ifriza, Yahya Nur, and Muhammad Sam’an. (2021).

“Performance Comparison of Support Vector Machine

and Gaussian Naive Bayes Classiﬁer for Youtube Spam

Comment Detection.” Journal of Soft Computing

Exploration 2 (2): 93–98.

Iqbal, Nazma, Afifa Mim Chowdhury, and Tanveer Ahsan.

(2018). “Enhancing the Performance of Sentiment

Analysis by Using Different Feature Combinations.” In

2018 International Conference on Computer,

Communication, Chemical, Material and Electronic

Engineering (IC4ME2), 1–4.

Karmiani, Divit, Ruman Kazi, Ameya Nambisan, Aastha

Shah, and Vijaya Kamble. (2019). “Comparison of

Predictive Algorithms: Backpropagation, SVM, LSTM

and Kalman Filter for Stock Market.” In 2019 Amity

International Conference on Artificial Intelligence

(AICAI), 228–34.

Liu, Hao, Anran Dai, Zhou Zhou, Xiaowen Xu, Kai Gao,

Qiuwen Li, Shouyu Xu, et al. (2023). “An Optimization

for Postpartum Depression Risk Assessment and

Preventive Intervention Strategy Based Machine

Learning Approaches.” Journal of Affective Disorders

328 (February): 163–74.

Nikmah, Tiara Lailatul, Muhammad Zhafran Ammar,

Yusuf Ridwan Allatif, RizkiMahjatiPrie Husna, Putu

Ayu Kurniasari, and Andi Syamsul Bahri. 2022.

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

114

“Comparison of LSTM, SVM, and Naive Bayes for

Classifying Sexual Harassment Tweets.” Journal of

Soft Computing Exploration 3 (2): 131–37.

O’Hara, Michael W. (1997). “The Nature of Postpartum

Depressive Disorders.” In Postpartum Depression and

Child Development , (pp, edited by Lynne Murray,

322:3–31. New York, NY, US: Guilford Press, xiv.

Padma, S., Vidhya Lakshmi, S., Prakash, R., Srividhya, S.,

Sivakumar, A. A., Divyah, N., ... & Saavedra Flores, E.

I. (2022). Simulation of land use/land cover dynamics

using Google Earth data and QGIS: a case study on

outer ring road, Southern India. Sustainability, 14(24),

16373.

Saqib, Kiran, Amber Fozia Khan, and Zahid Ahmad Butt.

2021. “Machine Learning Methods for Predicting

Postpartum Depression: Scoping Review.” JMIR

Mental Health 8 (11): e29838.

Stewart, Donna E., and Simone Vigod. (2016). “Postpartum

Depression.” The New England Journal of Medicine

375 (22): 2177–86.

Tan, Hong Hui, and King Hann Lim. (2019). “Vanishing

Gradient Mitigation with Deep Learning Neural

Network Optimization.” In 2019 7th International

Conference on Smart Computing & Communications

(ICSCC), 1–4.

Thurgood, Sara, B. S. Daniel M. Avery, and M. D. Lloyda

Williamson. (2009.) “Postpartum Depression (PPD).”

aapsus.org. 2009.

http://www.aapsus.org/articles/11.pdf.

Wisner, Katherine L., Barbara L. Parry, and Catherine M.

Piontek. 2002. “Postpartum Depression.” The New

England Journal of Medicine 347 (3): 194–99.

Xin, Lee Ker, and Nuraini Binti Abdul Rashid. (2021).

“Prediction of Depression among Women Using

Random Oversampling and Random Forest.” In 2021

International Conference of Women in Data Science at

Taif University (WiDSTaif ), 1–5.

Xu, Wen, and Mcclain Sampson. (2022). “Prenatal and

Childbirth Risk Factors of Postpartum Pain and

Depression: A Machine Learning Approach.” Maternal

and Child Health Journal, December.

https://doi.org/10.1007/s10995-022-03532-0.

Yeboa, Naomi Kyeremaa, Patience Muwanguzi, Connie

Olwit, Charles Peter Osingada, and Tom Denis

Ngabirano. (2023). “Prevalence and Associated Factor

of Postpartum Depression among Mothers Living with

HIV at an Urban Postnatal Clinic in Uganda.” Women’s

Health 19: 17455057231158471.

Yonkers, K. A., S. M. Ramin, A. J. Rush, C. A. Navarrete,

T. Carmody, D. March, S. F. Heartwell, and K. J.

Leveno. (2001). “Onset and Persistence of Postpartum

Depression in an Inner-City Maternal Health Clinic

System.” The American Journal of Psychiatry 158 (11):

1856–63.

Zhong, Minhui, Han Zhang, Chan Yu, Jinxia Jiang, and Xia

Duan. 2022. “Application of Machine Learning in

Predicting the Risk of Postpartum Depression: A

Systematic Review.” Journal of Affective Disorders

318 (December): 364–79.

Detecting Postpartum Depression Stages in New Mothers: A Comparative Study of Novel LSTM-CNN vs. Random Forest

115