2.1. Concept Related to Social Network Usage and Depression

Reference ¹⁷ found that individuals experiencing depression used social media to learn from others’ experiences and share personal stories. Similarly, reference ¹⁸ indicated that social media use among adolescents aged up to 25 years provides benefits; although it does not directly reduce psychological symptoms, users perceive social media as a safe space where they can seek information and find encouragement through online media. Conversely, this study shows that social media platforms have become spaces where researchers can infer what individuals are facing and identify barriers that prevent users from seeking mental health treatment. These insights were gathered by analyzing posts, shares, and comment interactions. However, excessive and continuous use of social media has been shown to negatively impact mental health and increase the risk of depression ^19,20. According to ²¹, high-volume social media usage, particularly Facebook, correlates with depressive symptoms. Therefore, the development of tools or strategies to mitigate these adverse effects is necessary. This study applied AI to predict the depression status of social media users, thereby enabling the prevention of potential negative impacts on Thai users. This aligns with the proposal of ²⁰, which emphasizes the need for cautious and mindful use of social media. The focus should be on detecting depression through social media, particularly utilizing visual and spatial data. Collaboration between mental health experts and computer scientists is essential for improving the methods of identifying depression on social media platforms, thereby transforming social media into a safe space for discussing sensitive topics such as mental health disorders ²². To develop this AI model, the researcher utilized words, messages, and images to create tools for predicting depression. This serves as a social monitoring system that enhances alert notifications via the X accounts of users who post high-risk content, aiming to prevent individuals from unknowingly experiencing depression. This approach integrates communication science, innovative technology, and mental health to provide a comprehensive solution.

2.2. Concept of Natural Language Processing

NLP is a branch of AI that enables computers to understand, interpret, and process human languages used in daily communication ²³. This involves segmenting the text or sentences and training the computer to comprehend their meanings. Subsequently, NLP analyzes the relationships within the data and processes this information to convey complex semantic content. Reference ²⁴ developed a depression detection system based on the Weibo platform using web scraping techniques. This study compared several models, including support vector machine (SVM), naïve Bayes (NB), convolutional neural network (CNN), and XLNet. The experimental results indicated that XLNet, which was based on the transformer architecture, produced the most accurate outcomes. However, this model has limitations related to device memory capacity and training time. Achieving higher accuracy required longer training periods, and the dataset did not distinguish between severe and mild depressive states. Therefore, this study aims to address this gap by incorporating psychiatric expert assessments to classify the severity level of messages prior to training the model, thereby enhancing its accuracy. References ^25,26 stated that transformer models are commonly trained on parallel corpora, which presents a limitation in requiring increased memory capacity with the length of the data ²⁷. When collecting posts from individuals at risk for depression, lengthy data such as expressions of emotional distress regarding their symptoms or sharing feelings about their experiences may be encountered. Therefore, models that offer high efficiency with relatively short processing times were selected. According to ²⁸, Transformer models require large volumes of data for training, posing challenges for scenarios with limited data. In this study, where data were collected from social media over the past three months, there was a higher risk of overfitting owing to the limited dataset size.

Reference ²⁹ developed methods for detecting depression in Thai using three neural network algorithms: 1) BiLSTM, 2) Thai-BERT, and 3) WangchanBERTa. The results showed that WangchanBERTa, which utilizes the Newmm word segmentation technique, achieved the best performance, with an F1-score of 74.2%, recall of 74.2%, precision of 74.7%, and accuracy of 74.2%. Although this study tested recent language learning models, their performance scores may have been insufficient for real-world applications. Additionally, the limitations of transformer architectures, such as high resource consumption and long training times, have motivated researchers to explore alternative models that are more resource-efficient and faster. Reference ³⁰ stated that the Bayes classifier demonstrated the highest accuracy in distinguishing depression from text on Twitter, because it yielded the most reliable results when tested on real-world datasets. Therefore, NB was chosen as the approach for developing the model. Similarly, reference ²⁶ noted that although Transformers are newer models, their complex decision-making processes make it difficult to explain, interpret, and predict, which is an important consideration for practical applications such as medical diagnostics. In this study, because collaboration with psychiatrists is essential, the prediction results need to be presented in an understandable manner that does not place an additional burden or take extra time for psychiatrists to interpret. Therefore, the researcher anticipates that NB is more suitable, given the available data volume, resources, and processing time allocated for the study, and the researcher’s capacity. Conversely, reference ³¹ employed four models: extreme gradient boosting (XGB) classifier, random forest, logistic regression, and SVM. They found that the support vector machine (SVM) and logistic regression produced more accurate depression predictions on Twitter when applied to existing datasets. Among these, logistic regression demonstrated higher accuracy than SVM and required less processing time, making it suitable for deployment in real-world scenarios. The results of this study indicate that different algorithms may exhibit varying levels of performance depending on the specific context. Therefore, it is essential to conduct testing within the research context to identify the model that offers the best performance, thereby ensuring optimal outcomes for subsequent R&D.

2.3. Concept of Sentiment Analysis

Sentiment analysis involves the examination of emotions expressed in textual data and is utilized to understand public opinions about daily life and various topics ³¹. Because emotions influence decision-making, detecting emotional states is a scientific approach that provides humans with an advantage by understanding how to approach or communicate with others, especially among groups at risk of depression ³². Therefore, detecting emotional sensitivity related to depression through words, messages, or images posted by users on social media can enhance the understanding between humans and computers, leading to greater empathy and improved detection of at-risk users.

Sentiment analysis categorized depression-related messages into positive (positive words), neutral (neutral words), and negative (negative words) ³³. Python was used to process these classifications, which often involve multi-dimensional sentiments that simultaneously encompass both positive and negative aspects. Reference³⁴ explained that the process involves segmenting text into individual words (using the PyThaiNLP library) and then analyzing the relationships between words and existing documents in vector form. The researcher developed the model using the NB algorithm (employing NLTK), which processes the input text by segmenting it into words and then applying the trained model to output a classification, such as the sentiment of the text. Example: “สู้ ๆ นะคะ คนเก่งทุก ๆ คนเลย #โรคซึมเศร้า” is a positive message, while “ต้องให้แตกสลายอีกกี่ครั้งหรอเราถึงจะได้จากไปโดยไม่รู้สึกผิดสักที #โรคซึมเศร้า #ครอบครัวไม่ใช่เซฟโซน” is a negative message, and so on.

This approach facilitates the analysis of textual messages and images to determine whether the expressed emotion is positive, negative, or neutral among users at risk of depression. Some studies have combined sentiment analysis and emotion detection within a single model to extract in-depth insights from opinionated content ³². For example, reference ⁸ used AI to predict depression on Twitter by detecting hashtags such as #depress, #hopeless, #nervous, #worthless, #tired, etc., posted daily. However, this method primarily compares levels of depression across days and cannot accurately identify individuals who are truly at risk for depression. Similarly, reference ⁷ applied ML to predict depression through voice analysis, showing that AI can make accurate predictions; however, it requires large amounts of voice data, must be supervised by medical professionals, and does not provide alerts to at-risk individuals to encourage self-awareness or prompt treatment. Additionally, reference ³⁵ developed a system for detecting potentially depressive sentences on X by comparing the accuracies of two algorithms, which achieved a prediction accuracy of over 94%. This recommendation suggests that depression should be studied at the individual level to improve accuracy in predicting depression status. Therefore, this study aims to address this gap by developing a system capable of sending alerts to each user’s X account when it detects that the user has posted words, messages, or images indicating a high risk.

2.4. Knowledge of Artificial Intelligence Systems

A model refers to a set of mathematical or probabilistic relationships between different variables and is commonly known as machine learning ^36,37. This enables computers to learn patterns and discover relationships within data to facilitate predictions or decision-making. Data are typically divided into training and test sets according to specified proportions, such as \(60 {:} 40\) or \(70 {:} 30\). The average accuracy tends to increase as the size of the training set increases ³⁸, as cited in ³⁹. Raw data for text classification using various neural network algorithms, such as NB, logistic regression, and SVM, are often complex and unstructured. Therefore, preprocessing steps are necessary, including data cleaning, tokenization into groups, removing meaningless words (stop words), reducing words to their root forms, and converting words into vector representations of numbers (text vectorization) to enable machines to understand textual meanings. After preprocessing, the performance of the model is evaluated by measuring prediction accuracy. If the results are unsatisfactory, then parameter tuning or alternative algorithms can be employed ²⁹. For example, in ⁴⁰ depression was predicted using user posts. Data were transformed into vectors using a CountVectorizer and TF-IDF Vectorizer, then trained with ML techniques such as NB, SVM, decision trees, random forest, and \(K\)-nearest neighbors (KNN). The words were processed using a dictionary to determine positive or negative sentiments. The data quality was further enhanced by removing special characters and unnecessary symbols using regular expressions, and preprocessing techniques such as stop-word removal, lemmatization, and tokenization were applied. This aligns with the research procedure use in the present study. However, if the training data for the model are of poor quality or insufficient quantity, it may lead to issues related to data imbalance. To address this issue, reference ⁴¹ proposed ERNIE 3.0 Hybrid Prompt-Suicide, which utilizes ERNIE 3.0—a language model capable of enhanced knowledge integration—designed to detect suicidal intent through a prompt-based classification suitable for resource-limited settings (mitigating overfitting issues with small datasets). The approach follows the “pre-train, prompt, predict” methodology, developing hybrid template prompts that transform original data into formats with unfilled slots. This method achieves an accuracy of 87.6%. The model exemplifies the application of advanced AI and DL techniques to improve performance in low-resource scenarios. However, this approach requires the integration of electronic health records (EHR) data. Because of the researcher did not utilize health data or involve patients due to ethical restrictions on human research in Thailand, this study proposes methods for detecting depression through users’ social media activity as an alternative.

For the Thai language, reference ⁴² developed a Thai depression dataset and compared it with four DL models. They used PyThaiNLP for sentence segmentation, divided the data into training, validation, and test sets, and applied ThaiVec for word embedding. The results showed that Thai BERT achieved the highest performance, with an overall accuracy of 77.53%.

Additionally, reference ⁴³ predicted depression signals in Thai tweets by collecting data via the TweePy library from hashtags such as #depression, #depressivedisorder, and #suicide, totaling 3,100 messages. Data labeling was performed by mental health experts, followed by data cleaning using PyThaiNLP. Oversampling was applied using the SMOTE technique. The study found that the LSTM \(+\) ReLU model offered good performance and accuracy and was then implemented in a prototype web application for model testing. However, these studies collected fewer data points and used different models. Therefore, this approach provides a basis for conducting research in the Thai language, where the optimal predictive model for identifying messages at risk for depression needs to be tested. The results help determine the extent to which the findings differ from previous research outcomes.

From the previous discussion, it is clear that before deploying a model in real-world applications, its performance must be evaluated to identify the most effective model for predicting depression. Common evaluation metrics include accuracy, precision, recall, and F1-score ³⁶. Additionally, cross-validation can be performed using test data to assess the model’s performance under conditions that closely resemble actual usage ²⁹. Generally, classification models are evaluated using a confusion matrix, which is a \(2\times 2\) table displaying the ability of the model to distinguish between classes, as shown in Table 1.

In Table 1, TP refers to the model predicting “depressed” when the actual outcome is “depressed.” TN indicates the model predicting “not depressed” when the actual outcome is “not depressed.” FP occurs when the model predicts “depressed” but the actual outcome is “not depressed.” FN means the model predicts “not depressed” but the actual outcome is “depressed.” These values were calculated using the formulas listed in Table 2.

Typically, models that predict “Yes” with only slight confidence tend to have high recall but low precision, while models that predict “Yes” only when they are highly confident tend to have low recall but high precision. Frequent positive predictions can result in numerous false positive, whereas frequent negative predictions can result in numerous false negative. Therefore, increasing the threshold in the model enhances the testing accuracy but can also reduce recall. In such cases, selecting an appropriate threshold involves determining a balanced trade-off between precision and recall ³⁶.

2.5. Bayesian Theory

Bayesian theory involves mathematical equations that utilize probability and statistics for computation. It is used in conjunction with neural networks to estimate parameters based on the likelihoods of different parameter occurrences. The fundamental principle of NB is to count (frequency) occurrences of items, calculate their probabilities, and compare them. The data used for training, including features (independent variables: \(x\)) and targets (dependent variable: \(y\)), should be suitable for classification or categorization into groups ³⁷. According to ⁴⁴, Bayesian models represent graphs that effectively describe the joint probability distributions of multiple random variables. They can capture conditional, dependencies clearly, thereby providing advantages over general classification models. These models can incorporate expert opinions into their graphical structures, generate interpretable predictions, provide uncertainty alerts, and handle missing data naturally. This has led the researchers to consider Bayesian methods as a feasible approach for predicting depression Bayesian methods have been found to effectively capture complex relationships between symptoms and various factors associated with depression. They demonstrated high performance in addressing misdiagnosis issues based on the DSM-V and ICD-10 criteria for predicting depression in Nigeria ⁴⁵. Similarly, the risk prediction model for depression (PDRM) developed in ⁴⁶, which compared prediction results with outpatient data diagnosed with depression, showed that the recall values were 0.92 for NB, 0.86 for logistic regression, and 0.84 for SVM, indicating a strong predictive capability. The aforementioned study indicate that Bayesian methods can be effectively integrated into existing clinical practices, rendering them viable AI models for further development. Therefore, this approach was adopted as a guiding framework for the development process. Reference ⁴⁷ analyzed opinions related to depression on Instagram using NB and developed an application that visualized analysis reports in a graphical format. They employed a lexicon-based technique to label data. The results indicated that the NB performed effectively, achieving an accuracy rate of 82.55%. The study provides a foundation for presenting prediction results through a web interface, making it easier for experts to interpret and understand the data.

In the context of Thai research, reference ⁴⁸ developed a sentiment analysis model for monkeypox misinformation on X, utilizing BERT for classification. Although capable of analyzing sentiments, this model demonstrated lower accuracy, precision, and recall than models employing NB. Meanwhile, reference ⁴⁹ applied NB for text classification to analyze tourists’ emotions toward visited locations and found that NB delivered good performance with a simple working procedure relative to other methods. Additionally, reference ⁵⁰ used NB to classify rubber tree species with unknown varieties and achieved an accuracy of 98.02%, indicating a high level of precision. However, these studies rarely report the use of NB to predict depression, highlighting a gap in the current research landscape. These studies demonstrated that NB possesses notable potential, exhibiting good performance and high efficiency. This makes it suitable for testing and practical application in real-world scenarios. According to ³⁷, NB offers several advantages, including a simple learning process and quick comprehension, low computational resource requirements, rapid output generation, flexibility in adding or removing data samples, support for both continuous (e.g., \(x_1, x_2,\ldots\)) and discrete data, immunity to overfitting, easy management of missing data, and the ability to handle large volumes of data efficiently. However, a notable disadvantage is that if the predicted value does not appear in the training data, the frequency will be zero, which can lead to prediction errors.

Based on a review of the above research concepts, the researcher designed a flowchart illustrating the research process to facilitate the study, as shown in Fig. 2.

3.1. Population and Sample Group

The researcher divided the study into two groups:

1. 1)

   Social media utilized by individuals at risk of depression to express their feelings through words, messages, or images, totaling 15,043 posts on three platforms:

   1. 1.1)

      Facebook:

      • Facebook Groups: The top five Thai depression-related groups with the largest membership are “If you have depression, we are friends,” “Depression,” “Depression, panic, bipolar, we will fight together” (Thailand Mental Health Support), “Recommendations for depression, bipolar disorder, schizophrenia, other psychiatric conditions,” “Depression relief without medication” ⁵¹ with 2,150 posts.

      • Facebook Pages: The top five Thai pages with the most followers or likes are “Mental health consultation page,” “Messages from depression patients,” “Depression, dear friend,” “Absorbing without getting depressed,” “When I got sick with depression” ⁵², totaling 574 posts. The combined number of posts from these sources was 2,724.

   2. 1.2)

      X selected tweets based on the most recent (latest) hashtags and those with the highest number of likes or retweets (top) each day. Specifically, negative tweets were collected from #โรคซึมเศร้า (#depression) and #ซึมเศร้า (#depressed), totaling 9,497 tweets, while positive tweets were gathered from #คนเก่ง (#talented), #ความสุข (#happiness), #ซึมเศร้าเพื่อนรัก (#depression my dear friend), #มีความสุข (#joyful), #สนุกสนาน (#fun), #สู้ต่อไป (#keep going), #สู้นะ (#cheer up), #ให้กำลังใจ (#support), and #ฮีลใจ (#healing) amounting to 1,594 tweets. The total number of tweets collected was 11,091.

   3. 1.3)

      Instagram selected posts based on images uploaded from various accounts with the hashtags #โรคซึมเศร้า (#depression) and #ซึมเศร้า (#depressed) ⁵³, totaling 1,228 posts.

2. 2)

   Experts practicing in government or private sector organizations in Thailand and possessing expertise in the fields of psychology, information technology/innovation, and communication sciences, evaluated the accuracy of the predictions. These professionals or academics have at least five years of practical experience and hold academic positions at the assistant professor level or higher, with doctoral degrees or equivalent academic qualifications. Purposive sampling was used to select 30 individuals based on the purpose of the study. This sampling was conducted through simple random selection focused on purposeful criteria ^54,55.

3.2. Research Procedure

This study aims to develop an AI model to predict depression based on social media usage. The implementation proceeded according to the following steps:

1. 1)

   Data were collected retrospectively over the past three months from the specified hashtags. Screenshots were used to extract words, messages, and images for the analysis. Keywords or phrases commonly posted by at-risk groups were identified, and messages were categorized accordingly. The number of posts in each category was then counted to support further analyses.

2. 2)

   The collected data were reviewed by three psychiatrists who independently assessed and categorized the messages, including rating the severity levels of the words. To reduce the risk of bias and enhance the reliability of the findings, a triangulation approach was employed, specifically interdisciplinary triangulation, in which the multiple expert opinions were compared and synthesized to achieve a consensus, ensuring more robust validity of the categorization and severity assessments.

3. 3)

   Sentiment analysis involves categorizing words, messages, and images into three emotional groups: positive, neutral, and negative ³³.

4. 4)

   The steps for developing the AI system are as follows:

   1. 4.1)

      Preparing the data for analysis: Data collection involved retrieving both positive and negative tweets, which were then organized into a DataFrame format using the Pandas library. The NLTK library in Python was used to analyze the data. The accuracy of the data was verified and validated by three experts to ensure the reliable identification of words or messages indicative of depression, thereby enhancing the effectiveness of the analysis.

   2. 4.2)

      Cleaning the data: The process begins with word segmentation and analyzes the relationships between words and a lexicon stored in vector form. Non-informative data, such as URLs, special symbols, numbers, and punctuation, were removed. Slang and its abbreviations were expanded into their full forms, and emojis and smiley faces were replaced with their corresponding true emotional expressions instead of being discarded. The images were classified into positive, negative, or neutral categories. These steps were performed by the researcher and validated for accuracy by a programmer.

   3. 4.3)

      Normalization: Words were transformed into canonical forms to standardize and group synonymous terms by reducing them to their root or base forms. This process employs techniques such as stemming and lemmatization. The steps were performed by the researcher, and the accuracy was subsequently verified by a programmer.

   4. 4.4)

      Building and evaluating model: Text data were transformed into feature vectors using CountVectorizer from the scikit-learn library, which converts text into a frequency-based vector representing how often each term appears. Important parameters such as “min_df” are applied to filter out terms that occur less than the specified threshold, while “max_features” is used to select only the most frequent words. Subsequently, a Gaussian NB model was employed to classify the texts based on these vectors.

   5. 4.5)

      Performance evaluation: The hold-out validation method was employed by randomly splitting the entire dataset into two parts: 70% for the training set and 30% for the testing set. This was implemented using the “train_test_split()” function from the scikit-learn library, with the “random_state” parameter set to ensure reproducibility of results. This approach increases the reliability of the model performance evaluation and reduces bias from random data splitting. Additionally, we applied \(k\)-fold cross-validation with \(k\) set to 5, which is a widely accepted standard in ML research. This process involvs dividing the complete dataset into five folds. The model was trained over five iterations, each time using four folds for training and one fold for testing, rotating through all folds until each fold served as a test set. In this study, TfidfVectorizer was combined with multinomial NB within a pipeline, and cross-validation was performed using the “cross_val_score()” function from scikit-learn. The results showed an average accuracy of 75.42%, with similar accuracy across all folds, indicating that the model was stable (robust) and reproducible.

   6. 4.6)

      Deploy and diagnose depression: The trained model is deployed to predict real-world data on X, as this platform does not impose restrictions related to data privacy. If continuous negative tweets are detected over a period of more than two weeks, indicating a potential risk of depression ⁵⁶, the system automatically sends a notification to the user advising them to consult a mental health professional.

5. 5)

   The accuracy of the system is evaluated by comparing its performance metrics with assessments provided by experts. The research process is illustrated in Fig. 2.

3.3. Data Collection

1. 1)

   Documents, theories, research studies, and various media were reviewed to establish the conceptual framework of this study.

2. 2)

   Data were collected through observation of digital ethnography across the three platforms, recorded, and organized in Microsoft Excel. Non-participant observation was employed to avoid disturbing the environment of the sample group and prevent sensitive mental health issues among individuals at risk of depression. This approach minimizes potential bias related to demographic factors such as age, sex, or background. In total, 15,043 posts were collected.

3. 3)

   The categorization and severity of the words were classified into three levels. To reduce expert bias, the severity assessment is based on the criteria established by ⁵⁷, which categorizes depression risk among suicidal patients into three levels: low, moderate, and high (based on depression assessment scores (\(\mathrm{9Q} \ge 7\)) and suicide risk evaluation (\(\mathrm{8Q} \ge 1\))). This classification aims to align psychiatrists’ evaluations (3 levels) with the three groups used in sentiment analysis inputs: positive (1), neutral (0), and negative (\(-1\)) ³³. Adopting standardized criteria helps minimize errors related to probabilistic predictions, thereby enhancing the accuracy and reliability of the model’s predictions, presenting data in tabular format.

4. 4)

   The validity of the AI evaluation form was assessed using the index of item objective congruence (IOC) by consulting three experts in fields related to information technology, innovation, and communication arts prior to data collection. The IOC values ranged from 0.66 to 1.00, which meets the content validity criteria ⁵⁸. The reliability of the assessment tool, it was tested by at least one psychologist. The expert conducted evaluation trials, and if the responses were consistent across all items, with answers aligned in the same direction and without challenges, the form was deemed reliable for use in the assessment.

3.4. Data Analysis

The researcher compiled a list of keywords and phrases frequently posted by individuals at risk of depression. These data were analyzed and synthesized through classification and organization using typology and taxonomy, along with inductive analytical methods. The analysis followed a predetermined observational framework to draw conclusions. The information provided to psychiatrists for expert assessment used a rating scale system—an interval scale widely employed in social science research—designed to develop questionnaires, interview formats, and attitude or characteristic measurement tools ⁵⁹. The severity was rated as follows:

1. Low or no severity

2. Moderate severity

3. High severity

Score interpretation was based on threshold criteria by calculating the mean across all assessment sets, with scores ranging from 1 to 3. The mean scores were categorized into three levels using the criteria established in ⁶⁰ and cited in ⁶¹:

• Scores between 2.36 and 3.00 indicate high severity

• Scores between 1.68 and 2.35 indicate moderate severity

• Scores between 1.00 and 1.67 indicate low or no severity

Data from all 30 assessment sets were coded and processed using statistical analysis software. Descriptive statistics, including means and standard deviations, were calculated to summarize the data. Inferential statistics were employed to compare the differences between more than two variables using one-way ANOVA, while pairwise comparisons were conducted using the least significant difference (LSD) test. An analysis of the intraclass correlation coefficient (ICC) among expert raters yielded a value of 0.973, indicating excellent inter-rater reliability and high consistency among evaluators ⁶².

3.5. Data Collection

The system developers collected training data from Facebook, X, and Instagram, totaling 15,043 posts, using digital ethnographic observation. The data were filtered from Thai users who tagged #depression #depressive disorder in the Thai language through the TweePy tool and retrieved via user IDs. The obtained data consisted of user IDs, tweet content, and posting timestamps, categorized into negative, neutral, and positive words, referencing the Thai sentiment lexicon database from ³⁴, which is an open-source resource for system developers. This list is presented in Table 3.

4.1. Results of Developing the Predictive Model for Words, Messages, or Images that Indicate Risk of Depression, Created Using Artificial Intelligence

Step 1: Preparing the Data for Analysis

The researcher compiled positive and negative texts into a single DataFrame using the Pandas library, consisting of 5,400 rows. The dataset was divided into 2,400 rows of positive tweets and 3,000 rows of negative tweets. The confidence values were encoded binarily: 0 indicated positive confidence and 1 indicated negative confidence, as illustrated in Fig. 3.

The researcher retrieved data from X using user ID, ensuring that the collected data could not be linked to individual identities. Therefore, it was not necessary to obtain permission from the original tweet owners for data use to comply with the Personal Data Protection Act (PDPA) regulation. Tweets were selected without specific temporal restrictions, focusing solely on those containing the hashtags #โรคซึมเศร้า (#depression) and #ซึมเศร้า (#depress). Data were collected using the TweePy tool.

Step 2: Cleaning Data

This process involves removing all URLs from the dataset and eliminating excessive text elements such as punctuation, numbers, and special characters. The “retweet” part is also deleted to reduce processing time, as larger datasets require longer model training times. For example:

• .*เราโดนคุณพ่อเราทำร้ายร่างกายค่ะ .*\r?\n (I was physically assaulted by my father.)

• .*เราอยากดำเนินคดีกับคุณพ่อเพราะเราไม่ได้โดน.*\r?\n (I want to file a legal complaint against my father because I was not assaulted.)

• ทำไมพอคนรอบข้างรู้ว่าเราเป็นซึมเศร้าทุกคนก็เริ่มเปลี่ยนไป ทำเหมือนกับว่าเราป่วย จริงๆเราก็แค่คนๆนึงที่เวลาเราเสียใจเราจะเสียใจมากกว่าคนอื ่นแค่นั้นเ อง เราแค่ต้องการให้คนรอบข้างปฏิบัติกับเราในแบบเดิม ไม่ใช่ Fade .*\r?\n (Why does everyone around us start to change once they learn that we are depressed? It’s as if we are ill. In fact, I am just a normal person who, when feeling sad, experiences deeper sorrow than others. I only want those around me to treat me the same as before, not to fade away.)

Approximately 40% of the originally collected data remained after cleaning. Next, create a test dataset by tagging each sentence as either “pos” or “neg” in the format [(sentence, pos or neg)]. Then, word segmentation was performed using the PyThaiNLP library version 4.1, employing the “pythainlp.tokenize.Tokenizer” class, for example: “ขอกำลังใจหน่อยได้ไหมครับทุกคน รู้สึกเหนื่อยเหลือเกิน” นำมาตัดคำจะได้ ‘ขอ’ ‘กำลังใจ’ ‘หน่อย’ ‘ได้’ ‘ไหม’ ‘ครับ’ ‘ทุกคน’ ‘รู้สึก’ ‘เหนื่อย’ ‘เหลือเกิน’.

The vocabulary data of the system will display the features as follows: [(‘โอ,’ ‘เค,’ ‘พวก,’ ‘เรา,’ ‘รัก,’ ‘ภาษา,’ ‘บ้าน,’ ‘เกิด,’ ‘pos’)].

All words were analyzed to determine their relationships with existing documents and converted them into numerical feature vectors based on word probability and frequency counts within a Bag-of-Words (BoW) model. All unique words in the dataset were used to create a dictionary, disregarding the order of the words. Each word is assigned a unique code, and each element in the vector represents the number of times the word appears in the text. For the “feature_set,” the data consists of word vectors, which are transformed into a format with a specified number of variables, as illustrated in Fig. 4.

A BoW is constructed as a table, in which each row represents a text sample, each column corresponds to a unique word in the vocabulary, and each cell contains the frequency of that word in the respective text. An example of this is shown in Fig. 5.

Step 3: Normalization of Data

This step involves processing vocabulary words, abbreviations, and misspelled words by converting them into their full forms to accurately capture their meanings. For example:

โรงบาล >> โรงพยาบาล (hospital)

อารม >> อารมณ์ (emotion)

เหนคนตาย >> เห็นคนตาย (seeing someone die)

หวัดดี >> สวัสดี (hello)

>> หน้าเศร้า (sad face/unhappy appearance)

: >> ยิ้ม (smile)

>> ยกมือไหว้ (respectful greeting)

??? >> สงสัย (uncertainty/doubt)

>> น้ำตาหยดบนหน้า (teardrops on the face/crying)

Step 4: Building the Model

Using CountVectorizer from the scikit-learn library in Python, the text data were transformed into vectors based on the frequency (count) of each word across all texts. The parameters used in the CountVectorizer are as follows:

• min_df: Excludes words whose document frequency is below a specified threshold during vocabulary creation.

• max_features: If not specified, the vocabulary includes the top max_feature_words ranked by frequency in the dataset.

Subsequently, each model type was tested to identify the one that demonstrated the best prediction performance. The detailed results of these tests are listed in Table 4.

Based on the results in Table 4, the NB classifier demonstrates the highest performance. Therefore, the NB classifier was selected to predict words or messages that indicated the risk of depression among social media users.

Step 5: Model Performance Evaluation

The dataset is divided into training and test datasets. The data are processed using TF-IDF vectorization and then imported into the “MultinomialNB” module. A multinomial NB classifier object is created using the function “MultinomialNB().” Additionally, 10% of the entire dataset was set aside as a test set for evaluation. The results were as follows:

• Positive tweets: 240 rows

• Negative tweets: 300 rows

• Neutral tweets: 0 rows

The “most informative features” were extracted from the model testing and analysis. These features are useful for predicting and classifying data, thereby enhancing our understanding of the learning models.

The three most frequently occurring words in the system were selected based on their highest probability of appearing in text data. For example:

“The ratio of positive (pos) to negative (neg) tweets was \(31.6:1.0\)."

This means that, when comparing the frequency of this word, the system predicts a cheerful (positive) sentiment in approximately 31.6 times more cases, whereas the system predicts depression (negative) in approximately 1 out of 31.6 cases.

““ขอให้” \(=\) True, \(\mathrm{pos:neg} =27.9:1.0\)"

This means that, based on the frequency ratio, the system encounters the word “ขอให้” in texts approximately 27.9 times more often in positive (cheerful) tweets than in negative (depressive) tweets. The system predicted positive sentiments (cheerfulness) in most cases, with a ratio of 27.9 to 1 compared to predicting depression (negative).

““คนเก่ง” \(=\) True, \(\mathrm{pos:neg} =21.7:1.0\)"

Similarly, this indicates that the word “คนเก่ง” appears approximately 21.7 times more frequently in positive tweets than in negative ones. The system tended to classify texts containing this word as beautiful (positive) rather than depressive (negative). Based on these analyses, a confusion matrix was create, as shown in Fig. 6.

The number of predictions made by the system was assessed by comparing the data with the confusion matrix of the developed model. These results can be used to evaluate the overall performance of the system across various metrics, as detailed in Table 5.

The results of the evaluation of the various performance metrics of the model developed by the researcher are summarized in Table 6.

Based on the results in Table 6, the accuracy of the training data for the model was 88.17%, whereas that of the test data was 75.00%.

An observation of the evaluation metrics reveals that, for the positive class (pos), precision is less than recall, whereas for the negative class (neg), precision exceeds recall. This indicates that the developed system demonstrates higher accuracy in predicting words or messages associated with depression in the negative class (neg) than in the positive class (pos). However, it is capable of covering more of the positive (pos) class, reflecting that the system is better at predicting words or messages that pose a risk for depression overall, especially because individuals at risk tend to post negative messages. The neg F1-measure was 76.70%.

Step 6: Deploy and Diagnose Depression

The model is deployed to predict and classify the emotional polarity of tweets over the past 14 days to determine datavariability and summarize whether the user exhibits symptoms consistent with depression, as illustrated in Figs. 7 and 8. The model was downloaded from https://github.com/aucifer16/AI-Depression-Prediction-Thai-NLP.

When the system predicts that an individual is at risk for depression, it sends an alert to the user’s X account, as shown in Fig. 9.

The researcher developed a web interface to present the data, facilitating easier understanding for general users without programming knowledge. The output pages are shown in Figs. 10 and 11.

4.2. Evaluation Results of Prediction System’s Accuracy in Screening Individuals at Risk of Depression

The developed AI system was submitted to 30 experts for evaluation. The performance results are listed in Table 7.

Based on the data in Table 7, most experts rated the performance of the AI-developed media innovation model as high, with an overall mean score of 4.02, representing 80.40%. When comparing the four aspects, it was found that perceptions of security and privacy (security and privacy tests) received the highest rating, with a mean score of 4.11, or 82.20%. Conversely, the aspect with the lowest rating was the accuracy of prediction (precision test), which maintained a high level with a mean score of 3.95, or 79.00%.

5.1. Results of Developing the Media Innovation Model

NLP and ML have demonstrated that the NB classifier is an effective algorithm for analyzing words or messages that reflect depressive states. The developed model can classify texts into positive (\(+1\)), neutral (0), and negative (\(-1\)) categories. The accuracy of the training dataset was 88.17%, while that of the test dataset was 85.00%. These findings indicate a high capacity for the efficient detection of negative messages. In addition, the F1-measure for negative (depressive) messages was 76.70%, confirming the effectiveness of the model in identifying such texts.

The development process of this model aligns with the research conducted by ⁶⁴, which employed sentiment analysis to evaluate emotional intensities from texts, videos, and audio, categorizing severity into three levels: severe, moderate, and mild. Similar to the model developed in this study, such categorization helps improve prediction accuracy in identifying negative messages related to depression.

The data processing steps, including URL removal, surveillance management, and transforming words into vectors using the BoW and TF-IDF techniques, contribute to enhancing the accuracy of the model. These procedures are consistent with ^40,65, which utilized data cleaning and transformation methods to effectively analyze depression-related texts. Moreover, reference ⁶⁶ demonstrated that applying TF-IDF and vectorization significantly increases the effectiveness of the model in accurately predicting negative sentiments.

Compared to other models such as KNN, decision tree, random forest, logistic regression, and AdaBoost, the NB classifier demonstrated superior performance, especially in identifying negative data, which is a key characteristic of individuals with depression. However, reference ⁶⁷ employed time-series forecasting methods to predict influenza outbreaks on Twitter using autoregressive integrated moving average (ARIMA) and Holt–Winters exponential smoothing (HWES) models. These findings indicate that ARIMA can rapidly detect influenza outbreaks in their early stages. Nevertheless, time-series models require a stationary time series, and the data utilized must exhibit stability. Furthermore, the aforementioned study employed tweet data spanning more than two years to ensure model robustness and reliability. This approach may be suitable for predicting chronic depression in patients with psychiatric disorders; however, researchers should allocate a longer period for data collection. This may not be appropriate for treating early stage depressive symptoms that require prompt intervention. A study by ⁶⁸ further confirmed that NB is well-suited for classifying negative texts and performs effectively with small and less complex datasets.

These results indicate that the model can be practically applied, particularly on the X platform, where user messages often reflect emotional states. Early notifications seeking medical advice can help mitigate the impact of depression at an early stage.

However, the accuracy of the model in identifying positive messages remained low, primarily because the training data were skewed toward negative messages. This highlights the need to increase the volume of positive data during the training process to improve the overall model coverage. This analysis also opens opportunities for future development by utilizing more diverse and comprehensive data that capture all dimensions of users’ emotions and sentiments on social media platforms. An important consideration for real-world applications is the actual emotional states of individuals experiencing depression. In other words, there may be discrepancies between the messages expressed and an individual’s actual feelings. Some words may appear to indicate risk but are not truly risky, while others may seem neutral or harmless yet carry significant risk, especially ambiguous or neutral expressions. This poses a challenge for AI systems that may struggle to accurately interpret human emotions or nuanced language. A possible solution to improve the performance of such automated systems is to include a notification or clearly stated message to users, such as: “This system serves only as a preliminary screening tool, and the results generated by artificial intelligence may be subject to error. If users have concerns, they are advised to retake the depression risk assessment before deciding to seek professional medical treatment.” In practical applications, if such a predictive system is to be implemented, psychiatrists or mental health professionals should review the AI-generated results before delivering alert messages to individuals at risk of depression. This is to mitigate the possibility of misclassification or unintended risk due to AI hallucination—cases in which the system encounters messages or language patterns not included in the training data.

A comparable issue was identified in the study by ⁶⁹, which reported that sentiment analysis using VADER and TextBlob was not well-suited for data in the Malay language. These limitations stem from the fact that most algorithms were trained primarily on English data, resulting in poor recognition of Malay language expressions. In addition, the size of the Malay language training dataset was insufficient. Therefore, the authors recommended experimenting with alternative or hybrid ML techniques and suggested that increasing the volume of training data, particularly for DL models, could significantly improve prediction accuracy.

5.2. Evaluation of Prediction Model’s Accuracy

Utilizing AI and NLP to analyze words, messages, or images from social media, revealed an average score of 4.01 (equivalent to 80.20%), indicating good performance. However, expert assessments from various fields have revealed different perspectives. Experts in information technology and computer science rated the model’s performance highest (mean score of 4.38), whereas experts in psychology and psychiatry assigned lower scores (mean score of 3.33). This discrepancy is attributed to the limitations of the model in predicting neutral messages, which yielded the lowest accuracy rate of 76.80%.

Identified issues include the model’s inaccuracy in detecting mildly negative emotional expressions, such as “สิ้นหวัง (hopeless)” or “หมดหวัง (despair),” which often result in high false negatives. This finding aligns with ⁶⁵, which noted that the NB model has limitations in identifying subtle negative sentiments. Furthermore, the complexity of texts containing sarcasm, metaphors, or emojis complicates the prediction of neutral sentiments. This challenge is supported by ³⁵, which emphasized that figurative language and symbolic emotional expressions on social media remain significant obstacles for AI-based models.

Nevertheless, the model demonstrated high accuracy in detecting explicit expressions, such as “อยากตาย (I want to die),” and correctly classified messages unrelated to depression, reflecting its capability when processing clearly defined inputs. However, to enhance the overall performance of the model, we recommend improving the handling of neutral expressions and expanding the keyword databases associated with depressive states. This recommendation is consistent with ⁶⁸, which suggested that incorporating a more diverse dataset would help mitigate the limitations in recognizing emotionally complex messages.