In our initial study [25], we reduced the complex criteria for major depression disorder from the DSM-IV-TR by proposing only three depression factors. These include negative emotions, symptoms, and negative thoughts. Table 2 shows how we mapped the depression factors to the major depressive disorder diagnosis guideline. We try to extract negative emotional terms in the web posts and find relationships between negative emotion terms and symptom terms, and between negative emotion terms and negative thought terms. We call this the Negative Emotion Evaluation model (NEE).
Table 2
Major depressive disorder mapping rules
Criteria | Description | Depression Factor |
1 | Depressed, sad, hopeless, discouraged most of the day | Negative Emotion |
2 | Loss of interest or pleasure in previously enjoyed activities |
3 | Impaired ability to think, concentrate, or make decisions | Symptom |
4 | Increased or reduced appetite. Loss or gain in weight. |
5 | Common sleep disturbance (insomnia/hypersomnia) |
6 | Psychomotor changes, including agitation or retardation |
7 | Decreased energy, tiredness, and fatigue |
8 | A sense of worthlessness or guilt. | Negative Thought |
9 | Frequent thoughts of death, suicidal ideation, or suicide attempts |
We faced a performance issue in the Negative Emotion Evaluation model. We found negative event terms presented in a majority of posts and proposed a new model to evaluate depression in web posts. We assume the event is the key factor and design an Event-Driven Depression Tendency Warning model (EDDTW) [26] to calculate the depression tendency score in a post. We propose a way to extend negative event terms via lexicon, part of speech pattern and the co-occurrence of negative event terms and negative emotion terms. In this study, we apply the Social Readjustment Rating Scale and Daily Hassles Scale to re-create the event lexicon and define the relations among depression factors to form the Stressful Life Event Driven Depression Analysis model (SLEDA).
We proposed two methods to analyze emotion-based and event-based post content for depressive mood detection. However, each model is designed for a specific type of post and not for general usage. In this study, we will combine these two models to analyze depressive moods from either emotion-based or event-based posts. Section 3.1 describes the Negative Emotion Evaluation model. Section 3.2 describes the Event-Driven Depression Tendency Warning model. Section 3.3 describes the Hybrid Depressive Mood analysis.
3.1 Negative Emotion Evaluation model
In previous work, we introduced the four depression factors of negative emotion, triggering event, symptom, and negative thinking in the proposed Negative Emotion Evaluation model. This study mainly investigates the adverse effects arising from emotional words, so we collect the negative emotion terms from web posts. We refer to Plutchik's wheel of emotions [27] to divide negative emotion terms into three levels and assign a strength value for each level (0.25, 0.5, and 0.75). For example, the words "sad" and "grief" differ in intensity. The word "grief" is more serious and assigned the highest score. Table 3 presents a sample of negative emotion related terms and their intensity level.
Table 3
Examples of Negative Emotion terms
Level 1 (0.25) | Level 2 (0.5) | Level 3 (0.75) |
Pensiveness | Sadness | Grief |
Boredom | Disgust | Loathing |
Apprehension | Fear | Terror |
Annoyance | Anger | Rage |
We collected symptom related terms from several resources. First, we extracted symptom terms from collected web posts based on the major depressive disorder diagnosis criteria according to DSM-IV-TR. Second, additional symptom terms were collected from the bilingual MeSH vocabulary words associated with depression symptoms. Third, we collected similar terms from Google suggestions and Wikipedia.
Negative thought is also one of the important indicators in the diagnosis of depression. Many studies and clinical experience note that more than 60% of depression patients might have suicidal thoughts and/or suicidal behavior. Therefore, we gathered the vocabulary of negative thoughts and behavior from the collected articles in order to establish a negative thoughts dictionary. The focus was on specific negative behavior terms, which is the vocabulary related to depression or suicide. For example, the specific terms of “jumping,” or “wrists,” and other words signaling negative behavior may be included in negative behavior narrative content. Some depressive disorder patients disclose suicidal intentions or even acts of attempted suicide in their posts, so we assume that negative thought terms are different from the negative emotion terms. We extracted concentration problem, memory difficulty, and other suicide-related terms from web forum posts, e.g. “guilt” and “suicide”, to generate the negative thought lexicon.
Figure 2 shows the modified framework. Accumulating an amount of negative emotion during a certain period could cause depression tendencies. Given web post B, we want to analyze depression tendency D from the above mentioned web resources by computing the probability \(P\left(D\right|B)\). We utilize depression tendency factors Pf to estimate depression tendency in a web post, and propose a probabilistic model as follows:
$$P\left(D|B\right)=P\left(M|B\right)\times P\left(S|M\right)\times P\left(T|M\right)$$
1
where M represents the negative emotion, S represents the symptom, and T represents the negative thought.
In Eq. 1, we extracted the negative emotion term in each post and summarize all emotion scores to calculate \(P\left(M|B\right)\). We find the distance relationship for all symptom and negative emotion term pairs for \(P\left(S|M\right)\). We also try to find the distance relationship for all negative thought and negative emotion pairs for \(P\left(T|M\right)\). The framework of the Negative Emotion Evaluation model is shown in Fig. 1.
3.2 Stressful life event driven depression analysis model
To enhance the event lexicon for the EDDTW [26] model, we apply the SRRS and Daily Hassles Scale to collect a set of seed event terms, which are called stressful life event terms. Initially, we obtained 135 stressful life event terms. We then separate the stressful life event terms into 5 categories and 46 topics. For each topic, we calculate the average score from the point values of the 43 events in the SRRS and the 63 events in the Daily Hassles Scale as the stress severity score of the topic. After calculating the stress severity score, we rescan the relationships among the stressful life event terms, negative emotion terms, symptom terms and negative behavior terms and replace EDDTW with the stressful life event driven depression analysis model (SLEDA).
We use Eq. (2) to calculate the probability\(P\left(D\right|B)\) in the stressful life event driven depression analysis model.
$$P\left(D|B\right)=P\left(E|B\right)\times P\left(M|E\right)\times P\left(S|E\right)\times P\left(T|E\right)$$
2
where E represents the stressful life event, M represents the negative emotion, S represents the symptom, and T represents the negative thought.
In Eq. (2), we utilize the average stress severity score of stressful life event terms in each post to compute the probability \(P\left(E|B\right).\) To calculate \(P\left(M|E\right),\) we selected negative emotion terms occurring in proximity to descriptions of stressful life events. Next, we also use the co-occurrence relation for all symptom and stressful life event term pairs for \(P\left(S|E\right)\). We also use the co-occurrence relation for all negative thought and stressful life event term pairs for \(P\left(T|E\right)\). The framework of the Stressful Life Event Driven Depression Analysis model is shown in Fig. 2.
3.3 The system architecture of hybrid depressive mood analysis model
In the collected dataset, not every post presented negative events. For these posts, we cannot provide the correct results of depression tendency detection. To solve this issue, we add the original idea to the diagnosis criteria of major depression disorder to calculate the depression tendency score if no event is present in a post. We separate all posts into event and non-event categories. In the category with events, we collect a new stressful life event lexicon and label event terms in the posts and negative emotion terms if the authors complain about something or exhibit depression. In the other group, we follow diagnosis criteria and assume the negative emotion terms are the key factor and trigger the depressive symptoms and/or negative thoughts in the post. In this work, we propose a complementary model to analyze these two categories of posts in Fig. 3.