The World Health Organization states that nearly 800,000 people die each year from suicide, one every 40 seconds [1]. Most patients who committed suicide had psychiatric disorders [2]. Among psychiatric disorders, schizophrenia and affective disorders demonstrate the highest risk for suicide [3], but few studies have examined the differences in the suicide-related features of psychiatric disorders [4].

Suicide attempts vary depending on the method, intent, and medical severity of the aftereffects [5]. The length of hospital stay is especially related to the outcome of a patient hospitalized for suicide attempts [3] with some studies stating that patients with longer admissions are at greater risk of postdischarge suicide [6]. Understanding the psychiatric features of patients who stay longer in the hospital might help reduce the length of stay and perhaps their postdischarge outcomes [7]. Several studies have explored the factors associated with the length of stay in suicidal patients, but the results have been inconsistent [8].

Meanwhile, the diagnosis of psychiatric patients so far has relied on categorical diagnostic systems. As the limitations of categorical diagnostic systems became increasingly apparent, the research domain criteria (RDoC) was introduced as an alternate nosology by the National Institute of Mental Health (NIMH) [9]. Natural language processing (NLP) was introduced as one of the ways to use RDoC, and hospital readmission could be predicted with RDoC domains extracted by NLP [10]. Thus, NLP can be used effectively to evaluate psychiatric notes as RDoC domains [11].

In this study, we aimed to explore the differences in the RDoC domains extracted by NLP among patients with depression, schizophrenia, and bipolar disorder who were hospitalized for suicide attempts. We sought to determine whether narrative clinical notes could identify suicide-related features of each disorder. We also investigated how these domains were associated with the length of hospital stay and compared them for each disorder.

Clinical and sociodemographic data were extracted from the electronic health records of patients in the psychiatry inpatient unit at Ajou University Hospital in South Korea between 2010 and 2020. All patients received a Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV-TR or DSM-5 diagnosis from a trained psychiatrist [12,13]. Clinical data included diagnosis (ie, depression, schizophrenia, and bipolar disorder) at admission and chief complaints at admission such as suicide attempts, suicide planning, and suicidal ideations. Sociodemographic data included the age, sex, length of stay, past medical history, and Charlson Comorbidity Index (CCI) score. Admission notes on the patients were extracted for estimating the RDoC scores by NLP. The data were encoded using the Observational Medical Outcomes Partnership (OMOP) common data model (CDM) (version 5) of the [14] in combination with a deidentification procedure. The OMOP-CDM is maintained by the Observational Health Data Sciences and Informatics network, which provides tools to facilitate data analysis. This study was approved by the Ajou University Hospital Institutional Review Board (AJIRB-MED-MDB-21-151), and the requirement for informed consent was waived owing to the deidentification.

McCoy et al [11] previously described a method for estimating RDoC scores from narrative text. In summary, the method evaluates a document using a predetermined set of terms belonging to a given research domain. This list of terms was developed through a group of clinical professionals, including the NIMH RDoC working group. The score can be calculated using a bag of words of the corpus and the count of predefined RDoC terms that appear in the document. For instance, if 10 terms comprise a predefined list and 2 appear in a document, the note would be assigned a score of 2/10 (20%). The list of terms predetermined by the NIMH RDoC working group is publicly available on the internet [11]. The patient admission notes in this study were written in English and Korean. The source texts were 33% in English and 67% in Korean, which were similar for each patient. However, important medical entities such as chief complaints, medical histories, medication prescriptions, and any other important descriptions that directly indicated the patients’ status are represented in English. Moreover, to minimize data loss in the corpus, we systematically translated the corpus into English that was generated by the googleLanguageR package of the R programming language (version 3.6.2) [15]. As a result, we were able to derive RDoC scores from the documents, regardless of the language, as shown in Figure 1.

We conducted a retrospective cohort study to explore the differences in the RDoC domains extracted by NLP among patients with depression, schizophrenia, and bipolar disorder who were hospitalized for suicide attempts. We also investigated how these domains were associated with the length of hospital stay and compared them for each disorder. Baseline demographic and clinical data are expressed as numbers (%) for categorical variables and means (SDs) for continuous variables. Differences between psychiatric disorders were compared using analysis of variance (ANOVA) for continuous variables and Cochran Q tests for categorical variables. Post hoc analysis of the RDoC domains was conducted using independent two-sample t tests. Linear regression modeling with adjustments for the sex, age, and CCI was used to analyze five domains (positive valence, negative valence, cognitive systems, systems for social processes, and arousal/regulatory systems) of the RDoC in different sociodemographic profiles. For each psychiatric disorder, a multiple logistic regression model analyzing the sociodemographic variables and RDoC domains was used to identify the factors associated with the length of hospital stay. For a secondary analysis, a Cox regression model with adjustments for demographics and categorical diagnosis was used to identify each domain associated with the length of hospital stay without controlling the other four domains. As the Mental Health Promotion and Welfare Act in South Korea defines involuntary psychiatric admission within 3 days [16], hospitalization for more than 3 days indicates serious psychiatric problems. Owing to this policy, the distribution of the length of stay was also divided into less than 3 days and more than 3 days. For these reasons, we defined the length of stay as less than3 days or more than 3 days. All analyses were performed using the R programming language (version 3.6.2, R Foundation for Statistical Computing) and the open-source R packages.

In this study, we identified statistically significant differences in the RDoC scores among psychiatric disorders and showed significant associations between the RDoC scores and length of stay for depression and bipolar disorder. The association between suicide and RDoC domains has been reported [17], but very few studies have analyzed how this relationship differs from disorder to disorder. In this regard, the present study investigated whether the RDoC scores derived by NLP differed for each disorder and whether they were related to clinical outcomes such as the length of hospital stay.

The depression group showed the highest negative valence scores among the disorder groups. Previous studies comparing the distribution of domains by diagnosis also showed differences in the negative valence scores for depression and bipolar disorder, and depression and schizophrenia [18]. Our findings not only confirm the previously established association between negative valence and suicide [19] but also suggest that negative valence is particularly associated with depression. Conversely, the schizophrenia and bipolar disorder groups scored higher than the depression group in the cognitive systems and systems for social processes, and no significant difference was found in the scores between the two groups. Several studies have suggested a significant overlap between schizophrenia and bipolar disorder [20], and similarities between these two disorders are prominent in cognitive and social functions [21,22]. Our results are in line with previous results and provide additional information that cognition and social function are important factors in suicide cases involving schizophrenia and bipolar disorder. In the arousal/regulatory systems, the bipolar disorder group scored significantly higher than the schizophrenia group, whereas no significant difference was found in the scores between the depression and bipolar disorder groups. Significant similarities in sleep features representing the arousal/regulatory systems between depression and bipolar disorder have been reported [23]. However, as schizophrenia and bipolar disorder are highly related to sleep disorders [24], it remains unclear whether the scores of the schizophrenia group in the arousal/regulatory systems are more significant than those of the bipolar disorder group.

In the current study, we found that negative valence scores were associated with a longer length of stay in the depression group (OR 1.058, 95% CI 1.006-1.114). The association found between negative valence and depression as well as suicide is consistent with previous findings [19,25]. In the schizophrenia group, no significant relationship was found between the RDoC domain and length of stay, whereas scores in the arousal/regulatory systems were associated with a shorter length of stay in the bipolar disorder group (OR 0.537, 95% CI 0.285-0.815). Contrary to our findings regarding bipolar disorder, a previous study has reported that higher arousal domain scores were associated with a longer length of stay for bipolar disorder [11]. However, sleep disturbance varies with the bipolar disorder phase [26]. These findings show that further consideration of the bipolar disorder phase is needed in interpreting the arousal domain scores with respect to bipolar disorder. Although previous studies found a significant association between the RDoC domain and length of hospital stay, some inconsistencies in prior significant relationships have been identified. For example, one study reported that a positive domain was associated with a shorter stay; however, another showed that a positive domain was associated with a longer stay [11,27]. Thus, the relationship may vary depending on the specific cohort. The association between the length of stay and RDoC domain is unclear for schizophrenia and differs from previous studies with respect to bipolar disorder because our work not only had a cohort different from that of previous studies but also analyzed the relationship between the length of stay and specific disorders.

Our findings show significant associations between the RDoC domains and length of hospital stay for depression and bipolar disorder. This result is consistent with those in existing literature reporting that the estimated RDoC domain scores were associated with the length of stay [27]. Moreover, significant differences and trends in the RDoC domains among depression, schizophrenia, and bipolar disorder were demonstrated. These findings are consistent with those of previous studies on the relationship between negative valence scores and major depressive disorders [25]. Recent publications suggest that cognitive and social functioning factors were observed in schizophrenia and bipolar disorder but not in depression, which is consistent with our findings [28]. On the other hand, using NLP to calculate the RDoC scores is more useful than using structural data alone [11]. For example, the RDoC of the cognitive domain extracted by NLP facilitated stratification of risk for dementia [29]. Our findings further validate of the usefulness and robustness of the RDoC scoring system, which identifies important clinical features in clinical notes. Furthermore, with this validated RDoC NLP tool, our study was conducted by integrating bilingual clinical notes into RDoC domains. Although prior research relied primarily on clinical notes written in English, our results show that the use of RDoC domains through NLP is appropriate for clinical notes that are not in English.

Our study has a few limitations. First, even though we extracted database records of suicidal patients from 2010 to 2020, we could identify only 732 psychiatric admissions in 571 patients. To validate our findings more accurately, a large data set is required. Second, our study has analyzed the conditions most highly related to suicide, but other suicide-related conditions, such as substance use disorders and personality disorders, have not been considered. Because substance use disorders and personality disorders frequently coexist with depression, schizophrenia, and bipolar disorder [30], it was difficult to distinguish between the disorders. Third, the accuracy of the translation of Korean text into English was not evaluated because this was beyond the scope of this study. The accuracy and effectiveness of using clinical NLP algorithms on multilingual clinical documents should also be investigated and validated in future.

Our study showed that the estimates of dimensional psychopathology derived from NLP are associated with the length of hospital stay in suicidal patients with depression or bipolar disorder and vary significantly among suicidal patients with depression, schizophrenia, and bipolar disorder. Therefore, our findings suggest that more attention might be paid to negative valence for depression and arousal/regulatory systems for bipolar disorder in relation to suicide. Additionally, our results may increase the understanding of the differences in dimensional psychopathology among suicidal patients with depression, schizophrenia, and bipolar disorder. We hope that further investigations will clarify the differences in the RDoC scores of suicidal patients and associations between the RDoC scores of suicidal patients and clinical outcomes.