According to the World Health Organization (WHO), suicide is the third leading cause of death among 15-29 years old people, while more than 726,000 individuals die due to suicide every year (WHO, 2024). The latter number is even higher for individuals who perform suicide attempts (Mental Health, Brain Health and Substance Use (MSD), 2021; WHO, 2024). Currently, given that suicidal behavior and suicide-related mental health impact of the COVID-19 pandemic may be long-lasting (Wang and Zou, 2022) and become even more prevalent with elevated attempts in the post-covid era (Guo et al., 2023), predicting clinically significant suicide risk to better inform suicide prevention appears to be a pressing need.

Suicidal behaviors are multi-faceted influenced by multiple and heterogenous factors, including biological, environmental, social, psychological, and cultural ones (Mannix et al., 2020). Taking into account its complex etiology, effective prevention and treatment of suicidality requires an early identification of both risk and protective factors (Franklin et al., 2017). Major depressive disorder (MDD) prevails as the most important factor that increases the risk of suicide, along with the occurrence of former suicide attempts (Agerbo et al., 2002; Eloi et al., 2021). To date, additional risk factors for suicide in people diagnosed with MDD include severity and duration of depression, a history of mental disorder in the family of origin, male gender, traumatic life experiences especially in childhood (e.g., sexual abuse), feelings of hopelessness, anxiety, agitation, drugs and alcohol misuse, and lacking interpersonal relationships (Li et al., 2022). Personality disorder comorbidity (Bolton et al., 2010) and self-injury as well as interpersonal problems with significant others (Ernst et al., 2020) have also been found to increase suicide risk in people diagnosed with depression. On the other hand, the most prevalent protective factors against suicidality among patients with MDD include hope and meaning in life (Luo et al., 2016), a spirituality-related coping style (Dervic et al., 2006), and being employed (Chiu et al., 2023).

In recent years, suicide-related research has shifted more towards modifiable psychological variables that could be part of suicide prevention to empower people against potential suicidal tendencies and minimize suicide risk. Particularly, current research sheds light on the protective role of self-compassion, that is a benign and compassionate approach toward one's own failures, inadequacies, or personal suffering (Neff, 2023). Evidence from a meta-analysis showed a positive relationship between self-compassion and psychological well-being (Zessin et al., 2015). Furthermore, self-compassion was found to be negatively associated with anxiety and depression (Dunne et al., 2018; Sirois et al., 2015), and have a protective effect against suicide risk (Rabon et al., 2019). Recent systematic reviews and meta-analyses revealed that self-compassion is a modifiable factor in depressed patients (Ferrari et al., 2019; Kirby et al., 2017; Wilson et al., 2019). In fact, therapies focusing on self-compassion (e.g., Compassion-Focused Therapy, Acceptance and Commitment Therapy) have been effective in increasing self-compassion and reducing anxiety and depression, mainly through developing emotion regulation, cognitive reappraisal of negative thoughts, and social connectedness (Kirby et al., 2017; Wilson et al., 2019). Therapeutic strategies for self-compassion reinforcement have also been linked with decreased levels of depression in follow-up studies (Ferrari et al., 2019). Similarly, an intrapersonal psychological variable that has been recently found to be associated with suicidality is adult attachment style, which pertains to the way people are predisposed toward significant others and behave in their close relationships. Secure attachment appears to play a protective role against suicidal thoughts and behaviors (Zortea et al., 2021). On the contrary, the two insecure attachment types, that is anxious and avoidant types, have both been positively associated with suicidal ideation (Turton et al., 2022), while anxious attachment (i.e., consisting of fearful and preoccupied subtypes) has been associated with suicide attempts (Macneil et al., 2023). Moreover, insecure attachment has been found to be associated with interpersonal difficulties, which mediate between the type of attachment and the type of suicidal behavior (Stepp et al., 2008). Attachment-based therapies, such as Attachment-Based Family Therapy, are well documented about their effectiveness in populations experiencing depression, suicidality, and trauma. Both in these therapies as well as in individual psychotherapy overall it has been found that the level of attachment insecurity can be modified in the course of treatment through the processing of perceived control and criticism by others, and enhancement of caring and secure interpersonal relationships, as these are initially cultivated in the therapist-patient relationship (Diamond et al., 2016; Ewing et al., 2015).

Yet, the relevant research that examines these two psychological variables (i.e., self-compassion and attachment style) in relation to suicidality remains limited (Amari et al., 2023). The existing evidence so far has indicated that self-compassion is negatively associated with insecure attachment (Huang and Wu, 2024) while the particular aspects of self-compassion (i.e., positive and negative aspects) mediated the relationship between depression and insecure attachment (i.e., anxious and avoidant attachment) (Yang et al., 2024) in different ways. Furthermore, a recent study (Yotsidi et al., 2024) found that avoidant attachment style mediated between self-compassion and suicidality in adult patients with MDD (Daniel, 2006; Travis et al., 2001). Still, little is known about the interplay of these two variables with other sociodemographic and clinical factors for suicidality prediction. Moreover, attachment insecurity and self-compassion have not been examined as yet in predictive artificial intelligence (AI) models for increased risk of suicide in depressed patients.

Artificial intelligence can be particularly helpful in predicting clinically significant suicide risk among patients with MDD. Currently, the use of Machine Learning (ML) techniques to analyze health-related data has gained a lot of attention to its promising results being employed to various applications, such as disease diagnosis, prediction and interpretation (Kokkotis et al., 2022; Moustakidis et al., 2022; Ntakolia et al., 2021a, 2020a). Particularly, in the mental health field, several models and tools for suicide risk prevention have been studied to date (Scatà et al., 2018). Most of these ML techniques focused on developing suicide-related categories through text-classification of collected relevant text, while few of them on feature engineering or sentiment analysis (Bauer et al., 2024; Castillo-Sánchez et al., 2020). For example, in the recent study of Bauer and colleagues (Bauer et al., 2024) large language models (NLPs) were used to understand suicidality via public web-based discussions (on Reddit) around topics related to suicidality. However, even if a plethora of ML models have been suggested, limited work has been done on adopting interpretable ML (IML) techniques and post hoc explainable models to interpret and understand deeper the factors that impact the progression of a mental disorder. Studies applying ML models for suicide-related problems include the use of various ML models for identifying several risk and protective factors of suicidality in US adults (García de la Garza et al., 2021), university students based on Suicidal Behaviors Questionnaire-Revised (SBQ-R) (Kirlic et al., 2023) and the development of ML models for predicting suicide risk in children and adolescents based on electronic health records (Su et al., 2020). Also, studies include the use of ML approaches for suicidal ideation identification (Heckler et al., 2022; Kusuma et al., 2022; McMullen et al., 2021; Roy et al., 2020). Nevertheless, the existing AI models for predicting suicidality have been conducted in the general population rather than in clinical populations (e.g., outpatients or inpatients with major depression disorder) as well as in specific cultural contexts. Given that most deaths due to suicide occurred in low- and middle-income countries (WHO, 2024), more research employing AI approaches for predicting suicide risk in countries with different socioeconomic status and cultural perspectives is needed.

To address this relative lack of empirical attention, the present study applies an interpretable machine learning approach for predicting clinically significant suicide risk in patients with major depressive disorder in Greece triggered by the increased suicide rate after the COVID-19 pandemic. Indeed, the suicide rate in Greece for 2024 was 5.1 per 100,000, with the proportion of men (8.4 per 100,000) being significantly higher compared to women (1.9 per 100,000) (O’Rourke et al., 2024). Concerning the burden on mental health after the COVID-19 pandemic, there was a worsening of depressive symptoms and an increase in suicidal thoughts (Fountoulakis et al., 2021), an increase in anxiety and depression (Efstathiou et al., 2022), as well as higher levels of somatization and distress in specific populations compared to the general population (Louvardi et al., 2020). In children, the deterioration in mental health was linked to unemployment of parents, family conflicts and the prevalence of health problems in children (Magklara et al., 2022; Ntakolia et al., 2022b). As for the student population, it appears that insomnia, poor sleep quality, and fatigue, as well as increases in anxiety, depression, and suicidal ideation were all associated with the pandemic (Kaparounaki et al., 2020). These data gear scientific interest in how the application of AI approaches may be beneficial to suicide prevention by providing new perspectives from cultural contexts that the particular research subject remains understudied.

To the best of our knowledge, this is the first study to apply an interpretable machine learning approach to examine jointly the variables of self-compassion, attachment style, and suicidality in individuals with major depressive disorder, aiming to fill the existing research gap by incorporating in the analyses a significant number of sociodemographic, clinical, and psychological characteristics. To build on previous recent research incorporating hypothesis-driven questions and statistical tools to interpret the data (Yotsidi et al., 2024), this study proposes an IML pipeline via comparative approach and post-hoc explainability for interpreting the suicidal behaviors of an MDD adult population by identifying the most important factors that lead to high suicidal risk. The study incorporates heterogenous data and self-reported psychological questionnaires covering a wide range of factors found in the literature regarding suicidality and MDD population. Additionally, in line with previous research (Brennan et al., 1998a; Neff et al., 2019), the psychological variables included in the study (i.e., self-compassion and attachment style) were examined as both overall constructs (i.e., total scores) and with regard to their particular dimensions (i.e., subscales) to depict potential nuances in predicting populations at risk for suicide. Specifically, the study contributes to the following:•

Collection and combination of data that pertain to suicidality, depression, self-compassion and adult attachment styles.

•

Creation of a dataset of MDD patients with heterogenous features covering sociodemographic, clinical, and psychological data acquired from self-report questionnaires.

•

Development of explainable ML pipeline to identify the most important features that increase suicide risk to patients with major depressive disorder.

•

Comparative and statistical analysis of 6 popular ML classifiers covering different types of classifiers (i.e., tree-based, neural network, regression and max-margin models).

•

Post hoc explainability for quantifying the contribution of each feature to the prediction output.

•

Identification of combined sociodemographic, clinical, and psychological features of patients with MDD in Greece with increased risk of suicide, using interpretable AI as a new research area for suicidality prediction and clinical intervention.