Abstract:

Depressed students are at high-risk for suicide. Psychological pain, especially pain avoidance, was a more robust predictor for suicide ideation than depression at the behavioral level. Due to suicide as a complex classification model, machine learning algorisms applied to integrate behavioral data and neural characteristic can advance suicide prediction, and the accuracy of multimodality features is superior than clinical interview. The present study aimed to integrate data-driven machine learning algorisms and the three-dimensional psychological pain model to figure out the optimal features in the prediction of suicide ideation.
Seventy-seven college students were recruited by advertisement and divided into three groups: depressed group with high levels of suicide ideation (HSI, n = 25), depressed group with low levels of suicide ideation (LSI, n = 20), and healthy controls (HC, n = 32). All participants completed the three-dimensional psychological pain scale (TDPPS), Beck depression inventory-I (BDI), Beck suicide ideation inventory (BSI), and the self-referential affective incentive delay task (SAID). The value of support vector based on machine-recursive feature elimination (RFE-SVM) algorithm applied to combine the scale scores, resting state and punitive-related EEG components for feature ranking in a nonlinear way. Firstly, in order to explore the importance and specificity of pain avoidance and pain processing related EEG characteristics in the classification model of suicidal ideation, the classification models of painful feeling and pain avoidance were established, which helped to understand the importance of EEG features under the self-referential punitive condition in above two models. Secondly, in order to differentiate the important prediction features of classification models of depression and suicide ideation, depression classification models with single- and multimodal features as input variables were established separately. Thirdly, the single- and multimodal classification models of suicidal ideation were established. Further, the important features of the classification model of suicide ideation, depression and pain avoidance would be compared with each other.
Results showed that: (1) Scores of pain avoidance in the HSI was higher than the LSI group (p < 0.001, Table 1). (2) The multimodal psychological pain-based model for suicide ideation classification (Accuracy = 85.66%, Precision = 0.82, Recall = 0.73, AUC = 0.92) was sufficient and superior than the EEG single-modal model (Accuracy = 65.43%, Precision = 0.40, Recall = 0.21, AUC = 0.57, Table 2). Importantly, the pain avoidance and BDI scores ranked the top two features in the classification model of suicide ideation, whereas painful feeling and pain arousal subscale scores ranked the top two features in the multimodal classification model of depression (Accuracy = 73.83%, Precision = 0.76, Recall = 0.83, AUC = 0.80, Table 2 and Table 3). The top four EEG features in the multimodal classification model of suicidal ideation are CNV, LPP, target-P3 and feedback-P3 under the self-referential punitive conditions, while the top four EEG features in the multimodal classification model of depression are FRN, LPP under self-referential reward condition and FRN, target-delta under self-referential punitive conditions. The EEG optimal features of overlap in the pain avoidance and suicide ideation classification models were the LPP and target-P3 under self-referential punitive conditions. (3) The powers of delta and beta band were negatively correlated with the BSI-W (r_delta = -0.26, p_delta < 0.05; r_beta = -0.24, p_beta < 0.05) and pain avoidance subscale scores (r_delta = -0.26, p_delta < 0.05; r_beta = -0.23, p_beta < 0.1). The FRN amplitude under other- and self-referential punitive conditions were negatively corelated with the pain avoidance subscale scores (r_other = -0.28, p_other < 0.05; r_self = -0.19, p_self < 0.05). In the HSI group, power of delta elicited by positive feedback under self-referential conditions was lower than those under other-referential conditions (p = 0.084). In the HSI group, the amplitude of LPP in other-referential punitive conditions was higher than those under reward (p = 0.003) and neutral conditions (p < 0.001), whereas in the LSI group, the amplitude of LPP under self-referential punitive conditions was higher than that under neutral conditions (p = 0.006).
As a pilot study, the current study provided a support for the prominent role of pain avoidance and its related neuroelectrophysiological correlates in the prediction of suicide. The clinical significance of this results will be discussed.

Key words: three-dimensional psychological pain, suicide ideation, machine learning, EEG