We analyzed data obtained from the Korea National Suicide Survey [17], which was a nationwide multicenter study of subjects from two cohorts comprising individuals who attempted suicide and were recruited by retrospective chart review and those who attempted suicide and completed psychiatric evaluations by on-call psychiatric residents. The subjects from the second dataset were used in this study. All individuals who attempted suicide visited the emergency room (ER), and they were evaluated in semistandardized interviews at 17 medical centers across South Korea from May 1, 2013, to November 7, 2013. Deaths in the ER were excluded from the data. Among 1359 individuals who attempted suicide, 1112 were included in the final analysis after excluding missing data from the Columbia Suicide Severity Rating Scale (C-SSRS) and Beck Suicide Intent Scale (SIS) (Figure 1).

The outcome of this study involving the medical lethality of suicide attempts was assessed by a clinician and classified based on the “actual lethality or medical damage” using the C-SSRS. The validated Korean version of the C-SSRS was used [18], and lethality was rated as follows: 1, no physical damage or very minor physical damage (eg, surface scratches); 2, minor physical damage (eg, lethargic speech and mild bleeding); 3, moderate physical damage (eg, conscious but sleepy); 4, moderately severe physical damage (eg, comatose with reflexes); 5, severe physical damage (eg, comatose without reflexes); and 6, death [19]. We used a lethality scale with the following two categories: score 3, a less lethal outcome of a suicide attempt and score >3, a lethal outcome of a suicide attempt.

The SIS was used to identify the elements of suicidal intent thoughts, behaviors, and communications [20]. The scale contains 15 questions (ie, SIS 1-15), and all of the items are scored on a scale from 0 to 2 for severity, where the total sum of the scores ranges from 0 to 30. In this study, we calibrated the SIS scale from 1 to 3 to calculate the relationships among the features. The SIS comprises the following two parts: objective circumstances of the attempt and the subject’s self-reported intentions and expectations regarding the attempt. The objective factors (SIS 1-8) are as follows: SIS 1, isolation; SIS 2, timing of intervention feasibility; SIS 3, active or passive precautions against discovery or intervention; SIS 4, acting to get help during or after the suicide attempt; SIS 5, final acts in anticipation of death; SIS 6, active preparation for the suicide attempt; SIS 7, suicide note; and SIS 8, overt communication of intent before the suicide attempt. The subjective factors (SIS 9-15) are as follows: SIS 9, alleged purpose of the suicide attempt; SIS 10, expectations of fatality; SIS 11, conception of a method’s lethality; SIS 12, seriousness of the suicide attempt; SIS 13, attitude toward living or dying; SIS 14, conception of medical rescuability; and SIS 15, degree of premeditation.

We also used clinical data reported from the ER as confounding variables in prediction modeling. In total, 14 confounders were considered, including sex, age, marital status, religion, monthly income, living status, educational level, urbanicity, ER visit date, ER visit on a weekend, ER visit time, admission route, admission transportation, and discharge date. All of the confounders were collected as numerical values, with coded indices or quantitative values as follows: sex (1=male, 2=female), marital status (1=single, 2=married, 3=living together, 4=separated, 5=divorced, 6=widowed), religion (1=Christian, 2=Buddhist, 3=Catholic, 4=Atheist, 5=other), living status (1=living with family, 2=living with somebody, 3=group facilities, 4=living alone), educational level (1=none, 2=elementary school, 3=middle school, 4=high school, 5=undergraduate or higher), ER visit on a weekend (1=yes, 2=no), admission route, admission transportation, and monthly income (self-reported in Korean currency). To represent the date records numerically (ie, year-month-day, ER visit date, and discharge date), we transformed the original date into a decimal year value (eg, 2013-6-30=2013.492). The detailed equation for the numerical transformation of the dates is presented in the supplementary source code [21].

For each pair of 15 suicidal intent items for each individual, we generated three relationship signatures comprising the interaction terms (I), harmonized average (H), and geometric angle differences using the tangent function (T). The definitions of the three relationships between the ith and jth element of SIS in the pth individual (R(I, H, T)) are represented by the following equations:

where S_i ^p indicates the ith SIS element in the pth individual, and I_i,j ^p determines the level of interaction between the ith and jth SIS elements. To represent the overall intensity in a sensitive manner, we utilized the harmonic mean of a pair of elements (H_i,j ^p). According to the differences in the sequential combination of a pair of elements, such as [S_i ^p=2 > S_j ^p=3] and [S_i ^p=3< S_j ^p=2], T_i,j ^p presents a single scalar value for the paired elements.

The chi-square test and Student t test were performed to compare variables for suicide attempts with lethal and nonlethal medical outcomes.

Among 1112 individuals who attempted suicide, 190 (17.1%) had suicide attempts categorized as lethal medical outcomes (Table 1). According to the C-SSRS–based fatality of attempt outcomes, we classified the individuals as those who had lethal attempts (n=190, C-SSRS severity 3) and those who had nonlethal attempts (n=922). More male individuals had lethal suicide attempts than nonlethal suicide attempts (107/190, 56.3% vs 357/922, 38.7%). The mean age of those who had lethal suicide attempts was higher than that of those who had nonlethal suicide attempts (47.3 years vs 42.3 years).

The mean total SIS score was higher for those who had lethal suicide attempts than those who had nonlethal suicide attempts (30.23, SD 6.19 vs 25.06, SD 5.61). The total SIS score is the sum of the 15 graded elements of the SIS, such as the scale of isolation from 1 to 3 (complete isolation).

Using GLASSO, we determined the weighted correlations among the assessed SIS elements according to suicide attempt fatalities [23]. We constructed two networks comprising nodes representing the SIS elements (eg, degree of isolation) and edges depicting the relationships among nodes, which indicated the distinct relationships between the elements of suicide in those who had lethal and those who had nonlethal suicide attempts (Figure 1). The statistical significance of the correlations assessed among the SIS elements using GLASSO were determined based on random distributions of the SIS elements (P<.05 for random distributions). Finally, we represented the distinct relationships between the suicidal intents of those who attempted suicide (lethal and nonlethal cases).

Fifteen nodes for suicidal intents were linked via 17 edges for lethal suicide attempts (n=190) (Figure 2A). Among the 922 individuals who had nonlethal suicide attempts, there were 16 relationships (ie, edges) among 14 suicidal intent elements (ie, nodes) in the nonlethal suicide attempts (Figure 2B). The edges between nodes represent the positive or negative relationships between nodes based on the GLASSO results (P<.05 for edges based on a random distribution). Among individuals who had lethal attempts, the fatal outcomes of suicide attempts were more tightly linked (ie, associated) with the suicidal intents compared with those who had nonlethal attempts, and the nodes for the SIS elements were loosely connected or separated in those who had nonlethal attempts (Figure 2A and B). Thus, the close connectedness of the suicidal intent elements, including the concept of lethality of the method, was stronger among those who had lethal attempts.

The topological properties of the network, such as the central node that had the largest number of relationships with other nodes and the average of the shortest paths (ie, degree of centrality), were employed to determine the central suicidal intent elements for the lethal and nonlethal attempts (Figure 2C-F). Among the lethal suicide attempts, the fatality of suicide attempt (“Fatality” node) and the conception of a suicide method’s lethality (“ConMeth” node) were strongly associated with other suicidal intent elements (Figure 2E). In Figure 2C, the y-axis denotes the average shortest path, which was a bottleneck and a central node, and the “ConMeth” node was ranked highly among those who had lethal attempts (Figure 2C). However, the alleged purpose of suicide attempt (“APurpo” node), including “to manipulate the environment,” was a crucial intent element in the minds of those who had nonlethal attempts (Figure 2D and F). Moreover, among those who had nonlethal attempts, suicide method-related features (eg, the conception of a method’s lethality and the expectation of fatality), which are closely linked to lethality among those who had lethal attempts, were completely disconnected from the other suicidal intent nodes (Figure 2D and F).

Thus, we elucidated the relationships between the suicidal intent elements in those who had lethal and those who had nonlethal suicide attempts. The conception of a method’s lethality (“ConMeth” node) was a central suicidal intent element, which was clearly related to lethal suicide attempts, and it was connected with the initiation of attempts, such as the nodes for expectation of fatality (“ExFatal” node) and seriousness of attempt (“SeriAtt” node).

Owing to the structural differences in the networks of antecedent behaviors according to the lethal or nonlethal outcomes of the suicide attempts, we generated three relationship signatures for each pair of SIS elements comprising the interaction terms (I), harmonized average (H), and geometric angle differences using the tangent function (T). In addition to the 15 SIS elements and 14 types of clinically reported data collected by the ER, including age, admission date, and living status, three relationship signatures were prepared for all possible SIS combinations for each individual. We represented the pairs of SIS elements as specific numeric values, and 315 relationship features were obtained among the 105 combinations of SIS elements for an individual. We built E-GONet based on a CNN, which is a subclass of deep learning. The overall structure of E-GONet comprises input and output layers, as well as multiple hidden layers (convolutional layers, pooling layers, and fully connected layers). The schematic structure of each layer is shown in Figure 3A. Multimedia Appendix 1 and Multimedia Appendix 2 describe the detailed structures of the E-GONet model.

Figure 3B shows the results obtained from the performance evaluation. We evaluated the predictive performance of the E-GONet model by 10-fold cross-validation. We used the same dataset for the performance comparison with E-GONet and to establish two prediction models with linear regression and random forest (ie, an aggregated decision tree model) [27] methods.

E-GONet performed better than the linear regression and random forest methods (E-GONet increased the F1 score up to 3.4%, and the mean increase in the F1 score was 2.1%; E-GONet also increased the area under the precision-recall gain curve [AUPRG] up to 32.1%, and the mean increase in the AUPRG was 18.1%) [28]. Besides, the positive-predictive value (PPV; precision) comprising the rate of correctly identifying lethal attempts was highest with the E-GONet predictions (0.59). As generally noted in the clinical field, our dataset was relatively imbalanced (lethal 190, nonlethal 922). In analysis involving imbalanced data, sensitivity, PPV, F1, and AUPRG have been used for performance evaluation instead of specificity, negative-predictive value, and accuracy.

The analysis of the contribution of learning features (Multimedia Appendix 3 and Multimedia Appendix 4) showed that most of the contributions of confounding variables (such as level of education) were negligible for the predictive performance of E-GONet. However, the relationship signatures between the SIS element pairs contributed greatly to the superior performance of the E-GONet model (the values of R(I, H, T)). As depicted in Multimedia Appendix 3 and Multimedia Appendix 4, the saliency heatmaps of our model highlighted the contribution of SIS relationship features. The first two rows of the feature importance matrixes were the confounders (age, sex, income levels, etc) and SIS elements (SIS 1-15). Out of those features, only age and income level contributed to E-GONet training. On the other hand, as presented in the figures, the developed relationship features of SIS elements were more important for CNN model training. Interestingly, out of all relationships among SIS elements, the relationship with SIS element 11 (conception of a suicide method’s lethality) was the biggest contributor to predictions. This is highly consistent with the network modeling of SIS elements in Figure 2.

Moreover, Figure 3C shows the evaluations of the performance of the models established without the SIS-based relationship features. The predictive models based on linear regression and random forest exhibited similar or lower performance after introducing the SIS-based relationship features, because these classical methods could not patternize the relationship features of suicide elements. However, E-GONet trained and improved performance via the vectorized relationships in high-dimensional spaces. As a result, the relationships between SIS elements increased the performance of the E-GONet model by 60% in precision, 6% in F1, and 8% in AUPRG (precision [without relationship/with relationship]=0.0/0.59, F1=0.75/0.81, and AUPRG=0.65/0.73). The full spectrum of the AUPRG displayed the training and fitting process of our deep learning approach in a very detailed manner (Multimedia Appendix 5). In Multimedia Appendix 6, the standard deviations of AUPRGs are presented via 100 trials of 10-fold cross-validation settings.

Therefore, identifying the lethality of attempt outcomes is feasible with deep learning through the major contributions of the relationships among SIS elements (ie, mutual interactions of antecedent behaviors). To allow the use of our method in clinics, we have made all of the source codes for the analytics available via the internet, including the network-based analytics, E-GONet model, and data preprocessing methods [21].