We gathered general information about each participant from interviews and the electronic patient registry. At baseline, we collected information on gender, education, marital status, living situation, children, work status, previous hospitalizations in mental health care, and the number of days in the current hospitalization. At 3 months follow-up, we counted the number of outpatient sessions the participants had received and the number of rehospitalizations since discharge.

Before discharge, participants were screened using the Mini International Neuropsychiatric Interview (version 7.0.2) [35]. If this was not feasible, the interview was conducted at the earliest possible time point within the first few days following discharge. We administered the Structured Clinical Interview for DSM-5 Personality Disorders (SCID-5-PD) [36] after 3 weeks or 3 months, depending on participants’ symptom severity and status, to ensure the validity of the screening results.

Beck Scale for Suicide Ideation-Current (BSSI-C) inquires about the worst level of SI the interviewee has experienced during the last 2 weeks. It consists of 19 items rated on a 0 to 2-point Likert scale, with a total score of 38. Higher total scores on the BSSI-C indicate higher suicide risk. The BSSI-C is structured as an interview and is considered to be a reliable measure of SI [37].

Suicide Status Form-IV (SSF-IV) is a multipurpose tracking and evaluation form for suicide-related cognitions. It inquires about the respondents’ current experiences of psychological pain, hopelessness, stress, agitation, and self-hate. Items are rated on a 5-point Likert scale (from low to high), with a total score of 25. The SSF-IV can be divided into 2 subscales or factors: acute and long-term or chronic SI. Items in the latter subscale contain 3 components: “psychological pain,” “hopelessness,” and “self-hate,” with a total score of 15, whereas the first subscale is represented by the 2 items “stress” and “agitation,” with a total score of 10 [38-40]. In addition, the SSF-IV asks the respondents to self-assess their suicide risk on a scale from 1 to 5, ranging from “low risk” to “very high risk” of suicide. The SSF-IV was administered both before discharge and after 3 weeks and 3 months. The form has been reported to have well-established psychometric properties in terms of reliability and validity [38].

We mapped participants’ previous experiences with self-inflicted harmful behaviors (including SAs) using the Suicide Attempt Self-Injury Count (SASIC) [41]. SASIC is a shorter version of the more extensive Suicide Attempt Self-Injury Interview (Linehan MM and Comtois K, Lifetime parasuicide history, unpublished work, University of Washington, 1996), which includes both a lifetime version and a recent version. We administered the lifetime version at baseline and the recent version at 3 weeks and 3 months.

We used the Outcome Questionnaire-45 (OQ-45) to measure general symptom severity before discharge and after 3 months. The OQ-45 consists of 45 questions about subjective discomfort, interpersonal relations, and social role performance. Items are rated on a 0 to 4-point Likert scale ranging from “never” to “always,” with a total possible score of 180. The questionnaire is sensitive to change, has demonstrated sound psychometric properties, and has been validated in a Norwegian sample [42,43].

The participants completed the Patient Health Questionnaire-9 (PHQ-9) [44] before discharge. This is a very brief assessment consisting of 9 questions about the severity of depression, with each question scored on a 4-point Likert scale ranging from 0 to 3, representing “not at all” to “almost every day,” with a total score of 27. The PHQ-9 is considered a reliable measure [45]. A score above 20 usually indicates severe depression when diagnostic criteria are fulfilled.

The EMA protocol consisted of 26 to 28 items inquiring about current thoughts and emotions, context or relations, and whereabouts, which were rated on Likert scales or with radio buttons. In this paper, we focused our analyses on the EMA item assessing current SI: “I am thinking about taking my life” [46], which was scored on a 5-point-scale, ranging from 1 (“not at all”) to 5 (“very much”) and distributed to participants 5 times a day over the first 10 days after discharge. Scores above “1” were defined as a positive report of SI. Single items are commonly used in the EMA literature to measure subjective state phenomena and may be particularly sensitive to change over short periods of time [17,47,48]. Assessments with single items reduce participants’ burden compared to administering a full traditional multi-item scale (eg, time spent) and may improve compliance and response rates. As the EMA item assessing SI was a single item, internal consistency (eg, Cronbach α) cannot be estimated. However, in a previous investigation, items from a larger EMA battery were psychometrically investigated for convergent validity and reliability [46], including the single item “I think about taking my life.” Reports indicated substantial within-person variability and acceptable prompt-level reliability [46].

Momentary SI and baseline variables were first characterized using descriptive statistics, including means with SDs, medians with quartiles, and frequencies with percentages, along with chi-square and Fisher exact tests.

We examined the temporal trajectory both in terms of time point of the day and over days in separate models. First, we examined how much of the overall variance in momentary SI was divided between and within participants by estimating ICC by running an intercept-only model. Next, a linear mixed-effects model for continuous outcomes was used to fix approximately 5 data points per participant over 10 days. In the analyses of momentary SI trajectory over days, we conducted 2 linear mixed-effects models: one random intercept model and one with random intercept and slope model, assuming missing data at random. The random intercept and slope variables were participant identifiers and time, respectively, and an unstructured variance-covariance structure was used for the random intercept and slope model. For the random intercept model, the participant identifier was specified as the random intercept. For the fixed part in both models, the time (days) variable was treated as a continuous variable, assuming SI followed a linear pattern. We compared the 2 linear mixed-effect models using a likelihood ratio test, and the random intercept and slope model was shown to fit our data best. In a separate analysis, we also modeled time (days) as a categorical variable to estimate the average SI score for each day and to explore the day-to-day pattern of change, as illustrated in Figure 1. In sensitivity analyses, we examined whether the temporal development of SI departed from a linear pattern by modeling SI as categorical, continuous, and with a spline model. In the latter, we defined 4-time intervals (3 knots) based on the small curves in Figure 1. The spline model trajectories overlapped substantially with trajectories modeling SI as categorical and linear, indicating that a parsimonious linear model was adequate for further analyses (Multimedia Appendix 2). For the analysis exploring temporal trajectory within days, we repeated the models’ specifications (random intercept and slope) and replaced the continuous variable (days) with the time point of the day.

In this section, we describe the methods used for each regression model.

Regression model a: We examined baseline differences before discharge between participants who experienced momentary SI during the 10-day period and those who did not. We used logistic regression to explore the association between baseline variables and momentary SI as a binary dependent variable (ie, yes or no momentary SI). The odds ratios (ORs) and 95% CIs were reported for the association.

Regression model b: We examined whether participants’ frequency of peak scores on momentary SI could be associated with baseline characteristics. We first transformed the SI scoring scale from 1‐5 to 0‐4 and continued by extracting the highest SI score for each participant over 10 days. To examine the burden of having high SI scores frequently, we calculated the frequency of the highest SI score during the 10-day period and multiplied that frequency by the highest score for each participant. Thereafter, we explored the association between the weighted frequencies and baseline characteristics using median regression due to the skewed distribution of the frequency data. We provided the estimated ICCs to reflect potential differences in SI variability between participants with an SA and those hospitalized because of suicide risk. Peak scores reflect the most severe symptom state. Although simplified, this approach complements mixed-effects models by emphasizing extreme intensity, a method that has been used in other clinical contexts as an indicator of severity [49,50].

Regression model c: We explored how baseline variables were associated with the intensity and trajectory of momentary SI over the 10-day assessment period. In a series of single mixed-effects linear regression models, each baseline variable, time (days), and baseline variable × time (days) interaction were added and tested. We specified both a random intercept and a random slope and retained the unstructured covariance structure. In all models using mixed-effects linear regression, the restricted maximum likelihood method was used. Given the small sample size, we used a parsimonious approach. Regression coefficients were reported, and P values with 95% CIs were provided for additional information.

For all regression models (a, b, and c), we examined each baseline variable separately to evaluate associations with SI while adjusting for gender and age. In the tables, results from unadjusted analyses are provided for additional information. We report only on associations that remained significant after correcting for multiple tests using a Bonferroni procedure.

We examined how missingness on the EMA surveys was associated with baseline characteristics to evaluate whether some participants were more prone to miss prompts than others. We observed that missingness was associated with gender (men missed more surveys compared to women) but not with age, symptom burden, suicide ideation or cognition, or depression. Additional analyses controlling for gender showed no impact. Consequently, we assumed missing at random (MAR) but not missing completely at random (MCAR). Mixed-effects models have been shown to be robust under MAR assumptions. Analyses were performed in SPSS (version 30; IBM Corp) and Stata/SE (version 19.5).

We observed no significant differences in any baseline symptoms between participants with (28/36, 77.8%) and without (8/36, 22.2%) momentary SI after discharge, after correcting for multiple comparisons (Table 2; Figure 2).

We observed a significant association between the frequency of peak scores of momentary SI and the reason for hospitalization (SA or suicidality; β=17.74, 95% CI 8.56‐26.91; P=.001), with higher mean scores in those hospitalized for suicidality (mean 26.6, SD 48.2) compared with those hospitalized for SA (mean 8.4, SD 12.6). Moreover, baseline levels of self-assessed suicide risk were significantly associated with the peak frequency of SI after discharge (β=6.95, 95% CI 3.95‐9.95; P=.001). See Multimedia Appendix 3 for a detailed overview of test statistics. ICC estimate for momentary SI was 0.59 in the SA group, indicating that 41% of the variance in SI over time was observed within participants. At the same time, the ICC estimate for participants hospitalized with suicidality was 0.79, suggesting that a larger proportion of the variance in momentary SI was due to between-participant differences (79%) in this group. Results indicate that momentary SI among the SA group was characterized by a state-dependent pattern, while momentary SI among participants hospitalized with suicidality was characterized by a more trait-like pattern. As sample sizes in the 2 groups were uneven and small, these estimates should be interpreted with caution.

We conducted a series of separate mixed-effects regression analyses to examine whether baseline variables were associated with the intensity and temporal trajectory of momentary SI over the 10-day assessment period. The SSF-IV chronic subscale was observed to significantly predict the intensity of momentary SI (β=0.13, 95% CI 0.05-0.20; P=.002). See Multimedia Appendix 4 for a detailed overview.

Previous research has indicated that one-third to one-half of patients with a high risk of suicide experience momentary SI after discharge from inpatient psychiatric care [22-29]. The higher prevalence of momentary SI in the present study (28/36, 77.8%) compared to reports from similar settings may reflect a relatively high level of clinical severity in this study. Moreover, overall, few studies have previously been conducted, study samples have been relatively moderate in size, women participants have typically been overrepresented, few suicide attempters have been included, and in some studies, adherence to the EMA protocol has been low [22,23,26-30]. Together, these factors may reflect uncertainty regarding the prevalence of SI after discharge in this population. Our observation of a high prevalence of momentary SI after discharge resonates with findings from a recent epidemiological cross-sectional study. Among 1004 ready-to-be-discharged patients from a psychiatric hospital, 48.9% reported that they experienced active SI [13], indicating that SI is common during the transition from inpatient to outpatient care.

Several previous EMA studies have reported a relative stability in momentary SI over time at the group level, but with considerable individual variations, in line with what we observed. Armey et al [23] followed suicide attempters, suicide ideators, and other patients (controls) hospitalized in psychiatric care with a 21-day EMA study where they inquired about SI. They observed an overall low likelihood of participants to experience SI, but this likelihood increased over the observation period around days 12 to 15. Wallace et al [29] followed newly discharged patients in a 65-day EMA protocol inquiring about the intensity of SI. They reported that the intensity of SI varied substantially between participants over the assessment period, indicating that momentary SI did not follow a uniform or consistent pattern over time. Sels et al [27] followed a similar patient group in a 4-week protocol and concluded that descriptive plots revealed considerable variability between participants and no clear pattern at the group level in terms of the intensity of momentary SI over time. These findings align with ours and indicate that individual differences may better explain the development of momentary SI than time (days and time within days). Further elaborating on this issue, in a longitudinal EMA study of momentary SI among participants with depression, Oquendo et al [51] found that individuals with highly variable and stress-induced momentary SI (state-like SI) differed from those with stable, persistent momentary SI (trait-like SI). When extrapolating upon their findings, the authors argued that patients with stress-sensitive SI are in the target group for rapid-response and emotion regulation interventions (eg, security planning interventions and dialectical behavioral therapy), while patients with trait-like SI may be in the target group for interventions focusing on maladaptive cognitive patterns. These observations and reflections are relevant in our context, as we observed different variances in SI in participants hospitalized with risk of suicide and in those hospitalized following an SA, respectively (see below).

A small subgroup reported no momentary SI after discharge. One interpretation is that these participants quickly resolved their problems upon admittance to the hospital and experienced a drop in symptoms. Such “turning points” have been reported in qualitative studies, explaining how contexts and events may help people change focus [52,53]. On the other hand, participants with no reports of momentary SI were hospitalized for the same duration as other patients, suggesting that clinicians considered them to have a similar need for support. It is possible that this subgroup was less inclined to report, accept, or recognize their own feelings, or that they were characterized by a response style involving denial of problems. Denying SI among high-risk individuals is more common than previously anticipated [54,55]. Multiple pathways to suicidal behavior outside the context of SI have been observed in groups of recent suicide attempters [56]. The Cusp catastrophe model [57] describes how the transition from SI to suicidal behavior may occur through sudden shifts from low-risk to high-risk states, even in the absence of conscious SI but in the presence of a high capability for suicidal acts. In this study, when exploring predischarge differences, we were unable to distinguish between those who later did vs did not endorse momentary SI after discharge. A 10-day observation period provides only a snapshot of the circumstances patients face after discharge. Ultimately, participants should be followed over extensive periods in observational studies to determine who goes on to exhibit suicidal behavior.

Armey et al [23] reported that participants hospitalized with SI were more likely to experience momentary SI compared to a group of controls. At the same time, participants hospitalized after an SA did not show different odds of experiencing SI in the postdischarge period compared with controls [23]. These findings are partly in line with the present results, where participants hospitalized with severe risk of suicide exhibited both a higher frequency of SI and a higher frequency of peak scores on SI. Our findings further suggested that momentary SI was characterized by a more state-dependent (reactive and context-dependent) pattern among those with a recent history of SA compared to those without. The theoretical ideation-to-action framework conceptualizes components that may explain why some people transition from thinking about suicide to acting on suicidal thoughts [58,59]. Erratic shifts in SI, driven by affective and situational or contextual factors in combination with a high capability for suicide, may represent a typical high-risk profile [16,23,58,59]. The 2 participant groups (those hospitalized following an SA and those hospitalized due to severe risk of suicide) may represent partly distinct and partly overlapping populations in terms of capability (attempt) and the mechanisms driving suicide risk and SI. Consistent with prior findings, we expected those with a recent history of SA to experience more frequent SI. These counterintuitive findings may be due to the small and heterogeneous sample available for analysis. Consequently, these observations should be replicated in future studies [23].

Two previous studies [23,30] have reported on the connection between SI before discharge and momentary SI after discharge. As noted earlier, Armey et al [23] observed a higher likelihood of SI after discharge among participants hospitalized with SI. Rogers et al [30] investigated how the frequency, perceived controllability, intensity, and duration of SI before discharge were associated with momentary SI after discharge. Participants’ perceived controllability of SI at baseline was associated with later momentary SI and SAs at a 3-week follow-up, whereas the frequency, intensity, and duration of baseline SI were not.

In the present study, we observed a significant association only between baseline levels on the self-assessed SSF-IV chronic subscale and momentary SI, but not between the clinician-administered BSSI-C and momentary SI. Interestingly, a few prior studies have reported low agreement between self-assessed and expert-assessed levels of SI [60,61], which may help explain this discrepancy. Evaluations of suicide risk guided by structured interviews may capture a phenomenon akin to “expert assessments of objective risk of suicide,” which may only partly correspond to patients’ lived experiences. A closer association between self-reported SI before and after discharge might be expected, as the EMA protocol samples lived experiences. The SSF-IV may be a valuable clinical assessment and treatment tool during hospitalization, as findings indicate a close relationship between its underlying components and postdischarge momentary SI. In this context, the SSF-IV may be used collaboratively by therapists and patients to inform understanding of the patient’s situation and to support treatment planning for the postdischarge period [62,63].