Modeling the Cost-Effectiveness of Interventions to Reduce Suicide Risk Among Hospital Emergency Department Patients

The United States’ Healthy People 2020 goals include a target suicide rate of 10.2 per 100,000 people, representing a 10% reduction from the 2007 rate of 11.3 and a 26% reduction from the 2015 rate of 13.8 (1–3). Reaching this goal requires multiple strategies across different populations and settings (4–7). One key setting is hospital emergency departments (EDs), where at least 500,000 people present annually with self-injury and many more with suicidal ideation. These numbers, along with evidence that self-injury and suicidal ideation are major suicide predictors, suggest that effective ED-initiated suicidality treatment is essential to meeting suicide prevention targets (8–10).

This study examined the current evidence regarding ED-initiated interventions to reduce suicide risk among individuals presenting to general hospital EDs in order to assess whether any such interventions are likely cost-effective enough to support widespread implementation. If so, there is a clinical and economic rationale for adopting such interventions as the new standard of practice, even while researchers develop and test additional suicide prevention interventions.

Several ED-initiated interventions to reduce suicide risk have been found to be efficacious, compared with usual care, in at least one randomized control trial. These vary in approach and intensity, from the “caring letters” approach, which provides messages of psychosocial support to individuals after discharge, to postdischarge telephone contacts that encourage follow-up treatment, to aftercare involving suicide-focused cognitive-behavioral therapy (CBT) (11–19). However, none is yet in widespread use, and the cost-effectiveness of population-based implementation, as well as the potential impact on suicide rates, is currently unknown (17–19). Therefore, we used decision analysis to address these questions, drawing on findings from existing efficacy trials, other available data, and expert opinion regarding suicide risk identification among ED patients (20). We also identified key gaps in existing evidence in order to provide guidance for future research.

Methods

A Markov state-transition model was created to evaluate the cost-effectiveness of three interventions designed to reduce postdischarge suicide risk among adults (ages 18 and older) presenting to general hospital EDs. Patients treated in psychiatric EDs were outside the scope of the study. Our study period was 54 weeks, from the initial (index) ED presentation, divided into nine six-week Markov cycles (for expositional convenience, we report annualized results). The model begins with initial ED presentation; for each subsequent period, each individual can have a new nonfatal suicide event (that is, an attempt or a reattempt), can die by suicide, can die by other manner, or none of these. Our endpoints are case identification, suicide attempts averted, life-years saved, and costs associated with the index visit and any subsequent ED visit and with inpatient and outpatient care following ED presentation. We used TreeAge Pro 15.2.1.0-v20150831 modeling software.

Triage and Case Identification

For tractability, and absent clear empirical guidance, we assumed that each ED patient falls into one of three latent (unobserved) states of suicide risk: high (2.8%), low (9.6%), and none (87.6%). High-risk patients have high near-term risk of suicidal acts, defined here as the next six weeks. Low-risk patients have somewhat elevated suicidal risk, which we define operationally as half the rate of high-risk patients. By definition, no-risk patients have zero near-term suicide risk. Details for all model parameters are presented in Table 1, based on available data and expert opinion (21–28). [A figure illustrating our understanding of patient flow through the ED for high-, low-, and no-risk patients is included in an online supplement to this article.]

We assumed that individuals receive a very brief suicide risk screen by the triage nurse and are triaged by whether they have acute or emergent medical problems (“medical branch”) or not (“psych branch”). Medical-branch patients are hospitalized immediately or treated in the ED; they may be rescreened for suicide risk on the basis of self-report of suicidality or nurse judgment. Medical-branch patients screening positive for suicide risk and all psych-branch patients undergo a full clinical suicide risk assessment. Patients who assess positive are hospitalized or are discharged and referred to outpatient treatment—that is, usual care or one of the alternative interventions. Patients screening or assessing negative are considered nonsuicidal, regardless of their “true” risk state. Screening and assessment aim to identify patients with any (high or low) versus no suicide risk. However, the assumed sensitivity and specificity of screening and assessment and probabilities of hospitalization versus discharge differ for high-, low-, and no-risk patients (and by the presence and severity of general medical conditions).

Outpatient Suicide Interventions

We assumed that all patients identified as having suicide risk are offered usual care after ED or hospital discharge. Our operationalized model of aftercare assumes the following: 65% would receive no specific treatment after discharge, and 35% would receive an average of one initial diagnostic evaluation plus two 45-minute psychotherapy sessions during the 12 weeks postdischarge (additional details are provided in Table 1). Our base estimate is that usual care reduces the rate of suicide attempts or reattempts by 5.25% in the target population over the 12 weeks postdischarge, compared with no treatment—that is, a 15% reduction among the 35% who received treatment. After 12 weeks, these effects decline linearly to zero at the end of the study period.

For patients identified as having suicide risk and discharged from the ED, we consider three ED-initiated outpatient interventions delivered in addition to usual care. When we began this research, these were the only three such interventions that had been tested in a randomized controlled trial. For each intervention, we assumed equal effectiveness for persons experiencing suicidal ideation and nonfatal suicidal acts within the respective risk categories. In this study, we assumed that these interventions are not available to patients hospitalized from the ED.

Postcards.

In this intervention, ED or other personnel send patients a total of eight follow-up postcards as psychosocial support, monthly for four months and then bimonthly (11). We assumed that all targeted patients receive this intervention. On the basis of the results of the most relevant trial, which was conducted in a medical ED among self-poisoning cases, our base estimate is that postcards reduce the rate of suicide attempts or reattempts by 45% compared with usual care alone (and approximately 48% compared with no treatment) (11).

Telephone outreach.

In this intervention, ED or other personnel conduct telephone outreach for one to three months after discharge as psychosocial support and to encourage engagement in follow-up outpatient treatment (12). Per the trial results, 70% of targeted patients receive these calls. Our base estimate is that telephone outreach reduces the rate of suicide attempts or reattempts by 34% in this subgroup of patients, compared with usual care alone, across the study period (12).

CBT.

In this intervention, ED or other personnel connect patients to a suicide-focused CBT program (13). We assumed that 65% of targeted patients participate, on the basis of the percentage of patients who agreed to participate in the clinical trial that tested the CBT intervention. Participants receive an average of nine weekly or biweekly psychotherapy sessions as needed. Our base estimate is that CBT reduces the rate of suicide attempts or reattempts by 50% among participants, compared with usual care alone, across our study period (13).

In each of these interventions, the ED has responsibility for initiating follow-up engagement with the patient. In the respective trials, this was done by ED staff, but in general it could also be done by non-ED staff, such as staff in the same health system (7,29,30).

Costs

Interventions can affect costs two ways: the direct cost of delivery, including the intervention per se and any associated health care use, and by altering the incidence of subsequent suicide attempts and deaths. For delivery costs, we used data on health service use reported in the corresponding clinical trial, inferring the relevant CPT codes and assigning costs based on national rates from the 2014 Medicare Physician Fee Schedule (11–13,31). For ED visits and hospitalizations, we calculated average costs on the basis of an analysis of corresponding events in the Healthcare Cost and Utilization Project (HCUP) database, for individuals discharged alive and for those who died while hospitalized (Table 1). We used HCUP data from four states (Arizona, Florida, Nebraska, and Utah) that either mandate reporting of ICD-9 external-cause-of-injury codes or are known to fully report these codes (32,33). General medical costs not associated with ED visits and related hospitalizations were outside this study’s scope.

Outcomes

At the end of each Markov cycle, individuals either die by suicide, die in another manner, experience a suicide attempt or reattempt, or survive the period with no suicide attempt or reattempt. We assumed that one in 13 suicide attempts results in death; the other 12 persons enter the next Markov cycle with a new ED presentation. Some hospitalized patients are at risk of suicide or of death in another manner or both while in the hospital, depending on their general medical and suicide risk state [see online supplement].

We assumed that individuals maintain the same suicide risk state (high, low, or none) across the study period and that risk of a new suicide event declines with each Markov cycle that does not include such an event. After a new attempt or reattempt, transition probabilities reset to the same levels as after the index event. We assumed that sensitivity and specificity of suicide screening and assessment are the same at the index and any subsequent ED visit.

Results

Cost-Effectiveness

Base case analysis.

Using the base parameter values in Table 1 and the ED patient flow figure in the online supplement, we computed the mean expected costs and life-years per person during the study period for each of the four interventions (usual care, postcards, telephone outreach, and CBT), as well as the incremental costs, life-years, and cost-effectiveness ratio of the three interventions, compared with usual care (Table 2). If there were no deaths in the cohort during the study period, mean life-years would be exactly one.

Mean costs per patient were $1,962 with usual care. Compared with usual care, mean costs were .07% lower with postcards and .05% and .25% higher with telephone outreach and CBT, respectively. Each intervention reduced mortality, on average. Compared with usual care, the postcards intervention was “dominant”—that is, it had both lower costs and better outcomes. The estimated mean incremental cost per life-year was $4,300 for telephone outreach and $18,800 for CBT.

Monte Carlo simulation.

Figure 1 presents data on incremental costs and outcomes of the enhanced interventions compared with usual care, based on Monte Carlo simulation that accounted for uncertainty across the model inputs. Table 1 lists ranges for each input, and values were assumed to follow a beta distribution. We drew 1,000 samples and computed the expected values of the respective outcomes for each vector of sampled parameters. Although there is no definitive benchmark regarding societal willingness to pay (WTP) to reduce mortality, the sloped line in Figure 1 marks the relatively conservative WTP threshold of $50,000 per life-year (34). Trials to the right of that line represent more favorable incremental cost-effectiveness (ICE). For each comparison, ellipses mark the 95% confidence interval (CI) for the estimated ICE.

Compared with usual care, postcards and CBT improved outcomes with ICE at or below $50,000 per life-year with certainty, whereas with telephone outreach the probability of improved outcomes at or below this threshold was 99.5%. The probability of postcards dominating usual care was 94%, and the postcards intervention was cost-effective compared with usual care even when WTP was set at $0 per life-year. The probability that telephone outreach was cost-effective compared with usual care was 96% when WTP was $20,000 per life-year and 80% when WTP was $10,000 per life-year [see online supplement], whereas the probability that CBT was cost-effective compared with usual care was 67% when WTP was $20,000 per life-year and only 1.6% when WTP was $10,000 per life-year.

Sensitivity analysis.

We conducted one-way sensitivity analyses for all inputs, across at least the full range in Table 1. For nearly all inputs, our primary finding—the cost-effectiveness of each intervention compared with usual care when WTP was ≥$50,000 per life-year—remained intact across this range. The sole exceptions were substantial increases in each intervention’s costs and the specificity of suicide risk screening among medical-branch patients. For CBT versus usual care, ICE exceeded $50,000 per life-year if specificity fell from our base rate of 99% (that is, 1% false positives among no-risk patients) to 91% or below. For telephone outreach versus usual care, ICE exceeded this WTP benchmark for specificity below 60%. For postcards versus usual care, incremental costs were positive for specificity below 79%, but ICE remained below $50,000 per life-year even at specificity below 40%. Given our other assumptions, even reducing specificity of suicide risk assessment among medical-branch patients who screened positive and all psych-branch patients to 30% from our base rate of 50% did not raise ICE above $50,000 per life-year for any of the three interventions, compared with usual care.

Given the obvious relevance of each intervention’s effect size and costs, we conducted two-way sensitivity analyses for these parameters (Table 3 and figure in online supplement). In Table 3, each entry represents an alternative assumption about the cost of the respective intervention; for each cost specified, the table reports the minimum risk-reduction ratio (effect size) necessary for the intervention to remain cost-effective at a WTP threshold of $50,000 per life-year or $100,000 per life-year. For instance, if the cost of telephone outreach tripled to $900 from the base value of $300, the ICE of telephone outreach versus usual care would remain below $50,000 per life-year as long as telephone outreach reduced the suicide rate by at least 26% relative to usual care, compared with our base value of 34%.

Population Impact

Our findings regarding the relative cost-effectiveness of each intervention were not sensitive to plausible variation in sensitivity of suicide screening or suicide risk assessment. In short, this was because false negatives do not present an opportunity for improved outcomes but also do not incur additional costs (whereas false positives also present no opportunity for improved outcomes but incur additional costs). However, sensitivity of screening and assessment was important for a different aspect of this analysis: the extent that enhanced intervention could reduce suicides in the target population.

On the basis of the parameters in Table 1, we estimated that 12.4% of individuals presenting for the index ED visit were suicidal, with 2.8% at high risk and 9.6% at low risk. Of these, we estimated that approximately 82% of high-risk patients and 26% of low-risk patients would be identified on the basis of our estimates for screening and assessment sensitivity; the rest were false negatives. However, only identified patients who were discharged from the ED were eligible for the enhanced interventions—that is, approximately 26% of high-risk patients and 3% of low-risk patients, following our assumptions about hospitalization—or about 8% of all individuals in the cohort with elevated suicide risk. This estimate represents an upper bound for impact in the target population. When additional estimates of each intervention’s relative effect size were applied, the Monte Carlo simulations yielded an estimated reduction of suicide deaths in the cohort from the interventions of around 2.5% (CI=0%−11%); the differences between postcards, telephone outreach, and CBT were not statistically significant at conventional levels.

Discussion and Conclusions

Our modeled analysis of ED-initiated suicide prevention interventions found that postcards improved outcomes and reduced costs, compared with usual care. Two other interventions, telephone outreach and CBT, improved outcomes at an incremental cost below a WTP of $50,000 per life-year, a conservative estimate of societal WTP for reducing mortality (34). These findings were largely insensitive to plausible variation in model inputs. In our view, this provides a compelling rationale for widespread implementation of any of these interventions, particularly postcards (or, more generally, letters, because postcards per se might raise privacy concerns).

However, even widespread implementation would have limited population impact because of the low sensitivity of detecting ED patients’ suicide risk and because of health system inefficiencies. Although this aspect of our findings should not deter adoption of interventions that are strongly cost-effective even under current circumstances, it suggests additional avenues for increasing intervention benefits, particularly increasing the sensitivity of detecting near-term suicide risk in the ED population, and for improving coordination and continuity of care across delivery settings after patients leave the ED (35–40). Our findings also highlight the importance of ensuring that evidence-based interventions are economically viable and that appropriate training and technical assistance are available.

This study had important limitations. EDs vary considerably in staffing, treatment protocols, and other characteristics. Given available information and the requirements for parsimony inherent in decision analysis, we tried to create a framework reflective of general hospital EDs in the United States. In any case, our main findings were quite robust to alternative assumptions about patient flow through an ED within the parameter ranges in Table 1. This makes our model useful for studying additional improvements in ED practice, such as efforts to improve sensitivity or specificity of suicide screening, and additional interventions that are found to be efficacious or effective via future trials.

We relied on our own opinions to estimate many inputs, because we could find no corresponding published sources or any data to use for primary estimation. Here, too, we are reassured by the robustness of our findings to alternative parameter values. We considered effects on mortality but not improvements in quality of life, which is a conservative assumption.

Only limited evidence exists regarding the effectiveness and costs of the three interventions (and trials have lacked power to assess their impacts on suicide deaths) (11–19). The trials we reference were conducted under conditions that differed somewhat from those in our model (11–13). The interventions using postcards and telephone outreach were tested outside the United States, and the CBT intervention was tested in a highly urban U.S. setting. All trials were small, and none explicitly reported effects on costs. The postcards trial focused on self-poisoning cases, a subset of ED patients with suicide risk, and it found significant reductions in the number of reattempts but not in the fraction of patients who reattempted (11). A separate study of the postcards intervention focused on all types of self-harm, with similar results (41). The telephone outreach trial also focused on self-poisoning, with null intent-to-treat findings but positive as-treated findings (12). In the CBT trial, one-third of invited patients declined to participate, and adherence to the therapy protocol approached 100% among participants (13). These trials found similar effect sizes across substantially different interventions, indicating that some caution is needed regarding the relative dominance of the postcards intervention, although not regarding its cost-effectiveness compared with usual care.

Despite the limitations, each of these interventions appears to be strongly cost-effective. Additional research on these and other interventions to reduce suicide risk among individuals presenting to general hospital EDs may further aid decision makers. For instance, the Emergency Department Safety and Follow-Up Evaluation (ED-SAFE) study reported that detection of at-risk individuals could be doubled by implementing universal screening in eight U.S. EDs (38). ED-SAFE also tested a form of telephone outreach aimed at supporting treatment adherence, with backup crisis support provided by a call center that is part of the National Suicide Prevention Lifeline, and found that universal screening and telephone outreach resulted in a 27% reduction in the risk of a composite outcome of attempts and deaths, compared with usual care alone (29,42). The Attempted Suicide Short Intervention Program trial in Switzerland, which combined elements of the CBT and postcards interventions, reported an 80% reduction in reattempts (43). Together, these findings support enhancing the standard of care for suicide risk in general hospital EDs with the goal of reaching national suicide prevention targets.