Introduction
Major depressive disorder (MDD) affects 1 in 5 adults over the life course (Hasin et al., Reference Hasin, Sarvet, Meyers, Saha, Ruan, Stohl and Grant2018), and relapse after remission is common and can be disabling (Monroe & Harkness, Reference Monroe and Harkness2022). Patients in remission on antidepressants face a key decision: whether to continue maintenance treatment or discontinue medication with symptom monitoring and planned restart if symptoms return (active monitoring). Maintenance lowers relapse risk (Kato et al., Reference Kato, Hori, Inoue, Iga, Iwata, Inagaki, Shinohara, Imai, Murata, Mishima and Tajika2021; Kishi et al., Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023), but may bring burdens such as sexual side effects, weight gain, emotional numbing, and an unwanted sense of dependence on medication (Cartwright et al., Reference Cartwright, Gibson, Read, Cowan and Dehar2016; Gafoor, Booth, & Gulliford, Reference Gafoor, Booth and Gulliford2018; Masdrakis, Markianos, & Baldwin, Reference Masdrakis, Markianos and Baldwin2023; Reichenpfader et al., Reference Reichenpfader, Gartlehner, Morgan, Greenblatt, Nussbaumer, Hansen, van Noord, Lux and Gaynes2014). Although relapse-prevention psychotherapy is an effective option for some patients (Zaccoletti et al., Reference Zaccoletti, Mosconi, Gastaldon, Benedetti, Gottardi, Papola, Ponzi, Purgato, Naudet, Cristea, Barbui and Ostuzzi2026), access and acceptability vary (Mojtabai et al., Reference Mojtabai, Olfson, Sampson, Jin, Druss, Wang, Wells, Pincus and Kessler2011; Zhu et al., Reference Zhu, Huntington, Haeder, Wolk and McConnell2024), and routine decisions often still center on maintenance versus active monitoring (Hockenberry, Joski, Yarbrough, & Druss, Reference Hockenberry, Joski, Yarbrough and Druss2019).
Clinical guidelines on antidepressant maintenance provide more prescriptive recommendations for patients at the extremes of relapse risk (Gelenberg et al., Reference Gelenberg, Freeman, Markowitz, Rosenbaum, Thase, Trivedi and Van Rhoads2010; Lam et al., Reference Lam, Kennedy, Adams, Bahji, Beaulieu, Bhat, Blier, Blumberger, Brietzke, Chakrabarty, do, Frey, Giacobbe, Gratzer, Grigoriadis, Habert, Ishrat Husain, Ismail, McGirr and Milev2024; McQuaid et al., Reference McQuaid, Buelt, Capaldi, Fuller, Issa, Lang, Hoge, Oslin, Sall, Wiechers and Williams2022; National Institute for Health and Care Excellence, 2022), with long-term maintenance generally discouraged for patients at lower risk and more strongly recommended for those at higher risk. However, the guidelines leave two important decision problems unresolved. Some patients at these extremes may have preferences that conflict with the guideline-based default and want to know how strong those preferences would need to be to justify choosing the alternative strategy. Others fall between the extremes, where guidance is less determinate, and the decision is inherently preference-sensitive. In both settings, quantifying the additional time spent taking antidepressants required to reduce future time spent depressed can help patients, in conversation with their clinicians, judge whether maintenance is worth the treatment burden.
Informed shared decisions require that patients understand relevant trade-offs and relate them to their own values (Stacey et al., Reference Stacey, Lewis, Smith, Carley, Volk, Douglas, Pacheco-Brousseau, Finderup, Gunderson, Barry, Bennett, Bravo, Steffensen, Gogovor, Graham, Kelly, Légaré, Sondergaard, Thomson and Trevena2024). Randomized maintenance-discontinuation trials and meta-analyses have reported that continued maintenance reduces 1-year relapse risk from roughly 40% to 20% (Hu et al., Reference Hu, Xue, Ni, Guo, Fan and Du2024; Kato et al., Reference Kato, Hori, Inoue, Iga, Iwata, Inagaki, Shinohara, Imai, Murata, Mishima and Tajika2021; Kishi et al., Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023). However, relapse is a coarse endpoint; decision-making also depends on time spent well versus unwell over months to years (Judd et al., Reference Judd, Akiskal, Maser, Zeller, Endicott, Coryell, Paulus, Kunovac, Leon, Mueller, Rice and Keller1998). These trials are typically enriched for recurrent illness (a mean of 4–5 prior episodes; Kishi et al., Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023), which may limit generalizability to patients with more moderate baseline risk. Prior decision-analytic models have focused largely on cost-effectiveness rather than clinical trade-offs for shared decision-making (Lokkerbol et al., Reference Lokkerbol, Wijnen, Ruhe, Spijker, Morad, Schoevers, de Boer, Cuijpers and Smit2021; Nuijten, Reference Nuijten2001; Sobocki et al., Reference Sobocki, Ekman, Ovanfors, Khandker and Jönsson2008). Existing evidence thus does not provide a practical framework for judging how the maintenance-versus-monitoring trade-off varies with recurrence probability or how strongly a patient would need to value avoiding depression relative to continued antidepressant use for one strategy to be preferred over the other.
To address this gap, we developed a health-state transition model (microsimulation) to quantify the maintenance-versus-monitoring trade-off across recurrence probabilities, together with a decision framework that links these model estimates to patient preferences. The model integrates trial and longitudinal cohort evidence and builds on our prior modeling work (Meyerson et al., Reference Meyerson, Ross, Kennedy, Hoyle, Chhatwal, Wang and Smoller2025). It estimates the additional time spent taking antidepressants required to avert 1 month of depression under continuous maintenance versus active monitoring. This follows a standard decision-analytic approach in which health-state burden accrues over time and preferences determine the relative value placed on different health states (Gold, Siegel, Russell, & Weinstein, Reference Gold, Siegel, Russell and Weinstein1996). The broader framework then compares the medication use required to avert depression with the amount a patient would consider acceptable, elicited through patient–clinician discussion informed by the patient’s lived experience of depression and antidepressant treatment. In the present analysis, the modeled trade-off varied nearly eightfold: maintenance required approximately 1.5 additional years of antidepressant use to avert 1 month of depression in the highest recurrence-probability group, compared with nearly 12 years in the lowest. Using the broader decision framework, we then examine how these trade-offs interact with patient preferences to influence when continuous maintenance versus active monitoring would be favored.
Methods
Overview
We developed an individual-level health state transition model (microsimulation) to inform antidepressant continuation decisions after remission from a major depressive episode. The simulated decision point was completion of an 8-month continuation phase after remission, when a patient and clinician may consider either ongoing antidepressant maintenance or guided discontinuation with active monitoring. The model compared continuous maintenance with active monitoring, defined as antidepressant discontinuation following the tapering practices represented in the underlying trials, symptom monitoring, and planned antidepressant restart after detected recurrence, over a 5-year horizon.
The model was designed to estimate how the trade-off between continuous maintenance and active monitoring varies across modeled recurrence-probability groups and patient preference thresholds. The model includes depressive health states and transitions among them, including remission, relapse into an active major depressive episode, and recovery after relapse (Supplementary Figure S1).
For model inputs, we selected sources to supply complementary quantities that no single study reported across the full decision horizon. Kishi et al. (Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023), a systematic review and meta-analysis of maintenance-discontinuation randomized trials provided the primary evidence on relapse under maintenance versus discontinuation in highly recurrent depression. These trials primarily represented highly recurrent populations, with a mean of approximately 4–5 prior depressive episodes. To represent lower recurrence probabilities and longer-term outcomes, we also used Bukh et al. (Reference Bukh, Andersen and Kessing2016), a prospective study of patients after a first lifetime depressive episode. Key characteristics and findings of these empirical anchors are summarized in Box 1. Selected model parameters and key structural assumptions are summarized in Table 1 and Box 2; full derivations are provided in the Supplementary Material.
Key Empirical Anchors for Model Parameterization
Maintenance versus discontinuation in recurrent depression. Kishi et al. (Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023) synthesized 35 double-blind maintenance-discontinuation randomized trials involving 9,442 adults with recurrent depression. Trial populations had a mean of approximately 4–5 prior depressive episodes. Cumulative relapse at 12 months was approximately 44% after discontinuation and 22% with maintenance. These data provided the primary randomized anchor for relapse under maintenance versus discontinuation and informed the high recurrence-probability group.
Recurrence after a first depressive episode. Bukh et al. (Reference Bukh, Andersen and Kessing2016) prospectively followed 301 patients after a first lifetime depressive episode. Cumulative recurrence after remission was 9.0% at 1 year and 31.5% at 5 years. These data anchored recurrence probabilities in the lower-probability groups and provided longer-term recurrence information beyond the follow-up available in most maintenance-discontinuation trials.
Selected Assumptions of the Model
Selected assumptions underlying quantitative validity:
-
• A patient in remission can relapse at any time. We model relapse as more likely shortly after remission, based on patterns observed in maintenance-discontinuation trials.
-
• A history of recurrence predicts future recurrence. We estimate this relationship by comparing relapse rates in highly recurrent trial populations with longitudinal cohorts following first depressive episodes.
-
• Antidepressant maintenance reduces relapse risk. The magnitude of this effect was estimated from randomized maintenance-discontinuation trials.
-
• Restarting antidepressants after relapse can help patients recover. We assume, with some support from re-initiation studies, that remission after restart follows a course broadly similar to remission after initial antidepressant treatment.
-
• Some depressive episodes lead to prolonged residual symptoms. We estimate their duration and residual symptom burden using long-term cohort studies.
Selected assumptions underlying clinical interpretation:
-
• Time depressed and time taking antidepressants are meaningful because they map onto patients’ lived experience.
-
• Patients can clarify how much time taking antidepressants they would accept to avoid 1 month of depression.
-
• More time depressed means proportionally greater depression burden.
-
• More time taking antidepressants means proportionally greater treatment burden.
-
• Past severity and functional consequences can inform how burdensome another month of depression would be.
-
• Past side effects and medication-related preferences can inform how burdensome continued antidepressant use would be.
Selected model parameters
Table 1. Long description
The table is organized into four columns: Parameter, Base case, Sensitivity range, and Source.
Section 1: Relapse probability by patient group (monthly probability).
* High (4 to 5 prior episodes): Base case 0.023; Sensitivity range 0.013 to 0.040; Source Kishi et al., 2023.
* Moderate (interpolated): Base case 0.013; Sensitivity range 0.009 to 0.018; Source Interpolated between Low and High groups.
* Low (first episode, moderate plus severity): Base case 0.0072; Sensitivity range 0.005 to 0.010; Source Bukh, Andersen, and Kessing, 2016.
* Very Low (first episode, mild severity): Base case 0.0033; Sensitivity range 0.0023 to 0.0046; Source Bukh et al., 2016.
Section 2: Treatment effect.
* Protective effect of maintenance antidepressant, observed composite hazard ratio: Base case 0.470; Sensitivity range 0.420 to 0.530; Source Zhou et al., 2022.
Section 3: Model structure.
* Continuation phase after remission, months: Base case 8; Sensitivity range Fixed; Source Lam et al., 2024 and McQuaid et al., 2022.
* Monitoring interval (active monitoring), months: Base case 3; Sensitivity range 1, 3, 6; Source Clinical practice and N I C E 2022.
* Time horizon, years: Base case 5; Sensitivity range 3, 5, 10; Source Bukh et al., 2016 follow-up duration.
a This composite hazard ratio reflects the combined effect of maintenance antidepressants on both non-chronic and chronic relapse; the model decomposes it into separate hazard ratios for simulation (see Supplementary Table S1).
b Scenario values tested in univariate sensitivity analysis. All other ranges are 95% CIs used in probabilistic sensitivity analysis. A complete parameter table is provided in the Supplementary Material.
We followed CHEERS 2022 guidance where applicable (Husereau et al., Reference Husereau, Drummond, Augustovski, de Bekker-Grob, Briggs, Carswell, Caulley, Chaiyakunapruk, Greenberg, Loder, Mauskopf, Mullins, Petrou, Pwu and Staniszewska2022). Analyses used R, version 4.4.1 (R Foundation for Statistical Computing). The Mass General Brigham Institutional Review Board determined that this work did not constitute human subjects research due to reliance on aggregate data. The analysis plan was not preregistered.
Treatment strategies and decision structure
We compared two treatment strategies, illustrated in Figure 1a. Under continuous maintenance, patients remained on antidepressants throughout follow-up. Under active monitoring, patients discontinued antidepressants at model entry after completing continuation treatment for the index episode; upon detected relapse, they restarted antidepressants and continued treatment for 8 months after remission before discontinuing again, consistent with guideline-recommended continuation periods in patients not pursuing maintenance treatment (Gelenberg et al., Reference Gelenberg, Freeman, Markowitz, Rosenbaum, Thase, Trivedi and Van Rhoads2010; Lam et al., Reference Lam, Kennedy, Adams, Bahji, Beaulieu, Bhat, Blier, Blumberger, Brietzke, Chakrabarty, do, Frey, Giacobbe, Gratzer, Grigoriadis, Habert, Ishrat Husain, Ismail, McGirr and Milev2024; McQuaid et al., Reference McQuaid, Buelt, Capaldi, Fuller, Issa, Lang, Hoge, Oslin, Sall, Wiechers and Williams2022; National Institute for Health and Care Excellence, 2022). Discontinuation was modeled according to the tapering practices represented in the underlying maintenance-discontinuation trials, which were heterogeneous and not consistently reported (Kishi et al., Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023).
Conceptual overview of the treatment strategies, model, and framework. (a) Comparison of continuous maintenance versus active monitoring over 5 years. Under continuous maintenance, antidepressant treatment continues throughout follow-up. Under active monitoring, antidepressants are discontinued while the patient is well and restarted after a detected relapse; treatment then continues through remission and for 8 months thereafter before discontinuation again. The interval between relapse and restart reflects the delay until relapse is detected. (b) Simplified monthly health-state transitions in the model. Each month, simulated patients may remain well, relapse from well to depressed, remain depressed, or remit from depressed to well. Continuous maintenance reduces the probability of relapse, whereas greater recurrence history increases it. Chronic residual symptoms are included in the full model but are not shown in this simplified schematic. (c) Clinical interpretation of the maintenance decision. The model uses recurrence history to estimate the medication-years required to avert 1 depression-month. This break-even quantity is compared with the medication-years a patient would consider acceptable to avert 1 depression-month, based on patient–clinician discussion of the expected consequences of depression, medication side effects, and preferences regarding depression and antidepressant use. Continuous maintenance is favored when the acceptable amount is at least as great as the required amount; otherwise, active monitoring is favored.
Figure 1. Long description
Panel A shows two horizontal timelines from 0 to 5 years. The top timeline, Continuous Maintenance, is a solid black bar indicating constant antidepressant use. The bottom timeline, Active Monitoring, shows a grey bar interrupted by a black bar. Vertical arrows indicate Relapse at month 18, Detected and restart at month 22, and Remission at month 33. The black bar continues for 8 months post-remission before returning to grey.
Panel B is a state-transition diagram. A Start box points to a Well state. A circular arrow on Well indicates Remains well. A horizontal arrow labeled Relapse points from Well to Depressed. A horizontal arrow labeled Remission points from Depressed back to Well. A circular arrow on Depressed indicates Remains depressed. Dashed arrows point to the Relapse transition: a minus sign for Continuous maintenance and a plus sign for Greater recurrence history.
Panel C features a horizontal axis from 0 to 5 labeled Medication-years per depression-month averted. On the left, Recurrence history feeds into a Model box, which outputs a Required value marked by a dot at approximately 1.4 on the axis. On the right, four factors feed into a Patient-Clinician Discussion box: plus signs for History of severe consequences and Preference to avoid depression; minus signs for Medication side effects and Preference to avoid antidepressant use. This box outputs an Acceptable value marked by a dot at approximately 3.6. A vertical dashed line at the Required dot marks the Break-even point. Below the axis, text states that if Acceptable is less than Required, Active Monitoring is favored. If Acceptable is greater than or equal to Required, Continuous Maintenance is favored, marked with a checkmark.
The model updated each patient’s depression and treatment status monthly. Relapse-detection delay depended on clinical monitoring frequency, which was quarterly in the base case; sensitivity analyses examined monthly and every-6-month monitoring. The 5-year horizon was chosen to capture longer-term trade-offs while remaining within the longest follow-up available in the source studies (Bukh et al., Reference Bukh, Andersen and Kessing2016).
Outcomes and preference thresholds
The primary outcomes were depression-months and medication-years over 5 years. We treated a month of depression (or antidepressant use) as equally important regardless of when it occurred during the 5-year horizon. Depression months were severity-weighted, with acute major depressive episodes counting fully and chronic residual symptoms contributing a lower weight derived from long-term cohort data (Judd et al., Reference Judd, Akiskal, Maser, Zeller, Endicott, Coryell, Paulus, Kunovac, Leon, Mueller, Rice and Keller1998). For each modeled recurrence-probability group, we report absolute outcomes under each strategy, between-strategy differences, and medication-years per depression-month averted, defined as the additional years taking antidepressants required to avert 1 month of depression.
We also computed the net benefit of continuous maintenance versus active monitoring across preference thresholds for antidepressant use, defined as the maximum additional medication-years a patient would accept to avert 1 depression-month. The thresholds profiled are illustrative preference parameters, not empirically measured patient preferences. They represent how much additional time taking antidepressants a hypothetical patient would consider worthwhile to avoid time depressed, given the side effects and other burdens of continued treatment. Notionally, this sort of quantity might be elicited in practice by asking a patient to recall a prior period of depression and consider how much additional time taking antidepressants while well they would accept to avoid 1 month of that experience, or through another clinically appropriate framing.
The break-even threshold for each modeled recurrence-probability group equals its medication-years per depression-month averted. For example, if maintenance requires 2 additional medication-years per depression-month averted, a hypothetical patient willing to accept at least 2 additional medication-years to avert 1 depression-month would prefer maintenance; otherwise, active monitoring would be preferred. The primary outcome is net benefit across the full range of continuous thresholds. For illustration purposes only, we highlight thresholds of 1, 2, and 3 medication-years per depression-month averted as illustrative anchors to showcase different parts of the decision curve.
Model structure and health states
At model entry, patients were in remission after completing continuation treatment. During monthly cycles, they could remain in remission, relapse, remit, or enter and later recover from chronic residual symptoms (Figure 1b; Supplementary Figure S1). The chronic residual-symptom state captures persistent, lower-severity depressive morbidity after the acute phase of some episodes and is not equivalent to persistent depressive disorder or dysthymia.
Recurrence probability groups and maintenance effect
We represented differences in recurrence probability using four clinically anchored groups: Very Low, Low, Moderate, and High modeled probability of recurrence. These groups are intended as clinically recognizable reference points rather than fixed categories or individualized predictions for a particular patient.
The Very Low and Low strata were anchored to a prospective first-lifetime depressive episode cohort stratified by baseline severity (Bukh et al., Reference Bukh, Andersen and Kessing2016): Very Low corresponded to the mild subgroup and Low to the subgroup with at least moderate baseline severity. The High stratum was anchored to a maintenance-discontinuation trial meta-analysis in recurrent depression (Kishi et al., Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023), representing populations with mean lifetime depressive episode counts of approximately 4 to 5. The Moderate stratum was interpolated between Low and High, notionally corresponding to about 2 to 3 lifetime depressive episodes.
Within each group, recurrence probability was modeled as higher closer to remission and lower later, consistent with patterns in maintenance-discontinuation trials (Kishi et al., Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023). Antidepressant maintenance reduced recurrence probability based on a meta-analysis of maintenance versus discontinuation (Zhou et al., Reference Zhou, Lv, Shi, Zhou, Lin, Chen, Wan, Li, Ran, Huang, Wang, Li, Wang, Liu and Kuang2022).
Remission after relapse
Once in an active depressive episode, patients faced monthly remission probabilities that depended on antidepressant status. Remission probabilities and antidepressant effects on remission were informed by 6-week remission rates from an acute-phase meta-analysis (Gibbons et al., Reference Gibbons, Hur, Brown, Davis and Mann2012). Because antidepressant benefit is not immediate after treatment initiation or restart, the model allowed a short delay before the full treatment effect on remission, informed by evidence on the early time course of antidepressant response (Henssler et al., Reference Henssler, Kurschus, Franklin, Bschor and Baethge2018). Full implementation details are provided in the Supplementary Material.
Chronic residual symptoms
To represent persistent depressive morbidity over longer horizons, the model allowed a subset of episodes to follow a chronic course. The frequency and persistence of these chronic-course episodes were informed by long-term episode-duration data (Spijker et al., Reference Spijker, de Graaf, Bijl, Beekman, Ormel and Nolen2002), and the relative daily burden of chronic residual symptoms compared with an acute major depressive episode was informed by a long-term cohort study (Judd et al., Reference Judd, Akiskal, Maser, Zeller, Endicott, Coryell, Paulus, Kunovac, Leon, Mueller, Rice and Keller1998).
In the base case, maintenance treatment was assumed to have half the relapse-prevention effect in chronic-course trajectories as in nonchronic episodes. Sensitivity analyses examined alternatives in which maintenance had either no relapse-prevention benefit or the same benefit as in nonchronic episodes.
Model verification and validation
We evaluated model performance by comparing model-projected outputs with published benchmarks across four domains: relapse after remission in maintenance-discontinuation trials (Kishi et al., Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023), longer-horizon cumulative relapse after a first depressive episode (Bukh et al., Reference Bukh, Andersen and Kessing2016), short-term remission after acute treatment (Gibbons et al., Reference Gibbons, Hur, Brown, Davis and Mann2012; Henssler et al., Reference Henssler, Kurschus, Franklin, Bschor and Baethge2018), and episode persistence over longer follow-up (Spijker et al., Reference Spijker, de Graaf, Bijl, Beekman, Ormel and Nolen2002). Full details are provided in the Supplementary Material.
Sensitivity analyses
Parameter uncertainty was evaluated using probabilistic sensitivity analysis. Ninety-five percent uncertainty intervals reflect variation across parameter draws. Full distributional assumptions and simulation specifications are provided in the Supplementary Material.
We also conducted univariate sensitivity analyses by varying key structural and design assumptions: chronic-relapse preventability with maintenance treatment, analytic time horizon, and monitoring frequency. For each scenario, we assessed whether strategy preference changed across modeled recurrence-probability groups and preference thresholds; scenario specifications are provided in the Supplementary Material.
Results
Verification and validation
We assessed whether the model reproduced the clinical patterns it is intended to represent by comparing model projections with published clinical targets (Figure 2). Overall, the model tracked these benchmarks closely, reproducing the key features: early post-remission relapse risk that attenuated over time, separation between maintenance and discontinuation arms, lower-risk cumulative relapse over multiple years, and persistence patterns consistent with episodic and chronic illness. For example, although the lower-risk recurrence parameters were constrained using data through 2 years after remission, the model generated a 32.0% 5-year cumulative relapse risk in the lower-risk group, compared with 31.5% (95% CI, 25.7%–37.3%) in a long-term cohort study of relapse following remission of a first lifetime depressive episode (Bukh et al., Reference Bukh, Andersen and Kessing2016). Point-by-point comparisons between published benchmarks and model projections are shown in Figure 2 and Supplementary Table S2.
Model verification and validation. Model projections (dashed lines) compared with published benchmarks (points with reported 95% CIs from the source studies) across four domains: (a) cumulative relapse over 12 months in patients with highly recurrent depression from maintenance-discontinuation trials, (b) cumulative relapse over 5 years in patients after a first lifetime depressive episode, (c) acute remission rates by time since treatment start, and (d) long-term episode persistence. Full details are provided in the Supplementary Material.
Figure 2. Long description
A multi-panel figure with four line graphs labeled A through D. A shared legend at the bottom identifies red circles as observed drug-off data, orange circles as observed drug-on data, and light red circles as observed mixed exposure. Dashed and dotted lines represent corresponding model projections.
* Panel A: High-risk relapse over 12 months. The X-axis shows months since remission (3 to 12) and the Y-axis shows cumulative relapse percentage (0 to 60 percent). Two upward-sloping trends are shown. The top dotted line (drug-off) rises from approximately 22 percent to 44 percent, closely matching four red data points with vertical error bars. The lower dashed line (drug-on) rises from 11 percent to 22 percent, matching four orange data points.
* Panel B: Lower-risk relapse over 5 years. The X-axis shows years since remission (1 to 5) and the Y-axis shows cumulative relapse percentage (0 to 50 percent). A single dashed line for mixed exposure shows a linear increase from roughly 9 percent at year 1 to 32 percent at year 5, passing through five light red data points with error bars.
* Panel C: Acute remission by time. The X-axis shows weeks since treatment start (0 to 12) and the Y-axis shows remission by time percentage (0 to 100 percent). Two logarithmic growth curves are shown. The upper dashed line (drug-on) reaches approximately 63 percent at week 12, matching orange data points. The lower dotted line (drug-off) reaches approximately 44 percent at week 12, matching red data points.
* Panel D: Long-term episode persistence. The X-axis shows months since episode onset (0 to 36) and the Y-axis shows the percentage unremitted (0 to 100 percent). A single dash-dot line shows a steep asymptotic decay, starting near 75 percent at month 2 and leveling off toward 15 percent by month 36, aligning with five light red data points.
Base case
In the base case, continuous maintenance resulted in fewer depression-months but substantially more time taking antidepressants compared with active monitoring. The magnitude of this trade-off varied markedly by modeled recurrence probability (Table 2). In the Very Low group (calibrated to patients 8 months after remission from a first lifetime mild depressive episode), continuous maintenance required 11.8 medication-years per depression-month averted. The corresponding values for the Low, Moderate, and High groups were 5.1, 2.8, and 1.5 medication-years per depression-month averted, respectively.
Base-case outcomes by recurrence-probability group
Table 2. Long description
The table presents base-case outcomes over a 5-year horizon. It is organized into seven columns: Recurrence probability group, Prior depressive episodes, Months of depression (Active monitoring vs Continuous), Years of antidepressant use (Active monitoring vs Continuous), and Efficiency of continuous strategy.
* High Recurrence Group (4 to 5 prior episodes): Active monitoring results in 6.0 months of depression and 1.2 years of medication. Continuous strategy results in 3.4 months of depression and 5 fixed years of medication. Efficiency is 1.5 additional med-years to avert 1 depression-month (most efficient).
* Moderate Recurrence Group (2 to 3 prior episodes): Active monitoring results in 3.4 months of depression and 0.7 years of medication. Continuous strategy results in 1.8 months of depression and 5 fixed years of medication. Efficiency is 2.8.
* Low Recurrence Group (1 moderate plus episode): Active monitoring results in 1.9 months of depression and 0.4 years of medication. Continuous strategy results in 1.0 months of depression and 5 fixed years of medication. Efficiency is 5.1.
* Very Low Recurrence Group (1 mild episode): Active monitoring results in 0.9 months of depression and 0.2 years of medication. Continuous strategy results in 0.4 months of depression and 5 fixed years of medication. Efficiency is 11.8 (least efficient).
All values for months and years include 95 percent uncertainty intervals in parentheses. Continuous medication use is fixed at 5 years for all groups.
Note: Values are means (95% uncertainty intervals) from probabilistic sensitivity analysis (1,000 parameter draws × 10,000 patients per recurrence-probability group).
a Approximate number of prior depressive episodes characterizing each recurrence-probability group. Very Low and Low were calibrated to a first-episode cohort stratified by baseline severity (Bukh et al., Reference Bukh, Andersen and Kessing2016); High was calibrated to maintenance-discontinuation trial populations with recurrent depression (Kishi et al., Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023); Moderate was interpolated.
b Severity-weighted depression-months. Acute major depressive episodes count fully; chronic residual symptoms are weighted at 0.44 per day relative to acute episodes.
c Additional medication-years of antidepressant use under the continuous strategy per severity-weighted depression-month averted, relative to the active monitoring strategy. Lower values indicate a more efficient trade-off. Technical note: Values in the final column were calculated separately within each probabilistic sensitivity analysis draw and then averaged across draws; because the resulting ratios are skewed, they are not exactly equal to ratios calculated from the mean values in the preceding columns.
Most of the additional depression burden under active monitoring resulted from episodes that were not prevented (82.3% of additional days depressed, averaged across recurrence-probability groups), with smaller contributions from longer episode duration (9.0%) and the combination of these effects (8.7%). Chronic residual symptoms accounted for the majority of depression burden under both strategies (63.8% under continuous maintenance; 53.1% under active monitoring).
Incremental net benefit analysis
Figure 3a presents the net benefit of continuous maintenance versus active monitoring, expressed in depression-months averted over 5 years, across varying preference thresholds for antidepressant use. Results are summarized at three illustrative thresholds (1, 2, and 3 medication-years per depression-month averted).
Net benefit and decision certainty for continuous maintenance versus active monitoring by preference threshold for antidepressant use. (a) Net benefit expressed as depression-months averted over 5 years. Blue shading indicates the region where continuous maintenance is preferred; warm shading indicates where active monitoring is preferred. Solid lines show mean net benefit across probabilistic sensitivity analysis draws; shaded bands show 95% uncertainty intervals reflecting between-draw parameter uncertainty. (b) Probability that continuous maintenance is the preferred strategy at each preference threshold for antidepressant use. Vertical dotted lines in both panels mark the three illustrative preference thresholds (1, 2, and 3 medication-years per depression-month averted).
Figure 3. Long description
Panel A is a line graph showing net benefit. The x-axis is Antidepressant acceptability threshold (medication-years per depression-month averted) from 0 to 10. The y-axis is Net benefit of continuous maintenance (depression-months averted; log sub 2 scale) from negative 16 to 8. Four curves represent recurrence risk groups: Very low (yellow), Low (orange), Moderate (red), and High (dark maroon). All curves show an asymptotic increase as the threshold increases. The High risk group crosses the 0 net benefit line earliest (threshold approximately 1.5), while the Very low risk group remains below 0 for the entire range. Shaded bands around each line represent 95 percent uncertainty intervals. The background is split horizontally at y equals 0: the top half is blue (Continuous preferred) and the bottom half is tan (Active monitoring preferred).
Panel B is a line graph showing probability. The x-axis is the same as Panel A. The y-axis is Probability continuous maintenance is preferred (percent) from 0 to 100. Four S-shaped curves correspond to the same risk groups. The High risk curve rises most steeply, reaching 100 percent probability near a threshold of 4. The Moderate, Low, and Very low curves are shifted progressively to the right. A horizontal line at 50 percent divides the graph into a blue top region (Continuous more likely preferred) and a tan bottom region (Active monitoring more likely preferred). Vertical dotted lines at thresholds 1, 2, and 3 intersect both panels.
At the lowest preference threshold (1 medication-year per depression-month averted), continuous maintenance produced net harm across all risk groups. At the intermediate threshold (2 medication-years per depression-month averted), maintenance was preferred only in the High group. At the highest threshold (3 medication-years per depression-month averted), maintenance was preferred in the High group and reached the break-even point in the Moderate group, although uncertainty in the Moderate group overlapped zero. Thus, even when patients strongly prioritized avoiding depression, expected gains from maintenance concentrated in the Moderate and High groups.
Figure 3b shows the probability that maintenance was preferred after accounting for parameter uncertainty; this probability increased with modeled recurrence probability and willingness to accept additional time taking antidepressants.
Univariate sensitivity analyses
To evaluate robustness, we examined 6 univariate sensitivity scenarios varying clinical monitoring frequency, analytic time horizon, and the preventability of chronic depression with maintenance treatment (Supplementary Table S3). We classified the base-case preferred strategy for each combination of modeled recurrence-probability group and preference threshold, then assessed how often these classifications changed across scenarios. Overall, 68 of 72 (94%) cell-level classifications were unchanged. The few shifts occurred near break-even regions and moved in expected directions: modeling chronic depression as more preventable and extending the time horizon favored continuous maintenance more strongly, whereas more frequent monitoring somewhat improved the favorability of active monitoring. The largest shifts arose in the no chronic-prevention-benefit scenario, in which, at 2 medication-years per depression-month averted, no group favored maintenance, whereas at 3 medication-years per depression-month averted, only the High group favored maintenance. Overall, these findings suggest that the model’s conclusions are robust, with strategy preference depending more strongly on patient preferences and modeled recurrence probability than on the structural assumptions varied here.
Discussion
In this decision-analytic model of adults who had completed an 8-month continuation phase after remission from MDD, continuous maintenance reduced depression burden over 5 years but required substantially more time taking antidepressants than active monitoring. The efficiency of this trade-off increased markedly with modeled recurrence probability: averting 1 month of depression required roughly 1.5 additional years of antidepressant use in the High recurrence-probability group, compared with more than a decade in the Very Low recurrence-probability group.
These findings are broadly consistent with current guideline logic (Gelenberg et al., Reference Gelenberg, Freeman, Markowitz, Rosenbaum, Thase, Trivedi and Van Rhoads2010; Lam et al., Reference Lam, Kennedy, Adams, Bahji, Beaulieu, Bhat, Blier, Blumberger, Brietzke, Chakrabarty, do, Frey, Giacobbe, Gratzer, Grigoriadis, Habert, Ishrat Husain, Ismail, McGirr and Milev2024; McQuaid et al., Reference McQuaid, Buelt, Capaldi, Fuller, Issa, Lang, Hoge, Oslin, Sall, Wiechers and Williams2022; National Institute for Health and Care Excellence, 2022): maintenance becomes more compelling as modeled recurrence probability rises, but the preferred strategy also depends on how much time a patient is willing to take antidepressants to avert future depression.
The framework may be particularly useful in two settings. First, in the Moderate group, notionally corresponding to patients with about 2 to 3 prior episodes, our analysis identified a break-even preference threshold of approximately 2.8 medication-years per depression-month averted. Because current guidance is less determinate for these patients, this threshold makes explicit how much willingness to continue antidepressants is needed for maintenance to be favored.
Second, the framework can inform decisions when patient preferences conflict with the usual recommendation at either end of the recurrence-probability spectrum. According to our model, maintenance may be a reasonable choice even after a single moderate-to-severe depressive episode if a patient places a sufficiently high value on avoiding future depression relative to the burden of continued antidepressant use. Conversely, active monitoring may remain reasonable even in highly recurrent depression if the burden of continued treatment exceeds what the patient is willing to accept. In the Low and High groups, the corresponding break-even thresholds were approximately 5.1 and 1.5 medication-years per depression-month averted, respectively.
The preference threshold represents how much additional time a patient would be willing to take antidepressants to avert a given amount of depression. That willingness may reflect how burdensome another depressive episode would be expected to be for the patient and how burdensome continued antidepressant use is to them. Prior suicidality, hospitalization, severe functional impairment, prolonged symptoms, or difficulty achieving remission may increase the value a patient places on avoiding depression, but such features are not required; a patient with an otherwise typical prior episode may simply strongly prefer not to experience depression again. Likewise, willingness to continue antidepressants may be high when treatment is well tolerated, but it may be lower because of side effects or because the patient simply prefers not to remain on medication or dependent on continued treatment. In practice, these preferences would likely need to be discussed and clarified with the patient rather than treated as a single unexamined stated preference.
To make this logic concrete, consider a 26-year-old woman with 2 years remaining in graduate school who is several months past remission from a first lifetime moderately severe depressive episode (Low recurrence probability group). After discussing the trade-offs with her clinician, she might conclude that she would take antidepressants for the remainder of graduate school to avoid another month of depression, but would not want to remain on them throughout her 20s for that same benefit. Her implied preference threshold is therefore between approximately 2 and 4 medication-years per depression-month averted. Because this falls below the Low-group break-even threshold of about 5, active monitoring would be favored.
Two variants of this example are worth considering for contrast. If, a decade later, she had experienced several relapses and resembled the High recurrence-probability group, the same underlying preference would favor continuous maintenance. Alternatively, suppose her single prior episode had included a suicide attempt, but she had since recovered her insight and capacity for treatment decisions. In that case, it would be consistent with the framework for the clinician to explore whether her initially stated preference fully reflects what another depressive episode could mean for her, including the possibility of similarly serious consequences. It would also be consistent with the framework to caution that the lower-probability group is only a reference point rather than an individualized estimate, so a close decision should not be over-interpreted if she differs importantly from the cohort used to anchor that group.
The modeled recurrence-probability groups are intended as clinically recognizable reference points rather than individualized predictions for particular patients. Population data suggest that the corresponding recurrence-history patterns are common: in the NCS-R, among adults with a major depressive episode in the prior year, 14.8% reported 1 lifetime episode, 24.6% reported 2 to 3, and 60.6% reported 4 or more (see Supplementary A3a). Patients will not necessarily correspond exactly to one group, and recurrence probability in practice may depend on factors beyond episode history that are not represented in the model. Even when a patient’s exact recurrence probability is uncertain, recognizing that the patient is closer to the Moderate or High benchmark than to the Low benchmark may be sufficient to inform shared decision-making. Future work could replace these group-based benchmarks with validated individualized estimates of recurrence probability.
The large contribution of chronic residual symptoms also highlights that recurrence probability is only one part of the maintenance decision. The model does not estimate which individual patients are more likely to experience prolonged, severe, or functionally consequential recurrences; instead, it applies average assumptions about depressive course once recurrence occurs. These expected consequences can nevertheless enter the decision through the patient’s preference threshold because a patient who expects another episode to be especially burdensome may place greater value on avoiding it. Future work could improve the framework by incorporating individualized estimates not only of recurrence probability but also of likely episode course and consequences.
This study has several limitations. The active-monitoring strategy inherits important limitations of the maintenance-discontinuation evidence base. Tapering practices in the underlying trials were heterogeneous and often incompletely reported (Kishi et al., Reference Kishi, Sakuma, Hatano, Okuya, Matsuda, Kato and Iwata2023), and there is some discussion in the literature that very early symptoms after antidepressant discontinuation may sometimes reflect withdrawal rather than true recurrence (Ghaemi & Selker, Reference Ghaemi and Selker2017; Hengartner & Plöderl, Reference Hengartner and Plöderl2021; Horowitz & Taylor, Reference Horowitz and Taylor2022). Contemporary gradual tapering, particularly when combined with relapse-prevention psychotherapy, may therefore make discontinuation more favorable than represented in our base case (Zaccoletti et al., Reference Zaccoletti, Mosconi, Gastaldon, Benedetti, Gottardi, Papola, Ponzi, Purgato, Naudet, Cristea, Barbui and Ostuzzi2026). In addition, remission after antidepressant restart was parameterized using general acute-treatment evidence rather than evidence specific to re-initiation of a previously effective and tolerated antidepressant, although what limited evidence is available regarding re-initiation does suggest broadly similar efficacy to that observed in acute-treatment samples (Fava et al., Reference Fava, Detke, Balestrieri, Wang, Raskin and Perahia2006).
The preference thresholds are hypothetical quantities rather than empirically measured patient preferences. Their distribution, stability, and optimal methods of elicitation and communication require empirical and stakeholder-engaged study. The model also isolates the antidepressant maintenance-versus-monitoring decision rather than representing the full range of long-term depression care. It does not quantitatively estimate individual differences in the expected severity, duration, or functional consequences of a future episode, nor does it incorporate comorbidities, adherence patterns, or changing clinical circumstances. Finally, the model combines parameters from randomized trials, meta-analyses, and observational cohorts that differed in patient selection, treatment context, outcome definitions, and follow-up; while in doing this we attempt to provide the best synthesis of all available data, this heterogeneity limits the precision and transportability of the estimates. The framework is therefore intended to inform, rather than replace, clinical judgment and shared decision-making.
Future work should test decision-support tools that combine individualized estimates of recurrence probability with patient preferences to guide maintenance decisions in real-world care. This will require stakeholder-engaged development of methods for eliciting and communicating preferences, as well as empirical study of their stability and clinical usefulness. Future models could also incorporate individualized estimates of the expected course and consequences of recurrence, rather than relying primarily on recurrence probability alone.
In this decision-analytic study, continuous maintenance reduced depression burden compared with active monitoring, but the efficiency of the trade-off between additional antidepressant use and depression-months averted varied substantially with modeled recurrence probability. This work provides a quantitative framework for maintenance decisions by making explicit how recurrence probability interacts with patient preferences regarding the relative burden of depression and continued antidepressant use.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291726106321.
Data availability statement
This decision-analytic modeling study used model inputs derived from published aggregate data. The underlying data sources are cited in the manuscript and Supplementary Material. No individual participant data were analyzed.
Author contribution
W.U.M. had full access to all model inputs, analytic code, and model outputs and takes responsibility for the integrity of the data, accuracy of the analysis, and fidelity of the manuscript to the reported results. W.U.M. conceived and designed the study, conducted the statistical analysis, and drafted the manuscript. All authors contributed to the acquisition, analysis, or interpretation of data and critically revised the manuscript for important intellectual content. J.W.S. provided administrative, technical, or material support. T.C. and J.W.S. supervised the study.
Funding statement
This study received no dedicated external funding. W.U.M. was supported by the National Library of Medicine, National Institutes of Health, under training grant T15LM007092. The funder had no role in the design, conduct, analysis, interpretation, or reporting of the study.
Competing interests
J.W.S. is a member of the Scientific Advisory Board of Sensorium Therapeutics and holds options in the company, has received consulting fees from Tempus, Inc., and has received grant support from Biogen, Inc. W.U.M. and T.C. declare no competing interests.
Ethical standard
This study used model inputs derived exclusively from published aggregate data and did not involve individual participant data. The Mass General Brigham Institutional Review Board determined that this study did not constitute human-subjects research.
Use of artificial intelligence
OpenAI ChatGPT GPT-5 Thinking-series models (OpenAI, San Francisco, CA) and Anthropic Claude Opus and Sonnet 4-series models (Anthropic, San Francisco, CA) were used intermittently from October 2025 through September 2026 through the publicly available ChatGPT and Claude interfaces. They were used in an assistive role during study development and manuscript preparation, including code refactoring and editing, discussion of analytic implementation details, and drafting, revising, and editing portions of the manuscript. The study concept, design, core analytic decisions, original code framework, interpretation of results, and final manuscript content were determined and verified by the authors. All AI-assisted output was reviewed, revised as needed, and checked by the authors, who take full responsibility for the accuracy and integrity of the work.
Open access
