Psilocybin Therapy for Treatment Resistant Depression: Prediction of Clinical Outcome by Natural Language Processing

Major depressive disorder (MDD) is highly prevalent and economically burdensome, and around one-third of people with MDD do not adequately respond to treatment, a condition labelled treatment-resistant depression (TRD). Earlier randomised trials of psilocybin-assisted therapy (PAT) demonstrated antidepressant effects: examples cited include a 6-week trial (N = 104) where a single 25-mg dose produced sustained remission in 25% versus 9.1% with active placebo, a smaller trial reporting a 2-week remission of 54% versus 12% in controls, and an RCT in TRD (N = 79) reporting a three-week remission rate of 29% compared to 13.7% for standard third-line antidepressants. Despite promising clinical signals, PAT requires substantial upfront resources for preparation, dosing and integration with trained professionals, and its comparative economic value versus existing third-line options is unclear. This study by Avanceña and colleagues set out to perform an exploratory, model-based economic evaluation of single-dose PAT as a third-line treatment for adults with TRD in the US. The investigators aimed to estimate costs, health outcomes (quality-adjusted life years, QALYs) and incremental cost-effectiveness, and to identify key determinants of PAT's value such as treatment cost and durability of effect. The analysis is presented as the first US-focused economic assessment of PAT to inform trial design, pricing and coverage discussions should PAT be approved.

The researchers built an individual-level, probabilistic decision-analytic simulation to compare single-dose PAT versus a population-weighted “standard of care” (SOC) mix of third-line treatments for TRD over a 12-month horizon. SOC was modelled predominantly as pharmacotherapy (more than 99%), often with psychotherapy, and included small proportions (<1%) receiving electroconvulsive therapy (ECT) or esketamine nasal spray; individuals who fail third-line options were assumed to receive pharmacotherapy for subsequent lines. The analysis used limited healthcare and limited societal perspectives, reported costs in 2022 US$, and did not discount outcomes because the time horizon was one year. Model structure comprised four mutually exclusive weekly health states: MDD, response, remission and death. TRD was defined as failure to respond to two prior MDD treatments. Transition probabilities for remission, response and relapse were drawn from published sources: STAR*D for pharmacotherapy/psychotherapy, a meta-analysis for ECT, esketamine trial data (noting a roughly 30% higher remission/response when combined with antidepressants) and a phase 2 RCT for PAT (single 25-mg dose with intensive psychosocial support: remission 29%, response 8% in that trial). Where needed, probabilities reported over other time intervals were converted to weekly transition probabilities. The authors acknowledged limited long-term PAT data and assumed PAT relapse rates equivalent to third-line therapies in STAR*D, noting this may overestimate relapse. Cost inputs were compiled from administrative claims analyses and prior costing studies. In the base case the total cost of a single PAT session was assumed to be $5000 (decomposed as $3500 for therapists and $1500 for psilocybin) and was deterministically varied from $3000 to $20,000 to explore sensitivity. The limited societal perspective added productivity losses (absenteeism and presenteeism) adjusted to 2022 earnings. Health utilities for QALY calculation came from a systematic review of MDD utilities measured by time-tradeoff methods and were transformed to the model’s weekly cycle length. To characterise parameter uncertainty, the investigators ran probabilistic analyses: 1000 independent model runs, each simulating 1000 individuals, with inputs sampled from assigned probability distributions to produce mean costs, mean QALYs and 95% predicted intervals. The primary economic metric was the incremental cost-effectiveness ratio (ICER), evaluated against a US-relevant threshold range of $100,000–150,000 per QALY. Conditional probabilistic analyses varied three key inputs deterministically—PAT remission probability, PAT relapse probability (each by ±5%, ±10% and ±15%), and PAT cost ($3000–$10,000)—while sampling all other parameters probabilistically. Results were summarised with cost-effectiveness acceptability curves and expected loss curves to quantify decision uncertainty and the potential value of further research.

In the base case assuming PAT cost of $5000, PAT produced an incremental mean cost of $3,638 per person from the limited healthcare perspective and an incremental mean benefit of 0.031 QALYs over 12 months (equivalent to 31 QALYs per 1000 individuals), reflecting an approximately 9% higher overall remission rate versus SOC. The expected ICER was $117,517 per QALY from the limited healthcare perspective and $111,069 per QALY from the limited societal perspective. The probabilistic analysis produced 95% predicted intervals around outcomes (the paper reports that these were constructed from the 1000 model runs, but specific interval bounds for the main outcomes are not provided in the extracted text). Cost-sensitivity was a dominant driver. PAT met common US cost-effectiveness thresholds when its per-treatment cost was $5000 or less. At a PAT cost of $10,000 the expected ICERs rose to $275,426 (limited healthcare) and $268,978 (limited societal), and the probability that PAT was cost-effective at a $150,000 threshold fell to 1%. Across thresholds, the probability that PAT was cost-effective increased from 31% at $100,000 to 74% at $150,000 under the $5000 cost assumption. Expected loss analysis at the $150,000 threshold favoured PAT: expected loss per person was $301 for PAT versus $1,307 for SOC. Conditional probabilistic analyses showed that changes in PAT effectiveness parameters shifted cost-effectiveness probabilities. Increasing the probability of remission by 15% raised PAT’s probability of being cost-effective from 74% to 89% (limited healthcare perspective, $5000 cost). Reducing the relapse probability by 15% increased that probability from 74% to 85%. When both remission increased by 15% and relapse decreased by 15%, PAT’s probability of being cost-effective at a $10,000 cost rose from 1% to 34%. Similar patterns were observed under the limited societal perspective.

Avanceña and colleagues interpret their modelled findings to mean that one-time, single-dose PAT could be a cost-effective third-line treatment for TRD in the US, but only under specific conditions—chiefly when per-treatment costs are relatively low (around $5,000 or less) or when effectiveness and durability of remission are materially better than in available trials. Cost was identified as the most influential determinant of economic value, though greater and more sustained remission also meaningfully improves cost-effectiveness. The authors compare their results to a prior UK analysis (McCrone et al.), which also found QALY gains for PAT but not cost-effectiveness from a UK healthcare perspective; differences between the studies arise from model structure (individual-level state-transition here versus deterministic decision-tree in the UK study), longer 12-month horizon, a defined TRD population, and explicit probabilistic uncertainty analysis. Practical strategies to reduce PAT costs are discussed, including group-based psychological support or alternative provider types, but the investigators note that the efficacy of lower-cost delivery models requires empirical testing. They also highlight multiple market and regulatory factors that could influence the eventual price of psilocybin. Several limitations are acknowledged. The 12-month horizon was chosen because long-term PAT effectiveness data are limited; exclusion of other TRD therapies (for example, repetitive transcranial magnetic stimulation and intravenous ketamine) may alter relative value; and the analysis focused narrowly on single-dose PAT as a third-line option rather than other lines or repeat/maintenance dosing. The model assumed universal uptake of PAT among the simulated cohort, did not incorporate subgroup heterogeneity beyond age- and sex-specific mortality adjustments, and used 2020 life tables which may be affected by the COVID-19 period. The authors also note the restricted perspectives used: they omitted caregiver, transportation and patient time costs and therefore did not capture the full societal consequences, including caregiver burden. Finally, equity and access concerns are emphasised: disparities in mental health treatment uptake mean distributional cost-effectiveness work is needed to understand how PAT’s introduction could affect vulnerable subgroups and health inequities.

The investigators conclude that single-dose PAT shows clinical effectiveness for MDD and TRD and has the potential to be cost-effective as a third-line therapy in the US if per-treatment costs are approximately $5,000 or lower. They recommend that future trials and implementation work examine strategies to reduce PAT costs and gather longer-term effectiveness data to better characterise its sustained benefits and value.