Study data were collected during Mindbloom’s clinical operations, a US-based direct-to-consumer organization specializing in ketamine therapy. Treatment is offered through a telehealth platform across 38 states, connecting patients with autonomous, licensed psychiatric clinicians. The Mindbloom platform can be accessed via an internet search or referral from an external physician/provider, while access to treatment through the platform is gated by rigorous screening from psychiatric clinicians.

Patients in the analysis were a subset of Mindbloom patients seeking treatment for elevated symptoms of depression, anxiety, or PTSD, defined as a Patient Health Questionnaire-9 (PHQ-9) or Generalized Anxiety Disorder-7 Scale (GAD-7) score of ≥10, or a PTSD Checklist for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (PCL-5) score of ≥33. Inclusion and exclusion criteria for treatment in the Mindbloom program are listed in Textbox 1.

In addition to the criteria in Textbox 1, patients with any of the following were evaluated for eligibility on a case-by-case basis and required to provide confirmation of concurrent, external therapeutic support to be considered for treatment: active moderate to severe alcohol or other substance use disorder, active moderate to severe opiate use disorder, and a history of severe trauma. In certain circumstances, additional medical clearance from outside providers and/or coordination with external mental health patients was required to ensure a patient was fit to proceed with treatment.

If a patient met initial eligibility requirements, they were invited to provide additional intake information. Patients filled out the new patient onboarding form through a patient portal, accessed via the web- and app-based Mindbloom platform, which included exclusion criteria screening tools and helped to prepare a clinician for an initial consultation. Clinicians assessed patient appropriateness for treatment using intake information, self-report assessments, a synchronous video consultation, and review of state prescription drug monitoring programs.

Patients were aged 44.7 (10.8) years on average, and 52.4% (1592/3041) of the sample was female (1438/3041, 47.3% male, 8/3041, 0.3% “Prefer not to answer,” and 3/3041, 0.1% “Other”). The sample was 43.5% (1323/3041) married/partnered. Additionally, 71.8% (2183/3041) reported having a prior diagnosis of a psychological disorder, and 91.6% (2786/3041) reported receiving some form of prior treatment for a mental illness. Due to the retrospective nature of the study, some demographic information, such as ethnicity, was not available, as these variables were not collected in the regular course of care.

This study was a retrospective cohort study based on routinely collected clinical data from patients receiving telehealth-supported at-home subcutaneous ketamine treatment through the Mindbloom platform.

Prior to treatment, patients and clinicians gave passive consent by agreeing to user terms and conditions that included the use of deidentified, aggregated data for research purposes. Data were deidentified prior to analysis to ensure patient anonymity in accordance with HIPAA (Health Insurance Portability and Accountability Act) standards, and no identifiable information or images of individual participants are included in this manuscript. No compensation was provided to patients for the use of their clinical data. The analysis in this paper constituted a retrospective analysis of Mindbloom’s routinely collected clinical data, and was deemed institutional review board (IRB) exempt by Biomedical Research Alliance of New York (BRANY) IRB (BRANY File #25-12-706-2414). Patients completed informed consent for treatment. However, as this was formally determined by the IRB to be exempt from IRB review, the study therefore did not require informed consent for research participation, in compliance with 45 CFR part 46 Category 4: Common Rule, Secondary research, for which consent is not required.

Clinicians had current prescription privileges in their state of licensure, completed a 30-hour training on Mindbloom's protocols, and passed an evaluation of their psychiatric evaluation skills with a standardized patient. Regular chart reviews were also conducted, as well as clinical quality metric monitoring to identify and provide coaching opportunities. Weekly ongoing training was provided for advanced skills.

Each patient was assigned a behavioral coach, known as a “guide,” to provide coaching, patient education, and logistical support to facilitate treatment adherence and serve as a communication bridge to the clinical team. Guides were required to have a coaching certification and/or to have provided one-on-one behavioral coaching for more than 1 year with at least 50 clients.

Side effects were assessed through a single-item self-report measure administered after session 2 and again after session 4: “Have you noticed any issues with your physical or mental health since beginning treatment?” Response options included chest pain, lower abdominal pain, increased BP, shortness of breath, cravings for ketamine, memory loss, pain urinating, suicidal thoughts, none of the above, and a field to specify anything else. Any adverse events reported to the clinician or guide outside of this measure were recorded in the electronic health record.

To account for missing data patterns and isolate the specific relationship between the subcutaneous treatment protocol and symptom reduction, several level-2 person-level constant variables were operationalized as statistical covariates in the primary analyses. These included demographic characteristics (baseline age and biological sex), clinical history (presence of a prior psychiatric diagnosis), and treatment parameters (final titrated medication dosage). Including these covariates ensured that the longitudinal symptom trajectories were adjusted for baseline imbalances and potential attrition biases.

First, we calculated descriptive statistics (eg, mean, SD, median, and range) at baseline, assessment wave 1 (after 2 sessions), assessment wave 2 (after 4 sessions), and assessment wave 3 (after 6 sessions) for the primary outcomes of interest (PHQ-9, GAD-7, and PCL-5) as well as number of side effects at each assessment wave. No transformations or filtering of data points beyond what was stated in the participant flow diagram/inclusion criteria were performed. Second, we computed the percent change in each of the primary outcomes, relative to baseline, and ran paired t tests to compare each percent difference to baseline. Finally, we built random-intercept mixed models to control for the nonlinear effect of time (days) from baseline. Models included a cubic spline term (df=3) for time and level-1 (time-varying) terms for assessment wave, as well as level-2 (person-level constant) terms for age (grand-mean centered), sex, final dosage (grand-mean centered), and history of a psychiatric diagnosis. These covariates were included to both account for missingness and to rule out between-person demographic and dosage effects. Spline models significantly improved fit compared to linear-time models (χ²₂>10.22; P<.007). By including the assessment wave in addition to the spline term for days, the model thus estimates a nonlinear change in outcome (days) as well as a linear change (assessment wave). Models were estimated for the 3 primary outcome variables: PHQ-9, GAD-7, and PCL-5. From these models, we estimated marginal means for each primary outcome at the median number of days for each assessment wave. This approach helps to account for the high variability in the number of days between assessment waves.

A final sample of 3041 patients taking subcutaneous ketamine injections was included in this analysis (see Figure 2 for the participant flow diagram).

We calculated the improvement in raw scores and percent change from baseline (Table 2). In this table, we calculated the number of people who experienced a minimal clinically important difference (MCID) in response to treatment in PHQ-9, GAD-7, and PCL-5. MCIDs were defined as improvements of 4 for PHQ-9 and GAD-7, and 5 or more points for the PCL-5.

Figure 3 shows the distribution of dose amount at baseline compared with session 6. Patients began the study at a median dose of 40.0 (IQR 10) mg and ended at a median dose of 70.0 (IQR 30) mg. This IQR calculation excludes the 0.4% of patients who transitioned from subcutaneous to sublingual administration prior to study completion.

A total of 93 (3.1%) people reported side effects after session 2 (wave 1 of surveys), 76 (3.2%) people reported side effects after session 4 (wave 2 of surveys), and 41 (2.8%) people reported side effects after session 6 (wave 3 of surveys). The most common side effects at each assessment wave were as follows: lower abdominal pain (19/93 after session 2), lower abdominal pain (12/76 after session 4), and memory loss (14/41 after session 6). There was no association between dosage and risk for experiencing a side effect at either of these assessment waves (P>.17). Additionally, only 19 (0.6%) patients switched from subcutaneous to sublingual administration of ketamine by the conclusion of the study, and only 6 (0.2%) reported side effects from the injections themselves (eg, pain, swelling, and discomfort).

A total of 3 SAEs were reported during the 22-month reporting period of the study, for an SAE rate of 0.08% (3/3943). One patient experienced a psychotic episode following their first treatment and was determined to be likely related to ketamine treatment. Two patients died by suicide during treatment—one following their first treatment, and one following their second treatment—but the attributability of the deaths to treatment could not be determined.

Separate models were run for PHQ-9, GAD-7, and PCL-5 scores, and each included random intercepts. All models controlled for nonlinear (spline) effects of time (days), since there were expected to be nonlinear and meaningful effects of time between each session. We also included the effect of assessment waves 1, 2, and 3, which occurred after sessions 2, 4, and 6, respectively. The full mixed model results are presented in Tables S1-S3 in Multimedia Appendix 1. Models showed that the assessment wave was significantly associated with consistent improvements in PHQ-9, GAD-7, and PCL-5 scores (see Table 3 for marginal means and Figure 4 for estimated marginal means over time). We also calculated cumulative within-person standardized differences in each outcome as the mean paired difference divided by the SD of each paired difference (ie, Cohen d–like effect sizes for mixed models, relative to baseline). Results indicate that effects were large and increased modestly across each assessment wave. CIs indicate relatively little variability in treatment outcomes, suggesting that the means are representative of what the majority of patients experienced.

With respect to covariate effects, history of a psychological disorder was associated with higher overall scores on all 3 measures (P<.002) and final dose was associated weakly with overall higher PHQ-9 (b=.003; P=.02) and GAD-7 scores (b=.003; P=.01). Additionally, age predicted lower overall GAD-7 scores (b=-0.03; P<.001) and females reported higher PHQ-9 scores (b=.48; P=.007).

With respect to attrition (ie, declining to complete surveys at future assessment waves), we found that female patients were slightly more likely to drop out (nonstatistically significant after 6 treatment sessions; 290/540, 53.8% of assessment wave 3 missing participants were female; P=.20). Reporting a psychological disorder at baseline was also associated with a slightly higher chance of attrition. Specifically, 73.9% (399/540) of those missing self-reports after 6 treatment sessions reported a disorder compared to 71.3% (1036/1452) of patients with a complete dataset across all assessment waves; P=.02. Given these associations with missingness, we can cautiously infer missingness at random; we included biological sex and previously reported psychological disorder diagnosis in our mixed models as covariates to account for this, as well as final dose and age, to rule out between-person effects of age and dosage. This statistical approach is consistent with established methodologies [31,38], and our specific analysis was modeled after the framework established by Hull et al [31]. Little missing completely at random (MCAR) test was run to assess if data were missing completely at random. The test revealed that outcome data were likely not MCAR (χ²₂₄=389; P<.001). From these analyses, we cautiously conclude that the data are missing at random, but not MCAR.

To address concerns about bias introduced by missing data, we completed a more rigorous multilevel multiple imputation on our full dataset. The imputed data yielded quite similar results as the original results (see Table S4 in Multimedia Appendix 1 for the sensitivity analysis of descriptive statistics and clinical outcomes using multiple imputation). For further assurance, we conducted a “worst-case imputation” by carrying the last observation forward to all missing cases within a person (ie, last observation carried forward). This approach—which assumes that all patients who dropped out did so because they failed to improve and would not have improved had they stayed—yielded results that were similar to the primary model estimates, though, as expected, the last observation carried forward results showed that participants responded less over time.

Given that the results did not change substantially in either of these increasingly conservative sensitivity analyses, the original analysis was retained to allow for the ease of interpretation that comes with reporting raw means.

Consistent with our primary aims, this large-scale retrospective study supports our core hypotheses regarding the implementation of at-home subcutaneous ketamine therapy and the potential clinical impact of a telehealth-mediated subcutaneous ketamine protocol in a large, real-world cohort. First, the protocol demonstrated a high safety ceiling characterized by a low incidence of adverse effects and an exceptionally low rate of serious complications. Second, the mode of administration was associated with rapid, statistically significant, and clinically meaningful symptom reductions across all 3 evaluated psychological domains: depression, anxiety, and trauma-related distress. Finally, the secondary feasibility hypothesis was validated by robust treatment completion rates, high program adherence, and minimal patient-initiated transitions to alternative administration routes. Taken together, these findings indicate that a supervised, remote subcutaneous ketamine treatment can deliver rapid clinical improvements comparable to traditional clinic-bound therapies while maintaining an acceptable safety profile.

The safety and tolerability profile observed in this study represents a meaningful contribution to the clinical management of treatment-resistant mood conditions. The low rate for side effects stands in contrast to standard-of-care pharmacological alternatives [56]; although an exact side-by-side comparison cannot be made, it is difficult to ignore that the reported value for the former is orders of magnitude lower than the latter. The high degree of tolerability likely reduced barriers to follow-through, as only a minority of patients choose to transition to the sublingual option. The low adverse event rates indicate that remote risk can be effectively mitigated when patient self-administration training is standardized through a combination of well-established best practices [48-50], including comprehensive video instruction, a verified live demonstration, a mandatory adult monitor, and active clinician-directed titration rules.

Crucially, subcutaneous administration was shown to be associated with rapid and clinically significant symptom reduction. The majority of both depression and anxiety patients achieved an MCID by the first assessment wave, and treatment response by the second assessment wave, suggesting that the observed benefits of subcutaneous ketamine were not only statistically significant but large enough to be meaningfully felt by participants [51,52,55]. The symptom improvement observed in the PTSD cohort is particularly noteworthy, as this population often faces some of the highest rates of treatment resistance and dropout in traditional psychotherapy and pharmacological models [15]. Further, symptom reduction effects were evident early in the treatment course, highlighting the rapid onset of action observed in the sample. In a clinical landscape where standard-of-care antidepressants often require 1-2 months to show effectiveness [19], an accelerated window for meaningful improvement could be impactful for patients in acute distress.

Finally, the results indicate that self-administered subcutaneous injections are a feasible modality for remote, at-home clinical settings. While sublingual tablets are an appealing option for some patients, prior literature notes high variability in ketamine effects due to inconsistent sublingual absorption and discomfort with sublingual administration [43]. Our findings demonstrate that patients can use higher-bioavailability subcutaneous administration [44] at home with a low incidence of administration challenges or side effects and high rates of adherence. Given that patients who used subcutaneous ketamine at home largely did so without issue, these data align with the broader literature [45,46] to suggest that subcutaneous has the potential to provide a viable option alongside other modes of administration. The existence of multiple effective at-home approaches further validates the overall feasibility of the at-home telehealth treatment model [12].

The analysis represents what is, to our knowledge, the first published study containing data on at-home subcutaneous ketamine use for depression, anxiety, and PTSD. Using a massive, nationwide cohort that spans the majority of US states and includes robust samples of multiple psychiatric disorders, these data, while limited by their observational nature, provide valuable insight into the safety, feasibility, and observed clinical outcomes associated with subcutaneous ketamine delivered at home.

This study is a retrospective, observational cohort study and lacks a comparison arm with which to make any claims of superiority over other treatments. There are inherent limitations to using real-world data, where patients navigate care freely, and provide data only while an active patients. We were unable to analyze the impact of ethnicity, as collecting this variable was not a part of the clinic’s standard workflow; without these data points, we cannot evaluate potential group-level differences in the speed or size of treatment response, limiting generalizability. Lack of data on patients who leave treatment means higher attrition than a controlled trial, limiting our ability to assess potential biases introduced by attrition that may or may not be random. Another important limitation relates to safety; while SAEs were infrequent and could not be conclusively tied to (or ruled out from being related to) ketamine treatment, the occurrence of 2 suicides early in treatment underscores the importance of rigorous monitoring and safety planning, particularly during the induction phase.

One important interpretation note on our findings concerns the variability in assessment wave timing. Because participants reached each assessment wave at variable times (ie, differing numbers of days), models included a nonlinear spline for days from baseline. The model-based expected symptom scores (Figure 4) reflect changes associated with progression through sessions, while simultaneously accounting for heterogeneous time between sessions via the spline term. This approach allowed us to detect trends in the marginal means across waves and disseminate highly interpretable findings. However, it does not allow us to pinpoint the exact impact of treatment timing itself on psychometric outcomes, highlighting the need for future research with stricter time-anchored treatment protocols.

While there are limitations to the approach used in this paper, there are also great benefits to real-world observational designs; they provide data representative of real-world practice, and focus on the types of patients who would naturally gravitate toward at-home ketamine, rather than a sample that may never have naturally chosen this treatment modality. Thus, this design is sufficient to establish a baseline of safety and feasibility, as well as initial evidence of clinical use and observed outcomes that can be built on in future research.

Additional studies may disentangle the effect of support from guides and clinicians from the impact of ketamine itself; the relative roles of therapeutic “integration” and the physiological impact of ketamine on psychiatric symptoms; and the extent to which the results observed in the users accessing Mindbloom may be biased due to the self-selected nature of the sample.

In summary, this paper demonstrates that telehealth protocols can safely implement a high-bioavailability injection model for home ketamine administration, at scale, without causing serious clinical complications. The broader implications of these findings challenge the assumption that exists, among some practitioners, that ketamine can only be used when closely supervised in a clinic setting [57]. Specifically, our findings suggest that remote guardrails, such as remote clinician check-ins, video-verified self-injection competency, and the provision of BP monitoring equipment at home, can effectively decentralize ketamine care without compromising patient safety. When administered at subanesthetic dosages, with strategies in place to provide clear instruction in a remote context, these data are suggestive of clinical benefit and acceptable safety outcomes, even in the absence of the costly and complex risk evaluation and mitigation strategy requirements that govern use of Spravato [58].