At-home, telehealth-supported ketamine treatment for depression: Findings from longitudinal, machine learning and symptom network analysis of real-world data

Abstract

Background:

Improving safe and effective access to ketamine therapy is of high priority given the growing burden of mental illness. Telehealth-supported administration of sublingual ketamine is being explored toward this goal.

Methods:

In this longitudinal study, moderately-to-severely depressed patients received four doses of ketamine at home over four weeks within a supportive digital health context. Treatment was structured to resemble methods of therapeutic psychedelic trials. Patients receiving a second course of treatment were also examined. Symptoms were assessed using the Patient Health Questionnaire (PHQ-9) for depression. We conducted preregistered machine learning and symptom network analyses to investigate outcomes (osf.io/v2rpx).

Results:

A sample of 11,441 patients was analyzed, demonstrating a modal antidepressant response from both non-severe (n = 6384, 55.8 %) and severe (n = 2070, 18.1 %) baseline depression levels. Adverse events were detected in 3.0–4.8 % of participants and predominantly neurologic or psychiatric in nature. A second course of treatment helped extend improvements in patients who responded favorably to initial treatment. Improvement was most strongly predicted by lower depression scores and age at baseline. Symptoms of Depressed mood and Anhedonia sustained depression despite ongoing treatment.

Limitations:

This study was limited by the absence of comparison or control groups and lack of a fixed-dose procedure for ketamine administration.

Conclusions:

At-home, telehealth-supported ketamine administration was largely safe, well-tolerated, and associated with improvement in patients with depression. Strategies for combining psychedelic-oriented therapies with rigorous telehealth models, as explored here, may uniquely address barriers to mental health treatment.

1. Introduction

Ketamine, an N-methyl-d-aspartate (NMDA) receptor-mediated dissociative drug, has received substantial attention in the last decade as a breakthrough mental health intervention (Sanacora et al., 2017). Though ketamine was approved for medical use by the United States Food and Drug Administration (FDA) as an anesthetic in 1970, its psychiatric value went largely unrecognized until 2000, when the first randomized controlled trial using a subanesthetic dose of ketamine for the treatment of depression indicated positive results (Berman et al., 2000). Numerous studies have since replicated and expanded upon these findings, providing evidence that ketamine can be effective in treating depression and shows promise as a pharmacotherapy for a wider range of mental health disorders (Walsh et al., 2021).

Despite interest in dissociative and psychedelic medicines as a novel category of rapidly-acting mental health treatments (Johnston et al., 2023; Lepow et al., 2023; Nayak et al., 2023; O’Donnell et al., 2023), several issues have limited broader adoption of ketamine as a psychiatric intervention. Most significantly, ketamine has not yet been approved by the FDA for use as an antidepressant. Racemic ketamine is widely available as a generic medication, and its lack of patent protection has deterred sponsorship of the large, resource-intensive trials that are required for the approval of ketamine for a new therapeutic indication (Wilkinson and Sanacora, 2017). Psychiatric applications of ketamine are considered “off-label,” limiting treatment accessibility and imposing significant cost barriers because insurance coverage is rarely available. Adoption is further complicated by a lack of consensus about optimal models of treatment delivery, including with respect to routes of administration, treatment setting, and the role of adjunctive psychotherapy or psychosocial support (Mathai et al., 2022a).

In 2019, the FDA approved intranasal esketamine (i.e., the patented S-enantiomer that forms one half of ketamine as a racemic mixture) for treatment-resistant depression. However, a growing body of evidence suggests that ketamine may benefit from an analogous safety profile, superior efficacy and lower dropout rates when compared with esketamine (Bahji et al., 2021; Correia-Melo et al., 2020; Nikolin et al., 2023; Singh et al., 2016). Financial analyses suggest that ketamine assisted treatment is a more affordable investment for the mental healthcare sector (Brendle et al., 2022; Dadiomov, 2020; Ottawa (ON): Canadian Agency for Drugs and Technologies in Health, 2021). Accordingly, improving safe access to generic, racemic ketamine is of high priority given the high rates of partial- or non- response observed with currently available antidepressants (Vasiliu, 2022). Oral or sublingual ketamine could uniquely address this need at low cost (Andrade, 2023; Dutton et al., 2023; Swainson and Khullar, 2020), particularly in cases where treatment with esketamine is not readily available, and may be of greater public interest than more invasive routes of ketamine administration (Mathai et al., 2021).

The use of telehealth presents an additional opportunity to expand access to mental health care by reducing treatment barriers including travel, time, and cost without substantially diminishing the quality of treatment (Cuijpers et al., 2019; Guinart et al., 2021; Hull et al., 2022). To this end, our group recently published an analysis of 1247 patients who used an online medical service (www.mindbloom.com) for at-home treatment with sublingual ketamine with access to a prescribing psychiatric clinician and behavioral coach through a secure telehealth and remote monitoring platform (Hull et al., 2022). The treatment protocol consisted of four medication doses (initiated at ~5 mg/kg) over four weeks with ongoing mental health support, given the anticipated safety and benefit of combining forms of psychosocial intervention with ketamine, and as seen with investigational psychedelic therapies involving MDMA and psilocybin (Dore et al., 2019; Mathai et al., 2022a; Tsang et al., 2023). These results provided initial evidence for the safety and effectiveness of at-home, sublingual ketamine telehealth for depression, as demonstrated by a low incidence of adverse reactions (3.8 %), and rates of response (62.8 %) and remission (32.6 %) that were consistent with laboratory- and clinic-administered ketamine treatment.

The current study was based on ongoing data collection using the same telehealth platform and focused on replication of the previous study within a substantially larger patient sample (n > 10,000) who received at-home, telehealth-supported ketamine treatment for psychiatric care. This report also included preliminary data on patients who received a second course of treatment with ketamine, used novel machine learning and symptom network analyses to examine treatment outcomes, and presented previously unpublished data on medication dosage. We had the following aims: 1) to assess clinical characteristics and treatment response for depressed patients who received one course of ketamine treatment; 2) to assess clinical characteristics and treatment response for a subsample of depressed patients who received a second course of treatment; 3) to explore if machine learning models could be used to organize heterogeneous subgroups of patients and identify predictors of treatment response; and 4) to use network analysis (Fonseca-Pedrero, 2017) to explore changes in individual symptoms of depression. These aims seek to enhance our understanding of several poorly understood features of ketamine as an intervention in real world settings.

2. Methods

2.1. Participants

2.1.1. Patients

Participants were self-referred adult patients who presented to the telehealth platform with a chief complaint of anxiety or depression and were seeking treatment through the web-based service. Inclusion criteria to access treatment through the platform consisted of: 1) being 18 years old or older, 2) having regular internet or cell phone access, 3) receiving a depression or anxiety diagnosis from their selected licensed clinician based on a video-based clinical intake interview, and 4) having access to a safe, private environment for dosing sessions and a trusted adult who could be physically present during treatment.

Exclusion criteria consisted of: 1) ketamine use disorder, 2) ketamine allergy or hypersensitivity, 3) active psychotic, manic or mixed symptoms, 4) history of primary psychotic disorder, 5) active suicidal ideation with method, intent, or plan in the past 3 months, 6) suicide attempt within the past year, 7) uncontrolled hypertension, 8) congestive heart failure or other impaired cardiac status, 9) severe and poorly-controlled respiratory problems (e. g., chronic obstructive pulmonary disease), 10) diagnosis of hyperthyroidism, 11) elevated intraocular pressure/glaucoma, 12) pregnancy, nursing, or currently trying to become pregnant, or 13) any other severe systemic disease, or other aspect of the patient’s psychiatric history, outpatient support system or home environment that would render at-home treatment psychologically unsafe in the opinion of the prescribing psychiatric clinician. Clinicians also screened for the following issues, and determined eligibility for treatment on a case-by-case basis: 1) active moderate or severe substance use disorder, and 2) history of severe trauma. When necessary, clinicians required labs, EKG, coordination with external mental health providers, confirmation of regular, ongoing third-party therapeutic support, and/or clearance from medical providers following a physical examination as appropriate to ensure patients were appropriate for treatment.

Analysis was limited to patients with moderate-to-severe depression as indicated by a baseline score ≥ 10 on the 9-item Patient Health Questionnaire. Study flowchart is reported in the Supplementary Materials (Supplementary Fig. 1).

2.1.2. Clinicians and guides

All clinicians in the provider network were Psychiatrists, Psychiatric-Mental Health Nurse Practitioners (PMHNPs), or Physician Associates (PAs) with psychiatry experience. Behavioral coaches, referred to as “Guides,” were required to have coaching certification and/or to have provided one-on-one behavioral coaching for more than one year with at least 50 clients. For more information on prescribing clinicians and behavioral coaches, see Supplement.

2.2. Procedures

Data collection for this study was conducted as part of organizational quality assurance and program management processes between January 21, 2021 and June 16, 2023. All patients and clinicians gave consent to the use of their data in a de-identified, aggregate format for research purposes as part of the user agreement before using the platform. Study analytic procedures were approved as exempt by the institutional review board at New York University (i21–01533). General procedures have been described elsewhere (Hull et al., 2022) and are detailed in the Supplement, including information on drug administration, treatment monitoring and general precautions. Briefly, patients completed a standardized medical and psychiatric intake evaluation with a psychiatric clinician through live video conference to determine if inclusion/exclusion criteria were met and appropriateness for treatment. Patients meeting criteria for treatment were mailed a limited supply of ketamine as needed for one course of treatment in the form of sublingual rapid dissolve tablets (RDTs) divided over two separate shipments. Dosing was initiated at 300 mg to 600 mg of ketamine RDT to approximate a weight-based dose of 5 mg/kg and adjusted using an ongoing, individualized dose discovery process that was informed by initial tolerability and targeted a mild to moderate level of dissociation for medication sessions. When taking ketamine, patients were asked to hold the tablets under the tongue or between the cheek and gums, without swallowing for 7 min, after which they were instructed to spit out all saliva. One course of treatment consisted of four medication sessions over four weeks with the availability of ongoing clinical, psychosocial, and peer support, organized within a structure of preparation, dosing, and integration that is considered best practice for psychedelic medicine trials (Johnson et al., 2008).

Patients had the option to elect an additional round of treatment. This required an additional consultation with the prescribing clinician and a determination of whether treatment was still appropriate, after which patients could be prescribed a second round of ketamine resembling the procedures above.

2.2.1. Assessments

Clinical assessments were used throughout treatment to track progress and facilitate care (Hull et al., 2022). Measures included the 9-item Patient Health Questionnaire (PHQ-9; (Kroenke and Spitzer, 2002)), the 7-item Generalized Anxiety Disorder questionnaire (GAD-7; (Spitzer et al., 2006)), the Columbia Suicide Severity Rating Scale (C-SSRS; (Posner et al., 2011)), the Alcohol Use Disorders Identification Test (AUDIT; (Bush et al., 1998)), the Drug Abuse Screening Test 10- item (DAST-10; (Villalobos-Gallegos et al., 2015)). Follow-up symptom measures for the first round of treatment were administered after the second and fourth medication sessions, along with an adverse event self-report measure, which asked, “Have you noticed any issues with your physical or mental health since beginning treatment?” Measures for the second round of treatment were administered before the first medication session and then after the fourth medication session. Any adverse events reported to the clinician or guide outside of this measure were recorded in the electronic health record (EHR) and were identified for inclusion in analysis as well. Dropout was measured by the number of individuals who canceled ongoing treatment or who were removed from treatment by the clinician due to adverse events or noncompliance.

2.3. Analyses

Analytic procedures for the study were pre-registered on the Open Science Framework (osf.io/v2rpx) and consisted of four steps. First, we assessed the treatment characteristics in the sample. Our analysis involved calculating clinical metrics and adverse events based on complete data. We adopted this approach to replicate previous findings (Hull et al., 2022) and to compare our results with other studies (Alnefeesi et al., 2022). We also examined these metrics in the subgroup of individuals undergoing a second round of treatment, as a preliminary finding on treatment maintenance. Second, we examined the treatment outcome trajectories. Our analysis involved using unsupervised machine learning to identify patterns of depression over time. We adopted this approach to distinguish subgroups of patients who were improving and not improving (Hull et al., 2020; Malgaroli et al., 2024), based on an intent-to-treat analysis. Third, we identified which patient and treatment characteristics increased the likelihood of improvement following treatment. Our analysis utilized supervised and explainable machine learning methods. We chose this approach due to the capability of these methods to identify and rank important features associated with the amelioration of depression (Lee et al., 2018; Schultebraucks et al., 2021a). Fourth, we identified which symptoms of depression played a role in treatment non-response. Our analysis utilized panel network models to longitudinally explore changes in individual symptoms over treatment. We chose this approach due to the ability of network analysis to pinpoint the most salient symptoms that sustain depression (Malgaroli et al., 2021).

2.3.1. Clinical metrics and adverse events

We assessed clinical and patient characteristics descriptively and measured treatment response, defined as a ≥50 % reduction in symptom score on the PHQ-9 from baseline to session 4 of treatment (week 0–4). We also measured clinically significant change (CSC; (Jacobson and Truax, 1991)) defined as a reduction of ≥5 points and a score that began above the clinical threshold of 10 and fell below it at follow-up, remission defined as beginning above the clinical threshold at baseline and having a follow-up score of <5 (Coley et al., 2020), changes in GAD-7 score, deterioration defined as a reliable increase in score of ≥5 points (Jacobson and Truax, 1991), and adverse events (AEs). PHQ-9 and GAD-7 outcomes were presented for individuals with complete survey responses at each timepoint of observation in order to facilitate interpretation and comparison with the prior study.

AEs were measured after initiation of treatment, and those reported from automated patient questionnaires were reviewed and manually coded by system organ class using the Common Terminology Criteria for Adverse Events (CTCAE) Version 5.0 (33). Incident report forms and clinician logs were analyzed for AEs of special interest (i.e., intense dissociation and sedation) and serious adverse events (SAEs). SAEs were defined in accordance with ICH Good Clinical Practice criteria (34).

We also examined the subsample of patients who received a second round of treatment. We conducted a sensitivity analysis to examine differences between the Round 1 and 2 samples, and a cohort analysis to separate trajectories for patients who did and did not achieve treatment response after Round 1 of treatment.

2.3.2. Outcome trajectories

Depression trajectories over treatment were identified using Latent Growth Mixture Modeling (LGMM) via MPLUS version 8 (Muthén and Muthén, 2017). LGMM is an unsupervised machine learning method that teases heterogeneity by identifying sub-groups of individuals who share similar patterns (i.e., PHQ-9 scores over time). We implemented an intent-to-treat approach for the LGMM by including all patients who had baseline data available, and handled cases of missing data through Full Information Maximum Likelihood (FIML) for the LGMM estimation. The number of trajectories was determined by comparing nested unconditional models with increasing classes based on recommendation from the literature (Nylund et al., 2007).

2.3.3. Improvement features: machine learning analyses

We used supervised machine learning to identify characteristics discerning patients more likely to be on an improving symptoms trajectory (Schultebraucks et al., 2021b). We used a Random Forest classification algorithm (Breiman, 2001) with default hyperparameters from the caret R package (Kuhn, 2008). Candidate variables of interest included demographics, baseline depression, ketamine dosage, and AEs at week 2. Training for the model consisted of nested-cross validation minimizing cross-entropy loss, with 10 inner- and 10 outer- folds via nestedcv (Lewis et al., 2023). Nested-cross validations offer more robust models for medical applications (Bates et al., 2023). The inner folds fine tune the model’s hyperparameter, while the outer folds offer unseen test data for model evaluation. We categorized trajectories into a binary classification problem to improve model interpretability. Random undersampling of the majority class was performed to rebalance class distribution in the data.

We interpreted feature importance in the classification model using SHapley Additive exPlanations (SHAP) values (Lundberg and Lee, 2017). Derived from game theory, SHAP values provide accurate and consistent local estimations to interpret the importance of each variable. SHAP values were generated using the package fastshap (Jethani et al., 2022). Large absolute SHAP values indicate important features for the model, with values closer to zero suggesting less significance. The sign of SHAP values indicates the direction of a feature’s impact on the prediction (i.e., high or low variable scores).

2.3.4. Non-response features: temporal panel network analyses

We examined symptoms patterns over treatment in non-responders using network analysis (Borsboom et al., 2021). The network approach identifies highly influential symptoms sustaining depression by examining their interactions (Malgaroli et al., 2021). Specifically, we explored symptom pathways sustaining non-response over treatment to identify treatment adjuncts. We estimated temporal networks using panel graphical vector-autoregression (panel-GVAR) from the package Psychonetrics v0.11.5 (Epskamp, 2020). In a panel-GVAR temporal network, directed edges at time t are contingent upon all nodes both at t and at the preceding time t-1. In addition to estimating within-person temporal effects, the panel GVAR model also computes contemporaneous and between-subjects effects, which can be used to estimate corresponding networks. As panel-GVAR models are based on the assumption of stationarity (Epskamp, 2020), we limited our analysis to patients on chronic trajectories, given relative stability of symptoms in treatment non-response.

We calculated panel-GVAR including symptoms of depression from baseline to week 4 using FIML. We first estimated a base model in which all edges were included. Subsequently, we estimated a pruned model and a step-up model using a significance threshold of α = 0.05 for the pruning or addition of individual edges respectively. We then identified the best fitting panel-GVAR based on Bayesian Information Criteria (BIC).

We used expected influence (Robinaugh et al., 2016) to identify the symptom in the temporal network that might play the greatest role in the activation and maintenance of depression over treatment. Expected influence in temporal networks is composed of In-Expected Influence and Out-Expected Influence. In-Expected Influence predicts the effect of other nodes on the activation of a specific symptom. Out-Expected Influence represents the influence of a symptom in activating the rest of the network.

3. Results

3.1. Sample characteristics

The sample consisted of 11,441 patients from ages 19 to 88, with an average of 42.0 years (SD = 10.5). Of these, 54.7 % were women (n = 6233), and 6.8 % (n = 778) lived in areas classified by the Center for Medicare and Medicaid Services as rural.

3.2. Clinical metrics and adverse events (AEs)

Table 1 presents the descriptive clinical outcomes for the entire sample for both depression and anxiety symptoms among individuals who completed survey responses at each timepoint. Average tablet dosage per session was 590 mg (SD =245 mg), or 7.3 mg/kg based on an average participant weight of 81.2 kg.

Table 1.

Clinical characteristics for full sample.

PHQ-9 observations	Available	Mean (SD)	Cohen’s d (95 % CI)	Response rate	Remission rate	Clin. sig. change	Deteriorated
Baseline	11,441	15.5 (4.1)	–	–	–	–	–
Session 2	7400	9.4 (5.5)	1.15 (1.12–1.18)	42.3 %	18.1 %	48.9 %	0.9 %
Session 4	4918	7.9 (5.2)	1.46 (1.41–1.50)	56.4 %	28.1 %	61.4 %	0.7 %
GAD-7 observations	Available	Mean (SD)	Cohen’s d (95 % CI)	Response rate	Remission rate	Clin. sig. change	Deteriorated
Baseline	7776	15.2 (3.3)	–	–	–	–	–
Session 2	5074	9.0 (5.4)	1.18 (1.14–1.21)	43.7 %	20.4 %	50.2 %	0.7 %
Session 4	3348	7.6 (5.1)	1.46 (1.41–1.51)	56.1 %	28.8 %	62.2 %	0.4 %

Note: Response rate defined as 50 % or larger reduction in symptoms. Remission defined as final symptom score below 5. Clinically significant change as moving below the clinical threshold (score of <10) AND improving at least 5 points. Deterioration as worsening of symptoms by 5 or more points. Abbreviations: PHQ-9 = Patient Health Questionnaire; GAD-7 = Generalized Anxiety Disorder questionnaire.

AEs were reported by 323 of 7496 patients (4.3 %) with valid responses after session 2 and by 242 of 5085 patients (4.8 %) after session 4. Across timepoints, the most common AEs were memory impairment (0.6–1.1 %), suicidal ideation (0.6–0.7 %), abdominal pain (0.4–0.6 %), dysuria (0.2–0.5 %), hypertension (0.1–0.4 %), chest discomfort (0–0.4 %), headache (0–0.4 %), dyspnea (0.2–0.4 %), and cravings (0.2–0.4 %). See Supplementary Tables S1–6 for full AE data.

3.3. Adverse events of special interest (AESI)

Intense dissociation and related experiences: Clinician reports and cancellation forms indicated 12 patients (0.1 %) who had treatment discontinuation (either self-initiated or initiated by treatment team) due to intense dissociation or other psychologically overwhelming drug experiences. Self-report AE forms indicated two patients with flashbacks and re-experiencing of prior trauma as a result of ketamine administration.

Sedation:

Clinician reports indicated two patients who experienced a depressed level of consciousness resulting in Emergency Department (ED) visits. Both recovered spontaneously and were discharged from the ED without the need for significant medical intervention or hospitalization.

3.4. Serious adverse events (SAEs) and treatment discontinuation

SAEs occurred in 6 patients, were all psychiatric by classification and consisted of severe depression (n = 1), suicidal ideation (n = 1), suicidal behavior (n = 2), and psychosis (n = 2), requiring inpatient admission in all but one of these cases. A total of 46 patients (0.4 %) had treatment discontinuation (either self-initiated or initiated by treatment team) due to AEs, generally for psychiatric reasons. The leading reasons for discontinuation were psychologically overwhelming experiences (n = 12), anxiety (n = 5), and agitation (n = 3). Treatment was discontinued for two individuals because of mania (n = 1) and hypomania (n = 1). Non-psychiatric reasons for discontinuation included vomiting (n = 3), syncope (n = 3), and headache (n = 3).

3.5. Cohort analysis for patients electing a second round of treatment

Two cohort pathway analyses were conducted to examine the effect of an additional round of treatment on clinical outcomes.

The first for participants who achieved Clinically Significant Change (CSC) at session 4 (n = 3020) in the first round of treatment and then elected to continue and receive a second round (n = 1241). Of these, 84 % of patients either maintained (if PHQ-9 scores remained ≤9 at the beginning of Round 2; 66 %) or recovered (if PHQ-9 scores increased to >9 at the beginning of Round 2; 18 %) CSC with a second round of treatment. The rest became symptomatic with PHQ-9 scores >9.

The second for participants who did not achieve CSC at session 4 (n = 1898) in the first round of treatment and then elected to continue and receive a second round (n = 665). From this cohort, 66 % of participants had PHQ-9 scores >9 at the beginning of Round 2, and 28 % of those met criteria for CSC with a second round of treatment.

Dosage was higher on average than the first round treatment at a mean of 758 mg (SD = 297 mg; Cohen’s d difference of 0.45, 95 % CI, 0.40–0.49), or 9.3 mg/kg based on average participant weight. Patients electing an additional treatment round were also more likely to be older and currently in outpatient therapy, and less likely to be divorced, separated, or widowed, live in a rural area, or have higher baseline scores on the DAST-10 or AUDIT than those who did not elect to continue for another round (see Table 2 for all comparisons and odds ratios). AEs were reported in 25 of 829 patients (3.0 %) with valid responses who pursued a second round of treatment.

Table 2.

Differences for patients electing a maintenance round.

			95% Confidence interval
	Odds Ratio	p	Lower bound Upper bound
Gender (female)	1.025	0.571	0.941	1.117
Age (older)	1.020	< .001	1.015	1.025
Divorced/Separated/Widowed (yes)	0.860	0.046	0.741	0.997
Rural zip code (yes)	0.842	0.049	0.709	0.999
BMI	1.001	0.778	0.994	1.008
Inpatient, History (yes)	1.077	0.230	0.954	1.215
Outpatient therapy, History (yes)	1.120	0.149	0.960	1.305
Outpatient therapy, Currently (yes)	1.152	0.002	1.054	1.259
AUDIT	0.985	0.002	0.975	0.995
DAST-10	0.954	0.015	0.919	0.991
CSSRS	0.886	0.225	0.729	1.077
PHQ-9 Baseline	1.000	0.893	0.994	1.007
GAD-7 Baseline	1.001	0.823	0.994	1.008

Abbreviations: p = P-value; BMI = Body Mass Index; AUDIT = Alcohol Use Disorder Identification Test; DAST-10 = Drug Abuse Screening Test; CSSRS = Columbia Suicide Severity Rating Scale; PHQ-9 = Patient Health Questionnaire; GAD-7 = Generalized Anxiety Disorder questionnaire.

3.6. Outcome trajectories

Fig. 1 displays the four trajectories identified to characterize depression symptoms patterns from baseline to Session 4. LGMM model fit information is reported in Supplementary Table S7. Two Improvement trajectories were identified, with PHQ-9 scores ameliorating below the clinical cutoff by the final observation. The first was Improvement, not severe (n = 6384, 55.8 %), the modal trajectory consisting of patients with moderate baseline depression, and the Improvement, severe trajectory (n = 2070, 18.1 %), characterized by more severe depressive symptoms at baseline. Two Chronic trajectories indicative of non-response to treatment were also identified: the Chronic, not severe (n = 2101, 18.4 %), and the Chronic, severe (n = 886, 7.4 %), with the latter characterized by severe baseline depression scores. In both Chronic trajectories, patients continued to endorse symptoms that did not subside over the course of treatment, remaining at moderate to severe depression levels. Sensitivity analysis examined survey non-adherence (See Supplementary Table S8), indicating that individuals not responding to surveys at week 2 or 4 were more likely to be assigned to the Improvement, not severe group compared to all other trajectories.

Fig. 1.

Outcome trajectories of Patient Health Questionnaire (PHQ-9) depression scores (n = 11,441).

3.7. Improvement features

We ran a random forest classifier to predict Improvement trajectories membership from patient baseline characteristics, ketamine dosage, and session 2 AEs (n = 11,441). The final model achieved an overall Accuracy of 0.804, and the Area Under the ROC Curve was 0.813, indicating good discriminative ability between Improvement and Chronic groups. The model Precision was 0.865, and Recall (or sensitivity) was 0.886, with a F1 score (harmonic mean of precision and recall) of 0.875. Further model characteristics are reported in Supplementary Figs. S2 and S3.

SHAP (SHapley Additive exPlanations) values were used to evaluate the impact of patients and treatment characteristics in the model’s prediction. The higher the SHAP value of a feature, the higher the importance in predicting likelihood of an Improvement trajectory. Examination of SHAP values (Fig. 2) suggests that depression symptoms endorsed at baseline (PHQ-9 scores) was the most important feature, indicating that lower initial scores were associated with higher likelihood of improvement. The age of the patients was also predictive of improvement, with a trend toward younger age increasing probability of improvement. Other demographic factors including patients’ body mass index (higher) and gender (female) also contributed to the model’s predictions, albeit to a lesser extent. Interestingly, ketamine dosage, while having a lower median SHAP value, showed some of the highest individual SHAP values, indicating specific cases where medication dosage was highly influential.

Fig. 2.

Feature importance for nested cross-validated Random Forest model of Improvement using SHAP values dot plot (n = 11,441).

Patient characteristics are ranked on the Y axis based on feature importance for the model (i.e., highest mean absolute SHAP values). Dots represent the attribution for each feature per patient, with values ranging from red (lower) to blue (higher). The X axis shows the effect of the feature in increasing (right) or decreasing (left) the likelihood of depression Improvement trajectories membership. Abbreviations: SHAP = SHapley Additive exPlanations; PHQ-9 = Patient Health Questionnaire; AUDIT = Alcohol Use Disorder Identification Test; DAST-10 = Drug Abuse Screening Test; CSSRS = Columbia Suicide Severity Rating Scale. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

3.8. Non-response features

The temporal network is estimated only for non-responder Chronic patients (n = 2987) using a panel-GVAR model. Model fit information is reported in Supplementary Table S9. Panel-GVAR also allows to estimate two additional networks aggregated across all time points: a within-subjects contemporaneous and a between-subject networks, which are reported in Supplementary Figures 4 and 5. Temporal Network stability metrics are reported in Supplementary Figs. S6 and S7.

The temporal network edges (Fig. 3) show mutual activations between Depressed Mood and Anhedonia. The two symptoms have the highest Out-Expected Influence (Fig. 4), suggesting their importance in sustaining the network of depressive treatments over time. Taken together, results from the temporal network indicate the influential role of Mood and Anhedonia in the maintenance of depression despite treatment.

Fig. 3.

Temporal network of depression symptoms over treatment in non-responder Chronic patients (n = 2987). Nodes in the temporal network represent individual depressive symptoms, while edges represent predictive temporal connections between symptoms. Blue and red edges represent positive and negative association respectively, with thicker edges representing stronger connections. Curved edges indicate the autoregressive stability of a symptom over treatment. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 4.

Expected Influence of depression symptoms in temporal network over treatment for non-responder Chronic patients (n = 2987).

4. Discussion

In this longitudinal study of at-home ketamine for depression, we found ongoing evidence for the safety and potential for improvement associated with treatment administered within a carefully supportive digital health context. Patients examined here received one or more courses of treatment, each consisting of four medication sessions with sublingual racemic ketamine over four weeks along with structured clinical, psychosocial, and peer support. To our knowledge, this is the largest safety and effectiveness study of any ketamine derivative or psychedelic-oriented intervention to date.

Preregistered analyses showed rates of clinically significant change, response, remission, deterioration, and adverse events comparable to those previously reported in a smaller sample with similar characteristics (Hull et al., 2022). Using a weight-based flexible-dosing procedure, patients were started on an initial dose of 300–600 mg ketamine RDT and titrated based on response to an average dose of 590 mg across treatment. This mean falls within the higher range of dosages previously reported for oral or sublingual ketamine when swallowed (Andrade, 2019; Hassan et al., 2022; Nuñez et al., 2020) but is difficult to compare because of differences in drug administration (e.g., lower bioavailability with spitting vs swallowing). It is likewise difficult to compare dosages here to those involving other routes of ketamine administration.

Of note, these procedures were unique relative to studies of intravenous ketamine or intranasal esketamine, which have generally pursued a dose optimization strategy that minimizes the subjective experience of dissociation (Mathai et al., 2022a). However, there has been little evidence to date that drug-induced dissociation negatively impacts therapeutic outcomes (Hull et al., 2022), with most studies conveying a neutral or positive relationship between dissociation and antidepressant efficacy for ketamine (Mathai et al., 2020, 2023b). Furthermore, it has been proposed that clinical outcomes with dissociative interventions might be optimized beyond the effects of the molecule alone within therapeutic frameworks that value and support the occurrence of psychologically meaningful drug experiences (Mathai et al., 2022a). Here, a flexible-dosing strategy was used targeting a mild to moderate level of dissociation with ketamine and found high rates of clinical response and tolerability, with only 12 of 11,441 patients with discontinuation of treatment due to intense dissociation or other psychologically overwhelming drug experiences.

Adverse events that occurred were predominantly neurologic or psychiatric in nature, including reports of memory impairment, suicidal ideation, headache, drug cravings, depression, and anxiety. All serious adverse events were psychiatric. With the exception of drug cravings, these events were consistent with established risks for commonly prescribed antidepressant medications (Gosmann et al., 2023; Jakobsen et al., 2017; Patel et al., 2015; Preda et al., 2001). Unlike conventional pharmacotherapies, ketamine is known for its prominent subjective effects that mimic intoxication (Mathai et al., 2022b), with significant qualitative experiences that can range from more pleasant to challenging (Breeksema et al., 2023, 2022; Mollaahmetoglu et al., 2021; Sumner et al., 2021). Risks of drug craving and misuse with ketamine have been identified previously, including the potential for a ketamine use disorder (Chubbs et al., 2022; Vines et al., 2022), but are mitigated in clinical settings that emphasize judicious prescribing and appropriate supervision of patients (Swainson et al., 2022). Together, these findings highlight the importance of ketamine administration within contexts that are equipped to provide behavioral support for subjective drug experiences that occur (Mathai et al., 2023a) and to adequately manage relevant psychiatric risks, as demonstrated here.

Some have argued for increased regulation of ketamine (Harding, 2023; Wilkinson et al., 2024), drawing comparisons to the use of a formal Risk Evaluation Mitigation Strategy (REMS) with esketamine, because of concerns for abuse, excessive dissociation, sedation, and respiratory depression. Furthermore, at least one case of at-home ketamine use involving excessive sedation and respiratory depression requiring serious medical intervention has been reported (Johnson et al., 2024). Importantly, these events of interest occurred minimally in our sample, were monitored closely when they did occur, and at no point met criteria for a serious adverse event. While the question of optimal regulation remains, our findings do not support the need for a REMS-type program for ketamine when administered at subanesthetic dosages, with clear instructions for use, and with the degree of supervision detailed here. Rather than providing significant clinical benefit, such programs are likely to increase administrative burden on healthcare systems, while decreasing access to treatment (Wilson and Milne, 2011). The clinical approach used here offers an example of how a rigorous telehealth program might be used to decrease the sizeable costs of psychedelic therapies (Hull et al., 2022; Williams et al., 2021) while maintaining a high quality of care.

Another critical area of study has involved extending the duration of benefit with ketamine, which has led to therapeutic strategies such as repeated medication administration (Kryst et al., 2020; McMullen et al., 2021; Phillips et al., 2019). The antidepressant benefits of a single dose of ketamine are virtually undetectable by two weeks, but even with repeated dosing, symptom relapse occurs at high rates within the first month of treatment cessation (Smith-Apeldoorn et al., 2022). Findings from our cohort analysis support the value of ongoing treatment among patients who demonstrate a favorable response to ketamine initially, as a way of maintaining improvement. For patients without response to an initial course of treatment, only a minority went on to achieve significant change pursuing a second trial, suggesting that the benefits of ongoing administration of ketamine may be diminished relative to risk for most of these individuals. This is especially worth considering given both the limited quality of data on the risks of long-term ketamine use (Meshkat et al., 2023; Nikayin et al., 2022), and our findings that the amount of ketamine administered tends to increase over time using flexible-dosing procedures. Importantly, even when maintenance ketamine treatment is appropriate for patients, clinicians are advised to actively monitor the risk-benefit ratio and need for ongoing treatment, the availability of alternative treatment options, and the potential for tolerance, relapse, or recurrence (McIntyre et al., 2021).

As before (Hull et al., 2022), we identified several unique sub-populations with different trajectories over the course of treatment. Trajectories of improvement were most strongly predicted by lower depression scores at baseline, as has been the case for many (Trivedi et al., 2006; Van et al., 2008) but not all studies (Fournier et al., 2010; Friedman et al., 2012) of antidepressant interventions. Other studies of ketamine and esketamine have similarly found lower illness severity to be predictive of response (Lucchese et al., 2021), and higher severity to be predictive of non-response (Chen et al., 2021; Jesus-Nunes et al., 2022).

Consistent with other studies of ketamine for depression, we also observed that younger age increased the probability of improvement (Chen et al., 2021; Turkoz et al., 2023). While not assessed here, it is possible that younger adults have greater drug potencies or are more sensitive to therapeutic effects driven by drug-induced changes in neuroplasticity (Strawn et al., 2023). Younger age has also been associated with more positive expectations involving psychedelics (Žuljević et al., 2022) and greater intensity of acute psychedelic effects (Aday et al., 2021), both of which may be relevant to therapeutic outcomes with ketamine and warrant further study.

Lastly, using network analysis of individual depressive symptoms, we found that mutual activation of depressed mood and anhedonia had a substantial role in maintaining depression despite ongoing treatment. This connection has been frequently demonstrated within symptom networks of depression (Malgaroli et al., 2021) and, for ketamine particularly, suggests the value of harnessing changes in neuroplasticity with specific interventions that are geared toward behavioral activation, and increasing engagement with activities and experiences that are rewarding for patients (Hasler, 2020; Lepow et al., 2021; Phillips et al., 2023; Serretti, 2023). Further research in this area is needed, but such combinations may enhance medication response and even lead to more durable forms of benefit (Mathai et al., 2022a).

4.1. Limitations

This study has several key limitations. First, the high number of participants with survey non-adherence at week 4, though expected for real-world studies, may have been relevant to outcomes. To mitigate this risk, we employed an intent-to-treat analysis and used likelihood-based adjustment to account for missing data when calculating outcome trajectories. Moreover, sensitivity analyses suggested a lower probability of poor clinical response for survey non-adherent patients. Next, the absence of a fixed-dose procedure, which is often considered gold standard for evaluation of dose response, may be seen as a drawback of the study and a source of bias caused by non-random treatment adjustment. However, it has also been argued that flexible-dose studies are more representative of actual clinical practice and risk/benefit considerations, since dose is often changed in accordance with patient response (Lipkovich et al., 2008). Finally, the lack of comparison or control groups makes it difficult to quantify the exact contribution of the treatment to reported outcomes. While these data support the ecological effectiveness (Singal et al., 2014) of sublingual ketamine through a sample of over 10,000 treatment-seeking patients, further study of the efficacy or comparative effectiveness of this intervention is warranted through randomized controlled trials comparing sublingual ketamine to placebo, other modes of ketamine administration, or conventional drug or talk therapies for depression.

5. Conclusion

We found that at-home ketamine administration within a supportive digital health infrastructure was largely safe, well-tolerated, and associated with improvement in patients with depression. Strategies for combining telehealth with ketamine and similar psychedelic therapies, as explored here, may uniquely address barriers to mental health treatment and increase access to care.