Published on 23 September 2026

Analysis · Anxiety & Depression · Autism Spectrum Disorder

In autism spectrum disorder, does the right non-drug intervention depend on the symptom?

◆ Collection
Frontiers in Psychiatry · 2025; 16: 1660412 · Ding et al.
DOI 10.3389/fpsyt.2025.1660412
PMID 41409332
Scientific 72
Editorial 78

Editorial collections The Monday Morning Patient PEB Academy PEB Decodes PEB Practice Change A Study Under the Microscope

The essentials

This network meta-analysis pools 67 randomized trials and 3,604 participants with autism spectrum disorder, comparing seven families of non-pharmacological interventions across three outcomes: anxiety, depression, and quality of life. Which intervention comes out on top depends entirely on which symptom is being treated. Mindfulness-based intervention ranks first for anxiety, with a cumulative SUCRA of 94.3%, and physical activity ranks first for quality of life, at 84.3%. For depression, cognitive behavioral therapy is the only modality whose effect against control reaches statistical significance, yet the ranking places it second, at 70.5%, behind a heterogeneous group of other approaches whose own estimate is not significant. Four caveats should govern how this is read. Certainty was graded using the GRADE approach, and none of the thirty-two comparisons reaches a high level. Inconsistency testing is not uniformly clean: one loop in the quality-of-life network is formally inconsistent. Heterogeneity reaches 93% for anxiety. And the population is nearly 93% male and predominantly high-functioning, which narrows how far the message travels. On top of this sits an internal reliability problem: the published abstract reports effect sizes and rankings for all three outcomes that neither the body of the article nor its supplementary tables reproduce.

Context

Anxiety and depression are the most common reasons an already-diagnosed autistic adult is referred for psychiatric consultation. The question raised is almost never diagnostic. It is a question of orientation: which type of non-drug care to refer the patient to, and in what order.

Until now the literature answered by intervention family, trial by trial, without setting these approaches against one another. Network meta-analysis is designed to do exactly that, comparing indirectly modalities that have never been tested head to head, provided the trial network is internally consistent. That is what this work attempts, and it is also what calls for the caution developed further down.

The study at a glance

Population
3,604 participants with autism spectrum disorder diagnosed under standardized DSM-5 or ICD-10 criteria, 1,909 of them in intervention arms, all ages combined, mean age 17.82 years (SD 10.82). Women made up 7.1% of the sample. Over 70% of trials, not of participants, explicitly recruited high-functioning autistic individuals, defined by normal or near-normal IQ and at least partial verbal communication.
Interventions compared
Seven families of non-pharmacological interventions: mindfulness-based intervention, cognitive behavioral therapy, behavioral and functional training, physical activity, sensory therapy, technology-assisted and family-based intervention, and a residual “other” group that includes, among others, animal-assisted therapy, drama therapy, and sleep training. Neither medication nor neuromodulation appears anywhere in the network.
Comparator
Control arms from the included trials, pooling no intervention, waitlist, treatment as usual, and nonspecific psychoeducation. The nature of the control therefore varies from trial to trial within a single network node.
Outcomes
Anxiety, depression, and quality of life, with no declared hierarchy among them, measured with heterogeneous scales across trials, hence the use of standardized mean differences. The three networks do not rest on the same trials: the publication reports 49 trials and 2,714 participants for anxiety, 20 trials and 1,052 participants for depression, and 17 trials and 861 participants for quality of life. Interventions are ranked by surface under the cumulative ranking curve (SUCRA).
Design
Systematic review and frequentist multivariate random-effects network meta-analysis, 67 randomized trials, five databases searched through March 28, 2025, PROSPERO registration CRD420251021423, conduct reported as compliant with the PRISMA extension for network meta-analyses. The search is restricted to English-language publications and excludes all grey literature. On the Oxford CEBM level-of-evidence scale used by this review, the design sits at level 1a.
Duration
Intervention duration from 1 to 52 weeks, mean 12.2 weeks (SD 7.4), sessions of 45 to 180 minutes, one to five times weekly. The signal is strongest in the 9-to-16-week window across all three outcomes. Average dropout was 13.6% in intervention arms and 11.2% in control arms.

Quality control

Point checked	Verdict
Protocol registration	Number provided
FindingThe protocol is registered with PROSPERO under number CRD420251021423, with conduct reported as compliant with the PRISMA extension for network meta-analyses. The registration date does not appear in the document reviewed: whether the deposited criteria match the reported ones cannot be established from the publication alone.
Data extraction	Duplicate, with arbitration
FindingSelection, extraction, and risk-of-bias assessment were each carried out independently by two reviewers, with a third reviewer resolving disagreements, meeting the expected methodological standard.
Network consistency	One loop formally inconsistent
FindingThree families of tests were run for each of the three outcomes: loop-specific, global, and local. The global tests are all negative. But Supplementary Table S11 reports, for quality of life, a loop connecting cognitive behavioral therapy, control, and physical activity with an inconsistency factor of 9.131, a z of 5.516, and a p value below 0.001, while the body of the publication states that every inconsistency test exceeds 0.05. The GRADE assessment separately downgrades three of the thirty-two comparisons for inconsistency. The condition without which an indirect comparison means nothing is therefore not met everywhere.
Publication bias	Symmetric funnels, GRADE downgraded
FindingFunnel plots are described as approximately symmetric, and for anxiety, symmetry is only achieved after removing two small outlying studies. In the opposite direction, the GRADE assessment downgrades 27 of the 32 comparisons for publication bias. An absence of visual asymmetry lowers suspicion without ruling it out, since this kind of test has low power.
Certainty of the estimates	No comparison rated high
FindingCertainty was assessed using the GRADE approach, comparison by comparison. Supplementary Table S17 lists 32 comparisons: none reaches the high level, 8 are moderate, 21 are low, and 3 are very low. Twenty-nine are downgraded for study limitations and 27 for publication bias. The summary table published in the body of the article does not match this detailed table: it cites imprecision and indirectness, neither ever invoked in the detailed table, and it rates as very low a comparison the detailed table rates as low. A high rank is therefore not a recommendation.
Allocation concealment	55.2% of trials
Finding37 of 67 trials report a procedure for allocation concealment. Separately, 20 trials are rated at high risk of performance bias, which is expected for non-pharmacological interventions that are difficult to blind.
Heterogeneity	Up to 93%
FindingIn pairwise comparison, heterogeneity reaches 93% for anxiety, 87% for quality of life, and 75% for depression. It reflects the diversity of scales, formats, and populations, and it widens uncertainty well beyond what the point intervals suggest.

Results

67
Randomized trials pooled in the network, covering 3,604 participants. The ranking of interventions is not the same depending on whether anxiety, depression, or quality of life is the target.

Outcome	What the source reports
Anxiety	Mindfulness-based intervention leads, cumulative rank 94.3%
ReadingAgainst control, the network estimate for mindfulness-based intervention is a standardized mean difference of -1.13 (95% CI -1.69 to -0.57). Next come cognitive behavioral therapy at -0.65 (-0.93 to -0.36) and behavioral and functional training at -0.62 (-1.09 to -0.14). Physical activity, at -0.55, does not cross the threshold, its interval running from -1.55 to 0.45. This is also the outcome with the highest heterogeneity. The published abstract reports -0.84 (-1.32 to -0.36) for mindfulness-based intervention and a cumulative rank of 91.4%: neither the body of the article nor its supplementary tables reproduce these values.
Depression	Cognitive behavioral therapy, the only significant modality
ReadingCognitive behavioral therapy is the only intervention whose interval does not cross zero: -0.33 (-0.64 to -0.03) as a network estimate and -0.35 (-0.61 to -0.09) in direct comparison, with low heterogeneity at 6%. The effect is modest, but it is the most homogeneous result in the whole study. The ranking, however, places the heterogeneous “other” group first at 73.4%, ahead of cognitive behavioral therapy at 70.5%, even though that group’s own estimate, -0.40 (-0.94 to 0.14), is not significant. The published abstract reports -0.77 and a rank of 90.1%, values that appear in neither table.
Quality of life	Physical activity leads, cumulative rank 84.3%
ReadingIn direct comparison, physical activity improves quality of life, standardized mean difference of 0.98 (0.61 to 1.35), with no detected heterogeneity, and technology-assisted and family-based intervention follows at 0.72 (0.43 to 1.01), with 87% heterogeneity. The network estimate against control, by contrast, does not cross the threshold, at 4.16 (-0.56 to 8.88), on a scale not comparable to the pairwise comparisons despite carrying the same label. The body of the publication attributes to physical activity and to technology-assisted and family-based intervention, against control, the values 5.32 and 4.66: in the supplementary league table, those are in fact the values for their comparisons against behavioral and functional training.
Modulators	Prespecified subgroups, results diverge by outcome
ReadingSubgroup analyses are described as prespecified on clinical and theoretical grounds, with no exploratory post hoc splitting. For anxiety, the effect is stronger in adults, -0.61 (-0.82 to -0.40), than in children and adolescents, -0.38 (-0.56 to -0.20). For depression, the difference between age groups is negligible, 0.25 versus 0.24, p = 0.948. For quality of life, it reverses: the effect is stronger under age 18, -0.59 versus -0.31. Protocols of 9 to 16 weeks stand out across all three outcomes. The direction of the estimates reported in these subgroups does not follow the measurement convention the authors state elsewhere, which limits how far they can be used.

Critical appraisal

Domain	Risk of bias
Randomization process	Uneven
Finding62 of 67 trials, 92.5%, describe sequence generation, but only 37, 55.2%, document allocation concealment. The authors’ own overall judgment rates 43 trials at low risk, 19 at moderate risk, and 5 at high risk. The tool is described as version 2.0 of the Cochrane risk-of-bias scale, yet the seven domains actually assessed are those of the earlier version.
Deviation from intended protocol	Blinding not possible
FindingA mindfulness or physical activity intervention cannot be concealed, and 20 trials are rated at high risk of performance bias. Comparison against a waitlist maximizes the expectancy effect, which weighs particularly heavily on self-reported outcomes.
Outcome measurement	Mixed scales, unblinded raters
FindingInstruments are both heterogeneous and of mixed nature: self-report questionnaires, parent-completed questionnaires, and clinician ratings coexist within the same outcome, and one of the quality-of-life scales measures the family’s quality of life rather than the participant’s. Only 34 of 67 trials report a blinded assessor. The standardized mean difference allows these measures to be pooled, it does not correct for this limitation.
Transitivity	Weakened
FindingIndirect comparison assumes trials are exchangeable on effect-modifying factors. Controls ranging from no intervention to psychoeducation within the same node weaken this assumption. For quality of life the assumption does not hold at all: one loop in the network is formally inconsistent.
External validity	Narrow
FindingPopulation nearly 93% male, and over 70% of trials explicitly recruited high-functioning individuals. The authors themselves state that transposing these findings to women, non-verbal individuals, and those with co-occurring intellectual disability is limited. Add to this the restriction to English-language publications and the complete exclusion of grey literature.
Interpretation of ranks	A classic trap
FindingA surface under the cumulative ranking curve expresses a probability of being among the most effective modalities, not a clinically measurable difference. Depression supplies its own illustration here: the “other” group is ranked first even though its estimate is not significant, while the only significant modality is ranked second. Two modalities separated by dozens of rank points can have effects that overlap substantially.

Level of evidence

Scientific72

Editorial78

Systematic review and network meta-analysis of randomized trials, which places the design at level 1a on the Oxford CEBM scale. Actual certainty runs well below what that rank suggests: of the thirty-two comparisons the authors graded using the GRADE approach, none reaches the high level, eight are moderate, twenty-one are low, and three are very low. Confidence is reasonable regarding the direction of the signal for cognitive behavioral therapy in depression, where the effect is modest but homogeneous and the comparison against control is rated moderate. It is low for the other rankings, owing to heterogeneity, near-systematic downgrading for publication bias, and, for quality of life, a loop inconsistency the authors do not address. It is very low for any autistic person whose profile departs from that of the included participants, namely women and those with substantial support needs. The study evaluates no pharmacological intervention and therefore permits no comparison with one. Finally, the published abstract agrees with none of its own tables on the three principal estimates and the three rankings: anyone citing this study should cite the tables, not the abstract.

The colleague test

What an experienced colleague might say about this study in two minutes, between two consultations.

“The message appeals to me, it matches what I already do without knowing whether I’m right about it. What bothers me is the ranking. When I’m handed a 94% probability of being the best option while the certainty behind that very comparison is rated low, I know it will end up on a slide as though it were settled. And look at who’s actually in these trials. Seven percent women. My autistic patients, the ones diagnosed at forty after three misdiagnoses, are almost absent from these sixty-seven trials.”

Translated for practice: the result is useful for choosing a first-line referral for a specific symptom in an autistic adult without substantial support needs. It says nothing solid beyond that, and the ranking should not be read as a hierarchy of quality between approaches.

What you can do with this

Orient by symptom rather than by diagnostic label: the useful question is not which therapy for autism, but which approach for this anxiety, this depression, or this quality-of-life complaint.
For comorbid depression, cognitive behavioral therapy remains the best-supported choice here, with a modest effect that should be presented as such.
For a complaint centered on quality of life rather than mood, consider structured physical activity, keeping in mind that the benefit rests on direct comparisons rather than the network estimate.
Plan an evaluation window of at least two to three months, consistent with the 9-to-16-week window that stands out across the subgroup analyses, rather than concluding after four weeks.
Tell the patient what is not known: these data concern overwhelmingly autistic men without substantial support needs, and they do not transpose automatically.
The course of action is set out in the NICE decision tree for generalised anxiety and panic disorder.

Frequently asked questions

Is mindfulness better than cognitive behavioral therapy?

For anxiety, it ranks higher. A better rank means a higher probability of being among the most effective modalities, which is not the same thing as demonstrated superiority between two approaches. The direct comparison between these two modalities does exist in the anxiety network, both direct and indirect, and it is inconclusive: the estimated difference is 0.49, with an interval running from -0.10 to 1.07. The certainty attached to the comparison between mindfulness and control is, moreover, rated low.

Do these results hold for children?

The authors ran prespecified age analyses on all three outcomes. For anxiety, the effect is stronger in adults than in children and adolescents, but it remains significant in the latter. For depression, the two age groups do not differ. For quality of life, it is those under 18 who respond most. The mean participant age, 17.82 years, places the sample at the boundary between adolescence and adulthood: these results are therefore not adult literature simply transposed onto children.

What about autistic people with substantial support needs?

They are very sparsely represented, and the authors themselves state this as the study’s first limitation. Nothing in these data says these interventions work for them, nor that they do not. Absence of data is not a result.

Was GRADE applied, and what does it show?

Yes, comparison by comparison. Of the thirty-two comparisons graded, none reaches the high level: eight are moderate, twenty-one are low, and three are very low. The downgrading is almost always due to study limitations and risk of publication bias. In other words, a ranking can be stable and still rest on fragile trials, which is exactly what the authors find. One caveat compounds this: the summary table published in the body of the article and the detailed table in the supplementary material do not give the same reasons for downgrading, nor the same comparison rated very low for quality of life.

Can these interventions replace medication?

The network contains no medication arm: protocols combining a drug or a neuromodulation technique were excluded from the outset. The question was therefore never asked, and no conclusion, in either direction, can be drawn from this study.

Annotated bibliography

Source study. Ding X, Luo H, Zhang J, Yang H, Fan Y, Wu J, Wu S. Comparative effectiveness of non-pharmacological interventions for anxiety, depression, and quality of life in individuals with autism spectrum disorder: A systematic review and network meta-analysis. Frontiers in Psychiatry, 2025; 16: article 1660412. Received July 6, 2025, revised November 12, 2025, accepted November 13, 2025, published December 2, 2025. DOI 10.3389/fpsyt.2025.1660412. PMID 41409332. Systematic review registered with PROSPERO, number CRD420251021423. Funding: the authors declare having received no financial support for the research or the publication of this article. Conflicts of interest: the authors declare that the research was conducted in the absence of any commercial or financial relationship that could constitute a conflict of interest. Open-access article under a Creative Commons Attribution license. Handling editor: Luca Simione; reviewers: Daniela Lopes Gomes and Christopher Keary. The authors declare having used no generative artificial intelligence in writing the manuscript; the publisher notes, however, that figure alt-text was produced with the assistance of artificial intelligence. The supplementary material, four methods appendices, seventeen tables, and thirty-eight figures, was consulted and serves as the reference whenever the body of the article and the appendices diverge.

Editorial collections