Comparative efficacy of pharmacological, adjunctive, and non-pharmacological interventions across behavioral domains in children and adolescents with autism spectrum disorder: a network meta-analysis

Abstract

Objective

To systematically compare the efficacy of pharmacological, adjunctive, and non-pharmacological interventions across multiple behavioral domains in children and adolescents with autism spectrum disorder (ASD), including irritability, hyperactivity, social withdrawal, stereotypic behavior, and inappropriate speech.

Methods

Following the 2020 PRISMA guidelines, five databases were searched from inception to March 2026 for randomized controlled trials (RCTs). A frequentist network meta-analysis was conducted to estimate the comparative efficacy and ranking probabilities of different interventions across behavioral domains (PROSPERO: CRD420261368925).

Results

Sixty-seven RCTs (N = 4,203) were included. For irritability, structured exercise ranked highest (SUCRA = 88.0%; SMD = − 1.48, 95% CI −2.24 to − 0.73), followed by risperidone plus adjunctive anti-inflammatory agents (SUCRA = 72.9%) and neuromodulatory agents (SUCRA = 67.8%). For hyperactivity, pharmacological interventions were most effective, with other antipsychotic/ADHD medications ranking highest (SUCRA = 95.1%; SMD = − 1.85, 95% CI −2.50 to − 1.20), followed by structured exercise (SUCRA = 76.1%). Similarly, for stereotypic behavior, pharmacological treatments dominated, with other antipsychotic/ADHD medications achieving the best performance (SUCRA = 92.0%; SMD = − 1.10, 95% CI −1.67 to − 0.52). In contrast, for social withdrawal, structured exercise emerged as the most effective intervention (SUCRA = 98.9%; SMD = − 1.28, 95% CI −1.96 to − 0.60), followed by risperidone plus neuromodulatory agents (SUCRA = 75.3%). For inappropriate speech, structured exercise (SUCRA = 80.1%) and probiotics plus oxytocin (SUCRA = 76.4%) demonstrated the highest rankings, with structured exercise showing a significant effect (SMD = − 0.80, 95% CI −1.40 to − 0.19).

Conclusion

This study establishes a domain-specific treatment hierarchy for ASD, demonstrating that intervention effects vary substantially across behavioral domains. Pharmacological approaches are more effective for hyperactivity and stereotypic behaviors, whereas structured exercise shows superior benefits for irritability and social withdrawal. Combination strategies consistently rank highly across multiple domains, supporting the value of multimodal interventions. These findings underscore the importance of individualized, symptom-targeted, and mechanism-informed treatment strategies to optimize clinical outcomes in ASD.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13034-026-01132-2.

Introduction

Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by persistent deficits in social communication, restricted interests, repetitive behaviors, and sensory abnormalities, as defined in the Diagnostic and Statistical Manual of Mental Disorders, fifth edition (DSM-5) [1]. Beyond these core features, many children and adolescents with ASD exhibit a range of maladaptive behavioral symptoms, including irritability, hyperactivity, social withdrawal, stereotypic behavior, and inappropriate speech [2]. These behavioral disturbances are associated with substantial functional impairment and increased caregiver burden. According to the Centers for Disease Control and Prevention (CDC), approximately one in 36 children has been diagnosed with ASD, and the prevalence has shown a steadily increasing trend [3]. Against this backdrop of rising prevalence and significant functional burden, effective interventions targeting these heterogeneous behavioral domains are urgently needed.

The clinical presentation of autism spectrum disorder (ASD) is inherently heterogeneous, necessitating multidimensional assessment frameworks to capture the complexity of behavioral manifestations [4]. Within this context, standardized instruments such as the Aberrant Behavior Checklist (ABC) are widely used to evaluate multiple behavioral domains, including irritability, social withdrawal, stereotypic behavior, hyperactivity, and inappropriate speech [5, 6]. These domains may represent clinically meaningful and potentially distinct therapeutic targets. Although the ABC is among the most commonly applied tools, other validated instruments have also been used across studies to assess similar behavioral dimensions, including the Social Responsiveness Scale (SRS), Gilliam Autism Rating Scale (GARS), Repetitive Behavior Scale (RBS), and Nisonger Child Behavior Rating Form (N-CBRF) [2]. However, many previous reviews and meta-analyses have focused on specific intervention classes, single outcomes, or overall behavioral improvement, and therefore tend to collapse distinct clinical outcomes into aggregate estimates, which may obscure domain-specific treatment effects [7–9]. In addition, few comparative syntheses have evaluated multiple intervention classes within the same domain-based framework. Establishing a domain-specific evidence framework is therefore important for clarifying which interventions may be more suitable for specific symptom profiles and for supporting more individualized clinical decision-making in children and adolescents with ASD.

A wide range of interventions have been investigated to address behavioral symptoms in autism spectrum disorder (ASD), encompassing pharmacological treatments, adjunctive strategies, and non-pharmacological approaches [10–12]. Pharmacological interventions, particularly antipsychotics and other psychotropic medications, are widely used in clinical practice and have demonstrated efficacy in reducing certain behavioral symptoms such as irritability and hyperactivity [13–15]. In the present review, adjunctive treatments refer to active pharmacological or biologically based agents added to risperidone or standard pharmacotherapy, with the aim of enhancing therapeutic effects through additional anti-inflammatory, immunomodulatory, metabolic, neuroprotective, or neuromodulatory mechanisms. Despite the expanding body of research, the overall evidence remains fragmented and inconclusive [1]. Pharmacological treatments are often associated with adverse effects and uncertain long-term tolerability, while findings across trials are frequently inconsistent [16]. Non-pharmacological interventions, including structured exercise, behavioral interventions, and sensory-based approaches, have also been investigated for children and adolescents with ASD. However, structured exercise was selected as the principal non-pharmacological node in the present network because eligible randomized controlled trials using structured exercise programs more consistently reported outcomes that could be mapped onto the predefined behavioral domains and connected to the evidence network. Other non-pharmacological interventions, such as behavioral or sensory-based therapies, were not incorporated as separate nodes because their intervention components, comparators, and outcome measures were highly heterogeneous and did not form sufficiently stable or connected networks for the present analysis.

Although previous systematic reviews and meta-analyses have examined specific pharmacological or non-pharmacological interventions for ASD, direct comparative evidence across intervention classes remains limited. Existing evidence syntheses have often focused on a single intervention class, a specific medication, or one broad behavioral outcome, rather than comparing pharmacological, adjunctive, and non-pharmacological interventions across multiple behavioral domains within a unified network framework. Network meta-analysis allows the integration of both direct and indirect evidence and can provide comparative estimates among interventions that have not been directly compared in head-to-head trials. Therefore, we aimed to conduct a network meta-analysis of randomized controlled trials to compare the efficacy of pharmacological, adjunctive, and non-pharmacological interventions across multiple behavioral domains in children and adolescents with ASD, including irritability, hyperactivity, social withdrawal, stereotypic behavior, and inappropriate speech. By integrating both direct and indirect evidence, this study aimed to establish a domain-specific hierarchy of treatment effects and to inform evidence-based, symptom-targeted clinical decision-making.

Methods

Registration

This systematic review and NMA were conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [17]. The study protocol has been prospectively registered in the International Prospective Register of Systematic Reviews, ID: CRD420261368925. Any deviations from the protocol will be reported and justified in the final manuscript.

Literature search strategy

A systematic literature search was conducted in PubMed, Cochrane Library, Web of Science, Embase, and EBSCO from database inception to March 2026. The search strategy was developed according to the PICOS framework and combined controlled vocabulary terms (e.g., MeSH terms) with free-text terms related to autism spectrum disorder, behavioral symptoms, and representative pharmacological, adjunctive, and non-pharmacological interventions. The full PubMed search strategy is presented in Table 1 and was adapted as appropriate for use in the other databases. Given the broad range of potentially relevant interventions for ASD and the difficulty of exhaustively listing every individual drug, supplement, or non-pharmacological approach in a single search formula, the electronic database search was supplemented by manual screening of the reference lists of relevant reviews and all included articles to identify additional eligible randomized controlled trials. This supplementary screening helped capture potentially eligible trials involving interventions such as antidepressants/SSRIs and other psychotropic or biologically based agents, even when some specific intervention names were not explicitly listed in the PubMed search formula. No restrictions were placed on publication year, but only studies published in English were included (Table 1).

Table 1.

Search strategies of pubmed database

Serial number	Search formula
#1	((((((((“Exercise“[Mesh]) OR ((((Physical Exercise) OR (Aerobic Exercise[Title/Abstract])) OR(Isometric Exercises[Title/Abstract]))) OR (“Risperidone“[Mesh])) OR ((Risperdal[Title/Abstract]) OR (Risperdal Consta[Title/Abstract]))) OR (“Antipsychotic Agents“[Mesh])) OR (((((Neuroleptics[Title/Abstract]) OR (Antipsychotic Agent[Title/Abstract])) OR (Neuroleptic[Title/Abstract])) OR (Antipsychotic[Title/Abstract])) OR (Neuroleptics[Title/Abstract]))) OR (“Dietary Supplements“[Mesh])) OR ((((Dietary Supplementations[Title/Abstract]) OR (Nutraceuticals[Title/Abstract])) OR (Nutriceutical[Title/Abstract])) OR (Food Supplementations[Title/Abstract]))) OR (((((((((((Aripiprazole[Title/Abstract]) OR (Lurasidone[Title/Abstract])) OR (Atomoxetine[Title/Abstract])) OR (Guanfacine[Title/Abstract])) OR (Omega-3[Title/Abstract])) OR (Arbaclofen[Title/Abstract])) OR (Prednisolone[Title/Abstract])) OR (Metformin[Title/Abstract])) OR (Vitamin D[Title/Abstract])) OR (Ginkgo[Title/Abstract])) OR (Folinic acid[Title/Abstract]))
#2	(((“Stereotyped Behavior“[Mesh]) OR ((((((Stereotyped Behaviors[Title/Abstract]) OR (Lethargy[Title/Abstract])) OR (Social Withdrawal[Title/Abstract])) OR (Irritability[Title/Abstract])) OR (Hyperactivity[Title/Abstract])) OR (Inappropriate Speech[Title/Abstract]))) OR (“Autism Spectrum Disorder“[Mesh])) OR ((Autistic Spectrum Disorder[Title/Abstract]) OR (Autism Spectrum Disorders[Title/Abstract]))
#3	#1 AND #2

Inclusion criteria

Eligibility criteria were defined using the PICOS framework:

Population (P): Children and adolescents with a diagnosis of ASD based on standardized diagnostic criteria (e.g., DSM or validated diagnostic instruments).

Intervention (I): Interventions were required to be classifiable into prespecified network nodes defined a priori, including structured exercise (EXE; e.g., aerobic exercise, yoga, horseback riding, martial arts); risperidone monotherapy (RIS; including risperidone administered with placebo); risperidone plus adjunctive treatments, including anti-inflammatory agents (RIS_AI; e.g., pentoxifylline, celecoxib, prednisolone, sulforaphane, resveratrol, palmitoylethanolamide) and neuromodulatory agents (RIS_NM; e.g., memantine, baclofen, pregnenolone, galantamine, riluzole, N-acetylcysteine, topiramate); other antipsychotic or ADHD medication monotherapy (OAP; e.g., aripiprazole, lurasidone, atomoxetine, guanfacine, arbaclofen); metabolic agents (MET; e.g., metformin, pioglitazone, L-carnitine); probiotics combined with oxytocin (PRO_OXT); and other monotherapies (NUT_BIO; e.g., omega-3 fatty acids, vitamin D, folinic acid, tetrahydrobiopterin, probiotics, fluoxetine, citalopram).For combination arms, interventions were categorized based on the active component(s), and placebo co-administration was not considered a separate intervention.Network nodes were defined a priori according to clinical use, treatment context, dominant therapeutic rationale, and the feasibility of forming connected evidence networks. Because many individual agents or specific exercise modalities were represented by only one or a small number of trials, excessively granular nodes would have resulted in sparse or disconnected networks and unstable comparative estimates. Therefore, clinically related interventions were grouped into broader pragmatic nodes while preserving clinically meaningful distinctions where possible. We acknowledge that some nodes, particularly NUT_BIO, OAP, and structured exercise, remained mechanistically heterogeneous; therefore, node-level findings were interpreted as estimates for intervention classes rather than for each individual agent or modality.

Comparison (C): Placebo, treatment as usual (TAU), or inactive control conditions (e.g., daily activities), with or without concomitant non-pharmacological care.

Outcomes (O): Quantitative measures of behavioral symptoms across multiple domains, including irritability, hyperactivity, social withdrawal, stereotypic behavior, and inappropriate speech, assessed using validated rating scales (e.g., Aberrant Behavior Checklist [ABC], Social Responsiveness Scale [SRS], or other established instruments capturing these domains).

Study design (S): Randomized controlled trials.

Trials were eligible if they reported sufficient data to calculate effect sizes for at least one of the predefined behavioral domains and allowed mapping of at least one active arm to a prespecified intervention node.

Exclusion criteria

Studies were excluded if they met any of the following criteria:

1. Not published in English;

2. Records without accessible full text (including theses, conference papers, or abstract-only reports);

3. Insufficient outcome data for effect-size calculation and failure to obtain missing data after contacting the corresponding author; or

4. Duplicate publications or overlapping datasets (for multiple reports from the same trial, only the most complete or most recent report was retained).

Literature screening and data extraction

Two reviewers independently screened titles and abstracts, followed by full texts, in accordance with the predefined eligibility criteria. The screening was conducted in a two-stage process (title/abstract followed by full text), with any disagreements resolved through consensus or, when necessary, consultation with a third reviewer. A PRISMA flow diagram was utilized to document the study selection process.

For each eligible RCT, data were extracted using a standardized form, including: first author, publication year, country, ASD diagnostic criteria, study design, sample size per arm, mean age and standard deviation, sex ratio, and setting. Intervention characteristics were extracted and mapped to the prespecified network nodes (structured exercise; risperidone monotherapy; risperidone plus adjuncts; other antipsychotic/ADHD monotherapy; and other monotherapies, such as nutraceutical, metabolic, or neuromodulatory agents). This included recording the specific agent and dosage for pharmacological interventions, or the type, intensity, frequency, and duration for structured exercise. Comparator types (placebo, TAU, or inactive control) were also recorded for each study.

Regarding outcomes, we extracted baseline and post-intervention means and standard deviations (or change scores, where available) for measures across the five core behavioral domains: irritability, hyperactivity, social withdrawal, stereotypic behavior, and inappropriate speech. Information on adverse events and overall treatment discontinuation was also extracted when reported. When necessary, study authors were contacted for clarification or to obtain missing outcome data.

Risk of bias evaluation of included studies

Two reviewers independently assessed the risk of bias for each included study using the original Cochrane risk-of-bias tool for randomized trials as described in the Cochrane Handbook for Systematic Reviews of Interventions [18]. The tool includes seven domains: (a) random sequence generation; (b) allocation concealment; (c) blinding of participants and personnel; (d) blinding of outcome assessment; (e) incomplete outcome data; (f) selective reporting of results; and (g) other sources of bias. Each domain was rated as low, unclear, or high risk of bias. The overall risk-of-bias category was derived using predefined decision rules based on domain-level judgments. Studies with no domains rated as high risk and three or fewer domains rated as unclear risk were classified as having a low overall risk of bias; studies with one domain rated as high risk or four or more domains rated as unclear risk were classified as having a moderate overall risk of bias; and studies with two or more domains rated as high risk were classified as having a high overall risk of bias. Discrepancies were resolved through discussion or consultation with a third reviewer.

Quality of evidence evaluation

The quality of evidence for each study was evaluated independently by two reviewers using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach. The GRADE framework considers five domains: study limitations (risk of bias), consistency of results, directness of evidence, precision of effect estimates, and publication bias. Based on these domains, the certainty of evidence for each outcome was rated as “high,” “moderate,” “low,” or “very low.” Any disagreements were resolved by consensus or adjudication by a third reviewer. Where appropriate, GRADE assessments were adapted to the context of network meta-analysis.

Statistical analysis

To minimize the impact of baseline differences, this study used the change in mean values and standard deviations before and after the intervention to combine effect sizes. The calculation method for the standard deviation of the change was performed using the formula provided in the Cochrane Handbook (version 6.3) [18]. Following the PRISMA guidelines for network meta-analysis [17], a random-effects model was used within a frequentist framework to combine effect sizes and calculate 95% confidence intervals (CI) in Stata 16.0 software [19],to assess the impact of multiple intervention classes on irritability, hyperactivity, social withdrawal, stereotypic behavior, and inappropriate speech in ASD. Due to the diversity of instruments used to assess these outcomes, the standardized mean difference (SMD) was used as the effect size for merging. The direction of effect was coded so that negative SMD values indicated greater improvement in behavioral symptoms. Because validated minimal clinically important differences are not consistently established across the different behavioral scales and domains included in this review, SMDs were interpreted primarily as standardized comparative effects rather than direct measures of clinically meaningful change. For multi-arm trials, each eligible intervention arm was extracted and assigned to its prespecified network node, and the potential influence of shared comparators was considered when interpreting estimates from nodes informed by multiple dose arms. When standard deviations of change scores were not directly reported, they were calculated using the formula recommended in the Cochrane Handbook where possible. If essential outcome data were unavailable or unclear, corresponding authors were contacted for clarification, and studies were excluded from the quantitative synthesis for a specific outcome only when sufficient data for effect-size calculation could not be obtained. To assess uncertainty, exploratory checks were conducted where feasible, including examination of transitivity, comparator type, blinding status, alternative node structures, sparse evidence, and local inconsistency.A network evidence plot was generated to describe the relationships between exercise interventions, where the lines connecting nodes represent direct comparisons between interventions, with line thickness proportional to the number of studies, and node size proportional to sample size. The consistency of each loop was evaluated by calculating the inconsistency factor and its 95% CI [20]. The inconsistency model was used to test for inconsistency, and when P > 0.05, the consistency model was applied for analysis [21]. Cumulative ranking probability plots (surface under the cumulative ranking, SUCRA) were used to rank and compare the effectiveness of different exercise interventions [22]. SUCRA values range from 0 to 100%, with 100% indicating the best intervention effect and 0 indicating the worst [23]. Funnel plots were used to check for publication bias or small sample effects. SUCRA rankings were interpreted as exploratory probabilistic rankings and were considered together with effect estimates, 95% confidence intervals, statistical significance, directness of evidence, and network sparsity. Therefore, a high SUCRA value was not interpreted as definitive evidence of superiority when the corresponding effect estimate was imprecise or not statistically significant.

Before interpreting indirect comparisons, we assessed the plausibility of the transitivity assumption by examining the distribution of potential effect modifiers across intervention nodes. These included participants’ age, sex distribution, diagnostic criteria for ASD, baseline behavioral symptom severity, outcome measurement instrument, intervention duration, comparator type, blinding status, concomitant treatment, and overall risk of bias. These characteristics were summarized descriptively across intervention nodes and are presented in Appendix (1) Detailed domain-level risk-of-bias judgments are presented in Appendix (2) Particular attention was paid to differences between pharmacological trials and structured exercise trials, because these intervention classes may differ in participant profiles, feasibility of blinding, and control conditions. Given the expected clinical and methodological heterogeneity across intervention classes, indirect comparisons were interpreted cautiously, especially when evidence was derived mainly from placebo-controlled comparisons, when direct evidence between active interventions was sparse, or when nodes were informed by a limited number of studies.

Results

Literature search results

A total of 12,961 records were identified through database searches (PubMed, n = 4,264; Embase, n = 4,434; Web of Science, n = 3,842; Cochrane Library, n = 338; EBSCO, n = 83). After removal of 6634 duplicates, 6,327 records remained for screening. Of these, 1024 were excluded as meta-analyses or systematic reviews (n = 786) and non-English-language articles (n = 238). The remaining 5303 records underwent title and abstract screening, and 5193 were excluded as irrelevant. We assessed 110 full-text articles for eligibility and excluded 43 (ineligible outcomes, n = 20; inclusion of adult participants, n = 12; lack of an appropriate control group, n = 1). Ultimately, 67 randomized controlled trials were included in the review. The study selection process is presented in Fig. 1.

Fig. 1.

Literature screening process and results

Basic characteristics of included studies

A total of 67 randomized controlled trials (RCTs), comprising 78 independent study comparisons or intervention datasets and involving 4,203 participants, were included in this analysis, with mean participant ages ranging from approximately 4 to 18 years. Across these trials, outcomes were reported across five behavioral domains, including irritability, hyperactivity, social withdrawal, stereotypic behavior, and inappropriate speech, with the majority of studies assessing multiple domains using standardized rating scales.Interventions were grouped into five active categories versus control: structured exercise (including yoga, mini-basketball training, aquatic exercise, horseback riding, and other physical activity-based programs); risperidone monotherapy; risperidone plus adjunctive treatments; other antipsychotic or ADHD medications (e.g., aripiprazole, lurasidone, atomoxetine, guanfacine, arbaclofen); and other monotherapies, including nutraceutical, metabolic, or neuromodulatory agents (e.g., omega-3 fatty acids, vitamin D, N-acetylcysteine [NAC], folinic acid, tetrahydrobiopterin [BH4], probiotics, metformin, and oxytocin).Control conditions included placebo and treatment-as-usual or standard care (TAU), as well as inactive control conditions such as daily activities or no-intervention controls, particularly in exercise-based trials.The duration of interventions ranged from 4 to 52 weeks, with most trials lasting between 8 and 12 weeks. Outcome measures were predominantly assessed using the Aberrant Behavior Checklist (ABC), while other validated instruments, including the Social Responsiveness Scale (SRS), Gilliam Autism Rating Scale (GARS), and Repetitive Behavior Scale (RBS), Nisonger-Child Behavior Rating Form (N-CBRF), were also used in a subset of studies.The basic characteristics of the included studies are summarized in Appendix 1.

The distribution of potential effect modifiers across intervention nodes is summarized in Appendix 1. Overall, participant age ranges, intervention duration, and assessed behavioral domains were broadly comparable across most nodes. However, methodological differences were observed across intervention classes. Pharmacological and adjunctive pharmacological trials were more commonly placebo-controlled and double-blinded, whereas structured exercise trials more frequently used usual-care, waitlist, or inactive control conditions and were generally unable to blind participants and personnel. Baseline symptom severity and outcome measurement instruments also varied across studies. Therefore, although the network was sufficiently connected for indirect comparisons, treatment rankings and indirect estimates were interpreted cautiously in light of these potential effect modifiers.

Evaluation results of risk of bias in included studies

Of the 67 included RCTs, 65 explicitly described their methods for group allocation, and 52 reported some form of blinding. Owing to the inherent difficulty of blinding participants and personnel in exercise interventions, most exercise trials in this review were rated as having a high or unclear risk of performance bias in this domain. Thirty-two studies reported blinding at the level of outcome assessment or data analysis. For incomplete outcome data, studies with a loss-to-follow-up rate greater than 20% and no use of an intention-to-treat approach were rated as high risk of bias. Among the included trials, 14 had a loss-to-follow-up rate exceeding 20%, whereas 50 studies reported attrition rates with broadly similar numbers and reasons for drop-out in the intervention and control groups.With respect to other potential sources of bias, studies were rated as high risk if they had very small sample sizes (fewer than 10 participants in any group), lacked adequate supervision of the intervention, or exhibited substantial measurement error in outcome assessments. Overall, risk of bias was assessed at the level of independent study comparisons or intervention datasets included in the network meta-analysis. Among the 78 independent study comparisons derived from 67 randomized controlled trials, 53 were classified as having a low overall risk of bias, 17 as having a moderate overall risk of bias, and 8 as having a high overall risk of bias. Detailed domain-level risk-of-bias judgments are presented in Appendix 2.

Network meta-analysis results

Network evidence diagram

In the network evidence plot (Fig. 2), the size of the nodes is positively correlated with the sample size of each intervention, and the thickness of the lines between nodes represents the number of studies comparing the connected interventions. The placebo (PBO) node was the largest and most connected, indicating that most comparisons were conducted against placebo, while direct comparisons between active interventions were relatively limited. We examined whether several heterogeneous nodes could be further divided into more homogeneous categories. However, further splitting of nodes such as NUT_BIO, OAP, and structured exercise resulted in sparse or disconnected networks for several behavioral domains, because many individual agents or specific exercise modalities were represented by only one or a small number of trials. Therefore, the prespecified pragmatic node structure was retained as the primary analysis, and findings for these nodes were interpreted at the intervention-class level rather than as evidence for any single agent or exercise modality.

Fig. 2.

Network meta-analysis comparing different interventions across multiple behavioral outcomes in children with ASD

Inconsistency test

For irritability, three of the four closed loops demonstrated good consistency (P > 0.05), whereas the PBO–RIS–RIS_NM loop showed significant inconsistency (P = 0.002). The inconsistency model indicated no significant global inconsistency (P = 0.0885). Node-splitting analysis further revealed that only the comparison between placebo and risperidone showed significant inconsistency (P = 0.004).

For hyperactivity, two loops (PBO–RIS–NUT_BIO, P = 0.01; PBO–RIS–RIS_NM, P = 0.032) exhibited significant inconsistency, while the remaining loops were consistent. No significant global inconsistency was observed (P = 0.5966), and node-splitting analyses showed no evidence of local inconsistency.

For inappropriate speech, all loops demonstrated good consistency (all P > 0.05), with no evidence of either global (P = 0.81) or local inconsistency.

For stereotypic behavior, two loops (PBO–RIS–RIS_NM, P = 0.013; PBO–RIS–MET, P = 0.031) showed significant inconsistency, whereas the global inconsistency was not significant (P = 0.6811). Node-splitting results indicated no local inconsistency.

For social withdrawal, one loop (PBO–RIS–RIS_NM, P = 0.004) showed significant inconsistency, while no significant global inconsistency was detected (P = 0.2887). Node-splitting analysis identified inconsistency in the comparison between placebo and risperidone (P = 0.04), whereas all other comparisons remained consistent.

Overall, several local inconsistencies were identified, particularly in loops involving risperidone-based interventions, such as PBO–RIS–RIS_NM and PBO–RIS–MET. These inconsistencies may reflect differences in baseline symptom severity, adjunctive agents, dosing regimens, outcome measurement instruments, comparator conditions, and study design characteristics across the included trials. Therefore, estimates from loops involving risperidone-based adjunctive interventions were interpreted cautiously.

Combining effect size analysis and ranking results

Irritability

Regarding irritability, the network meta-analysis results indicated that several active interventions were significantly more effective than placebo in reducing symptoms. Specifically, compared with the placebo group, structured exercise (EXE) (SMD = −1.48, 95% CI −2.24 to −0.73), risperidone plus adjunctive anti-inflammatory agents (RIS_AI) (SMD = −1.12, 95% CI −1.70 to −0.54), risperidone plus adjunctive neuromodulatory agents (RIS_NM) (SMD = −1.04, 95% CI −1.58 to −0.50), and other antipsychotic or ADHD medications (OAP) (SMD = −0.87, 95% CI −1.41 to −0.33) all demonstrated significant therapeutic advantages (Fig. 3). Risperidone monotherapy (RIS) showed a favorable direction of effect but did not reach statistical significance compared with placebo in the present network. This finding should be interpreted cautiously given the detected local inconsistency involving the PBO–RIS comparison and risperidone-based loops. No statistically significant differences were observed for other monotherapies (NUT_BIO), metabolic agents (MET), or the combination of probiotics and oxytocin (PRO_OXT) when compared with placebo. However, given the relatively large magnitude of this estimate and the methodological characteristics of structured exercise trials, including limited feasibility of participant blinding and frequent use of inactive or usual-care controls, this finding should be interpreted cautiously.

Fig. 3.

Results of network meta-analysis

The SUCRA probability ranking results further elucidated the relative hierarchy of these interventions (Fig. 4). Structured exercise (EXE) had the highest probability of being the most effective treatment for irritability, followed by RIS_AI, RIS_NM, and OAP. Probiotics plus oxytocin (PRO_OXT), metabolic agents (MET), other monotherapies (NUT_BIO), and risperidone monotherapy (RIS) showed lower probabilities of effectiveness, with placebo (PBO) consistently ranking as the least effective intervention.

Fig. 4.

Area under the cumulative probability ranking curve

Hyperactivity

Regarding hyperactivity, the network meta-analysis demonstrated that several interventions showed significant improvements compared with placebo. Specifically, structured exercise (EXE) (SMD = −1.31, 95% CI −2.23 to −0.39), risperidone plus adjunctive anti-inflammatory agents (RIS_AI) (SMD = −0.95, 95% CI −1.76 to −0.14), risperidone plus adjunctive neuromodulatory agents (RIS_NM) (SMD = −0.86, 95% CI −1.61 to −0.10), and other antipsychotic or ADHD medications (OAP) (SMD = −1.85, 95% CI −2.50 to −1.20) were all significantly more effective than placebo (Fig. 5). In contrast, no statistically significant differences were observed for risperidone monotherapy (RIS), other monotherapies (NUT_BIO), or metabolic agents (MET) when compared with placebo, as their confidence intervals crossed zero. Because some estimates for structured exercise were relatively large, particularly for irritability and hyperactivity, these results were interpreted cautiously. Structured exercise trials were generally less able to blind participants and personnel and often used inactive, usual-care, or waitlist control conditions, which may have contributed to larger observed effects.

Fig. 5.

Results of network meta-analysis

The SUCRA probability ranking results further elucidated the relative hierarchy of these interventions (Fig. 6). OAP had the highest probability of being the most effective treatment for hyperactivity, followed by EXE, RIS_AI, and RIS_NM. Metabolic agents (MET), other monotherapies (NUT_BIO), and risperidone monotherapy (RIS) showed lower probabilities of effectiveness, with placebo (PBO) consistently ranking as the least effective intervention.

Fig. 6.

Area under the cumulative probability ranking curve

Inappropriate speech

Regarding inappropriate speech, the network meta-analysis showed that several interventions demonstrated significant improvements compared with placebo. Specifically, structured exercise (EXE) (SMD = −0.80, 95% CI −1.40 to −0.19), risperidone plus adjunctive anti-inflammatory agents (RIS_AI) (SMD = −0.56, 95% CI −1.00 to −0.13), and other antipsychotic or ADHD medications (OAP) (SMD = −0.56, 95% CI −0.97 to −0.15) were all significantly more effective than placebo in reducing inappropriate speech symptoms (Fig. 7). In contrast, no statistically significant differences were observed for risperidone monotherapy (RIS), risperidone plus adjunctive neuromodulatory agents (RIS_NM), other monotherapies (NUT_BIO), metabolic agents (MET), or probiotics combined with oxytocin (PRO_OXT), as their confidence intervals crossed zero.

Fig. 7.

Results of network meta-analysis

The SUCRA probability ranking results suggested that probiotics plus oxytocin (PRO_OXT) and structured exercise (EXE) ranked highly for inappropriate speech (Fig. 8). However, the PRO_OXT estimate versus placebo was not statistically significant and was based on sparse evidence; therefore, this ranking should be interpreted cautiously and should not be regarded as definitive evidence of superiority. In contrast, structured exercise showed both a relatively high ranking and a statistically significant effect compared with placebo.

Fig. 8.

Area under the cumulative probability ranking curve

Stereotypic behavior

Regarding stereotypic behavior, the network meta-analysis demonstrated that several interventions were significantly more effective than placebo in reducing symptoms. Specifically, structured exercise (EXE) (SMD = −0.72, 95% CI −1.18 to −0.26), risperidone plus adjunctive anti-inflammatory agents (RIS_AI) (SMD = −0.58, 95% CI −1.10 to −0.06), risperidone plus adjunctive neuromodulatory agents (RIS_NM) (SMD = −0.65, 95% CI −1.25 to −0.05), and other antipsychotic or ADHD medications (OAP) (SMD = −1.10, 95% CI −1.67 to −0.52) all showed significant therapeutic advantages compared with placebo (Fig. 9).In addition, other monotherapies (NUT_BIO) also demonstrated a modest but statistically significant effect (SMD = −0.32, 95% CI −0.75 to −0.12).In contrast, no statistically significant differences were observed for risperidone monotherapy (RIS) or metabolic agents (MET), as their confidence intervals crossed zero.

Fig. 9.

Results of network meta-analysis

The SUCRA probability ranking results further elucidated the relative hierarchy of these interventions (Fig. 10). OAP had the highest probability of being the most effective treatment for stereotypic behavior, followed by EXE, RIS_NM, and RIS_AI. Other monotherapies (NUT_BIO), metabolic agents (MET), and risperidone monotherapy (RIS) showed lower probabilities of effectiveness, with placebo (PBO) consistently ranking as the least effective intervention.

Fig. 10.

Area under the cumulative probability ranking curve

Social withdrawal

Regarding social withdrawal, the network meta-analysis demonstrated that several interventions were significantly more effective than placebo in improving symptoms. Specifically, structured exercise (EXE) (SMD = −1.28, 95% CI −1.96 to −0.60), risperidone (RIS) (SMD = −0.95, 95% CI −1.74 to −0.16), risperidone plus adjunctive neuromodulatory agents (RIS_NM) (SMD = −0.54, 95% CI −0.95 to −0.13), and other antipsychotic or ADHD medications (OAP) (SMD = −1.19, 95% CI −2.26 to −0.12) all showed significant therapeutic advantages compared with placebo (Fig. 11).In contrast, no statistically significant differences were observed for risperidone plus adjunctive anti-inflammatory agents (RIS_AI), other monotherapies (NUT_BIO), or metabolic agents (MET), as their confidence intervals crossed zero.

Fig. 11.

Results of network Meta-analysis

The SUCRA probability ranking results further elucidated the relative hierarchy of these interventions (Fig. 12). Structured exercise (EXE) had the highest probability of being the most effective treatment for social withdrawal, followed by RIS_NM, RIS_AI, and NUT_BIO. Other antipsychotic or ADHD medications (OAP), metabolic agents (MET), and risperidone monotherapy (RIS) showed lower probabilities of effectiveness, with placebo (PBO) consistently ranking as the least effective intervention.

Fig. 12.

Area under the cumulative probability ranking curve

Publication bias test

Publication bias across the included studies was assessed using comparison-adjusted funnel plots (Fig. 13). Overall, the distributions of studies for irritability and inappropriate speech appeared relatively symmetrical around the vertical line, suggesting a low risk of publication bias.For hyperactivity and stereotypic behavior, a certain degree of asymmetry was observed, with smaller studies tending to be distributed on one side of the funnel, indicating potential small-study effects or publication bias.Similarly, the funnel plot for social withdrawal showed slight asymmetry, suggesting a possible risk of publication bias. However, given the relatively limited number of studies for some comparisons, these findings should be interpreted with caution.

Fig. 13.

Publication bias analysis of the included literature

Discussion

This network meta-analysis systematically evaluated the comparative efficacy of multiple interventions for core behavioral symptoms in children with ASD, including irritability, hyperactivity, inappropriate speech, stereotypic behavior, and social withdrawal. Overall, this study demonstrates that the effectiveness of interventions for ASD is highly domain-specific rather than uniform across behavioral outcomes. Among all interventions, structured exercise and combination pharmacological strategies (particularly risperidone combined with adjunctive agents) consistently showed relatively greater effectiveness across multiple outcomes. In contrast, monotherapies and certain metabolic or complementary interventions generally exhibited more limited or inconsistent benefits. Notably, the ranking results further supported these findings, indicating that structured exercise and combination therapies were more likely to be among the most effective treatment options, whereas placebo consistently ranked lowest across all outcomes.Despite some local inconsistencies identified in specific comparisons, the overall network structure demonstrated acceptable consistency, and the results were generally robust. These findings provide comprehensive evidence to inform clinical decision-making and highlight the potential advantages of multimodal and non-pharmacological interventions in managing ASD-related behavioral symptoms. The relatively large effect estimates observed for structured exercise, particularly for irritability and hyperactivity, should be interpreted with caution. Unlike pharmacological trials, structured exercise interventions are difficult to blind, and many exercise studies used inactive, usual-care, or waitlist control conditions. In addition, several exercise trials had relatively small sample sizes and relied on caregiver- or observer-rated behavioral outcomes, which may be susceptible to expectancy effects or detection bias. These methodological characteristics may have contributed to larger observed effects and should be considered when interpreting the apparent advantages of structured exercise in the ranking results.

A domain-specific examination of the SUCRA rankings reveals distinct therapeutic patterns across behavioral outcomes. For irritability and social withdrawal, structured exercise consistently ranked among the most effective interventions, suggesting a particularly strong benefit for affective and socially oriented symptoms. In contrast, pharmacological interventions—especially other antipsychotic or ADHD medications—demonstrated greater efficacy in managing hyperactivity and stereotypic behavior, domains that are more closely related to behavioral dysregulation and motor activity. Adjunctive strategies combining risperidone with additional agents also showed consistently high rankings across multiple domains, indicating potential advantages of multimodal approaches over monotherapy.Taken together, these findings highlight that treatment effects in ASD are not uniform but vary substantially depending on the targeted behavioral domain. This pattern supports a domain-specific and mechanism-informed treatment framework, in which pharmacological interventions may be prioritized for symptoms such as hyperactivity and stereotypy, while structured exercise may serve as a particularly effective strategy for improving irritability and social withdrawal. Importantly, the consistently lower performance of monotherapies further underscores the potential value of combined or integrative approaches. From a clinical perspective, these results advocate for individualized and symptom-targeted treatment strategies, rather than a one-size-fits-all approach, to optimize therapeutic outcomes in children and adolescents with ASD.

To further interpret these findings, particularly the superior performance of pharmacological interventions in certain domains, it is important to consider their underlying neurobiological mechanisms. Consistent with our SUCRA rankings, which identified pharmacological interventions as the most effective approach for stereotyped behaviors, the off-label use of psychotropic medications—particularly risperidone and aripiprazole—has become increasingly common in managing maladaptive behaviors among preschool-aged children with ASD [24–29].Accumulating evidence indicates that early administration of these agents can significantly improve Clinical Global Impression-Improvement (CGI-I) scores, primarily through the reduction of severe behavioral symptoms [30]. For instance, Posey et al. [31] observed that in young autistic boys, risperidone treatment resulted in marked reductions in severe aggression, irritability, and repetitive behaviors, which in turn facilitated improvements in eye contact and social interaction.Similarly, Fayyazi et al. [32] demonstrated that risperidone effectively reduced hyperactivity, disruptive behaviors, and stereotypy. These findings are consistent with our results, which showed that pharmacological interventions were particularly effective in domains such as hyperactivity and stereotypic behavior. Mechanistically, the therapeutic effects of these atypical antipsychotics are largely attributed to their antagonistic actions on postsynaptic dopamine and serotonin receptors [33]. In particular, dopamine D2 receptor blockade is closely associated with reductions in motor stereotypy, aggression, and agitation, whereas serotonin 5-HT2A receptor antagonism may contribute to improvements in restricted behavioral patterns and aspects of social communication [34–36]. However, these agents also exhibit affinity for histamine H1 receptors, which is associated with sedative effects and may limit their tolerability in clinical practice [37, 38].

In contrast, the strong performance of structured exercise observed in this study may be explained by its multifaceted physiological and psychosocial mechanisms [39]. Exercise has been shown to regulate autonomic nervous system activity, reduce stress-related hormonal responses, and enhance neuroplasticity through increased levels of brain-derived neurotrophic factor (BDNF) [40–42]. In addition, exercise-based interventions often involve structured social interaction and environmental engagement, which may be particularly beneficial for improving symptoms such as irritability and social withdrawal. These multimodal effects may account for the consistently high ranking of structured exercise in these domains, highlighting its unique role as a non-pharmacological intervention targeting both biological and behavioral dimensions of ASD. Furthermore, the superior performance of adjunctive pharmacological strategies suggests the presence of synergistic effects across multiple biological pathways [43, 44]. By combining core pharmacological mechanisms with anti-inflammatory, metabolic, or neuromodulatory actions, these approaches may achieve broader and more robust therapeutic benefits than monotherapy [45, 46]. Taken together, these findings emphasize that the effectiveness of interventions in ASD is closely linked to their underlying mechanisms and reinforce the importance of mechanism-informed, domain-specific treatment strategies.

From a clinical perspective, these findings have important implications for the management of behavioral symptoms in children and adolescents with ASD. The observed domain-specific differences in treatment efficacy suggest that intervention strategies should be tailored to the predominant symptom profile rather than relying on a uniform treatment approach. For example, pharmacological interventions may be preferentially considered for symptoms characterized by behavioral dysregulation, such as hyperactivity and stereotypic behaviors, whereas structured exercise may serve as a valuable first-line or adjunctive option for addressing irritability and social withdrawal [38, 47]. Moreover, the consistently favorable performance of combination strategies highlights the potential benefits of integrating pharmacological and non-pharmacological approaches in clinical practice. Such multimodal interventions may offer synergistic effects by simultaneously targeting multiple underlying mechanisms, thereby enhancing overall treatment effectiveness. Importantly, given the potential adverse effects associated with long-term psychotropic pharmacotherapy, incorporating structured exercise and other non-pharmacological interventions may also help reduce medication burden and enhance overall treatment tolerability. Taken together, these findings support a shift toward individualized, mechanism-informed, and domain-specific treatment planning in ASD, with an emphasis on optimizing both efficacy and safety in real-world clinical settings.

Limitation

Several limitations should be considered when interpreting the findings of this network meta-analysis.First, substantial clinical and methodological heterogeneity existed across studies, including differences in intervention types, comparator conditions, blinding feasibility, intervention duration, and outcome assessment instruments. Although outcomes were mapped onto predefined behavioral domains, different scales such as the ABC, SRS, GARS, RBS, and N-CBRF may capture overlapping but not fully identical constructs. Therefore, pooled SMD estimates should be interpreted as standardized effects across related behavioral dimensions rather than as direct evidence of identical clinical changes across scales.In particular, structured exercise trials were difficult to blind and often used usual-care, waitlist, or inactive controls, which may have contributed to relatively large effect estimates and should be considered when interpreting their rankings. Second, although potential effect modifiers were examined to assess the plausibility of transitivity, residual imbalance across intervention classes cannot be excluded, and most direct evidence came from comparisons with placebo or inactive controls rather than head-to-head comparisons between active treatments. Third, some network nodes, particularly NUT_BIO, OAP, and structured exercise, included interventions with different mechanisms or delivery formats; although further splitting of these nodes resulted in sparse or disconnected networks, findings for these nodes should be interpreted as class-level estimates rather than definitive evidence for any individual intervention. Fourth, several local inconsistencies were observed, particularly in loops involving risperidone-based interventions, and SUCRA rankings should therefore be interpreted cautiously alongside effect estimates and confidence intervals. Finally, because only English-language studies were included and unpublished or grey literature may have been missed, potential language and publication bias cannot be excluded. Further high-quality, large-scale randomized controlled trials with longer follow-up periods and more direct active-comparator designs are warranted.

Conclusion

In conclusion, this network meta-analysis provides a comprehensive and domain-specific comparison of multiple interventions for behavioral symptoms in children and adolescents with ASD. The findings demonstrate that treatment effects vary substantially across different behavioral domains, with pharmacological interventions showing greater efficacy for hyperactivity and stereotypic behaviors, while structured exercise appears particularly beneficial for irritability and social withdrawal. In addition, combination strategies consistently demonstrated favorable performance across multiple outcomes, suggesting potential advantages of multimodal approaches.These results underscore the importance of individualized, symptom-targeted, and mechanism-informed treatment strategies in ASD, rather than a uniform approach. Integrating pharmacological and non-pharmacological interventions may represent a promising direction for optimizing both efficacy and safety in clinical practice. Future research should focus on high-quality head-to-head trials and long-term follow-up to further strengthen the evidence base and guide clinical decision-making.

Clinical relevance and implications

This study highlights the importance of domain-specific treatment strategies for behavioral symptoms in autism spectrum disorder (ASD). The findings suggest that pharmacological interventions may be more suitable for managing hyperactivity and stereotypic behaviors, whereas structured exercise appears particularly beneficial for irritability and social withdrawal.These results support a shift toward individualized, symptom-targeted care rather than a uniform treatment approach. In addition, the consistently favorable performance of combination strategies indicates that integrating pharmacological and non-pharmacological interventions may enhance overall treatment effectiveness.From a clinical perspective, incorporating structured exercise into routine care may help reduce reliance on long-term medication and improve overall outcomes, supporting a more balanced and multimodal treatment framework for ASD.

Abbreviations

EXE

Structured Exercise

RIS

Risperidone

RIS_AI

Risperidone plus Adjunctive Anti-inflammatory Agents

RIS_NM

Risperidone plus Adjunctive Neuromodulatory Agents

OAP

Other Antipsychotic or ADHD Medications

MET

Metabolic Agents

NUT_BIO

Nutraceutical, Biological, or Other Monotherapies

PRO_OXT

Probiotics plus Oxytocin

PBO

Placebo

TAU

Treatment As Usual

Author contributions

[Longyan Liu]: Conceptualization, Methodology, Writing - Original Draft.[Xianyang Xin]: Data Curation, Formal Analysis. [ying zhang]: Supervision, Writing - Review & Editing.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

No datasets were generated or analysed during the current study.

Materials availability

Not applicable.

Code availability

Not applicable.

Declarations

Ethics approval and consent to participate

As this study was based on previously published data, no ethical approval or informed consent was required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.