Sensory processing differences and behavioural problems in children with autism: a retrospective study using statistical modelling and multi-output machine learning Authors: Chenyu Yan, Jialu Xu, Kexin Duan, Yifei Xiang, Haifeng Li Published: 2026 ABSTRACT AbstractBackgroundSensory processing differences (SPDs) are common in children with autism, yet their specific contributions to behavioural problems remain insufficiently explored. To examine how SPDs relate to behavioural problems in children with autism and to identify key sensory predictors and sensory-based subtypes associated with behavioural risk.MethodsThe retrospective study included 127 children with autism (1–7 years) who received rehabilitation training at a tertiary children’s hospital between 2020 and 2024. Fifteen SPD features across visual, auditory, tactile, gustatory/olfactory and vestibular/proprioceptive modalities (covering sensitivity, hyporesponsivity and seeking behaviours) and six behavioural problems were coded as binary variables based on parent-reported questionnaires. Associations were analysed using Phi correlations and logistic regression. A multi-output random forest with Shapley Additive Explanations (SHAP) evaluated the joint predictive value of SPDs. Hierarchical clustering using Jaccard distance was used to identify sensory subtypes and compare behavioural profiles.ResultsVestibular/proprioceptive seeking was positively associated with frequent tantrums (φ=0.30) and emerged as an independent predictor (OR=3.83, 95% CI 1.65 to 8.90). Visual seeking was negatively associated with difficulty adapting to routine changes (φ=−0.23; OR=0.34, 95% CI 0.15 to 0.78). The multi-output random forest showed moderate performance. SHAP analysis highlighted auditory sensitivity and multiple seeking behaviours as major contributors. Clustering revealed a ‘multi-sensory seeking’ subtype with higher rates of tantrums.ConclusionsSPDs are meaningfully linked to behavioural problems in autism. Vestibular/proprioceptive seeking and auditory sensitivity are key behavioural risk indicators, while visual seeking may support adaptability. Sensory-based profiling has the potential to identify children at elevated behavioural risk and guide personalised intervention planning.Keywords: Machine Learning, Child WHAT IS ALREADY KNOWN ON THIS TOPIC Sensory processing differences are linked to behavioural problems in autistic children. Sensory processing differences are linked to behavioural problems in autistic children. WHAT THIS STUDY ADDS This study identifies distinct behaviour-specific sensory associations and highlights auditory hyper-responsivity and multisensory seeking as key contributors to behavioural risk. This study identifies distinct behaviour-specific sensory associations and highlights auditory hyper-responsivity and multisensory seeking as key contributors to behavioural risk. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY These findings support more modality-focused sensory assessment and inform the development of sensory-informed screening and intervention approaches. These findings support more modality-focused sensory assessment and inform the development of sensory-informed screening and intervention approaches. Introduction Autism spectrum disorder (ASD) is a lifelong neurodevelopmental disorder characterised by difficulties in social communication and repetitive or restricted behavioural patterns.1 Behavioural problems, especially externalising behaviours such as aggression and self-injury, are highly prevalent among children with ASD and have a devastating impact on their adaptive functioning and social acceptance.2 3 Understanding the underlying mechanisms of these behavioural problems is therefore a key priority for clinical assessment and early intervention. Although extensive studies and clinical interventions have focused on high-level contributors such as impaired cognitive and social functioning,4 5 evidence from studies on neurodevelopmental disorders (eg, attention-deficit/hyperactivity disorder) shows that even children with relatively intact social abilities also experience significant behavioural problems.6 This highlights the need to consider some lower-level contributors more closely related to neurobiological underpinnings, such as sensory processing differences (SPDs). SPDs include atypical sensory responses—such as hyper-responsiveness (eg, covering ears to moderate sounds), hyporesponsivity (eg, ignoring name calling) or seeking behaviours (eg, repeatedly turning lights on/off for visual stimulation) across multiple modalities such as vision, hearing, touch, taste/smell and vestibular/proprioceptive systems.7 8 SPDs are highly prevalent in children with neurodevelopmental disorders, especially for ASD, and have become an ASD diagnostic criterion in DSM-5 (Diagnostic and statistical manual of mental disorders, Fifth edition).1 This atypical way of perceiving the world can in turn give rise to atypical behavioural responses. Early studies have reported a general correlation between SPDs and greater sensory difficulties, which have been linked to more substantial social challenges and reduced adaptive functioning.9 10 More recently, a three-level meta-analysis including over 11 000 autistic participants reported that higher levels of all major SPD subtypes were associated with greater internalising and externalising problems; hypersensitivity and sensory seeking showed particularly strong links with externalising difficulties.11 These findings illustrate a shift from demonstrating general correlations to examining more fine-grained sensory subtypes and their distinct behavioural correlates. However, important gaps remain. First, most studies have examined the link between SPDs in a single modality and a single behavioural problem,11 which may underestimate joint contributions from multiple modalities and shared sensory mechanisms underlying multiple co-occurring behaviours. Second, although multiple works have identified distinct sensory subtypes in autism,12 13 few studies have explored whether such sensory-based patterns correspond to specific profiles of everyday behavioural problems in clinical samples. Third, research involving Chinese samples remains limited, even though cultural factors such as higher expectations for behavioural conformity might influence how behaviours are interpreted and reported. Clarifying how concrete sensory features and their combinations relate to common behavioural problems in ASD is crucial for refining risk assessment and tailoring interventions. The present study therefore aimed to: (1) examine associations between multiple SPD features and a set of clinically relevant behavioural problems in children with ASD; (2) evaluate the joint predictive value of SPD for these behaviours using multi-output machine learning (ML) with explainable models; and (3) identify sensory-based subtypes and compare their behavioural risk profiles. By adopting a multidimensional, data-driven approach in a real-world clinical cohort, we sought to generate clinically interpretable evidence on the sensory correlates of behavioural difficulties in ASD. Methods Participants This retrospective observational study reviewed clinical records and parent-reported questionnaires from children who received rehabilitation between January 2020 and December 2024 in the Rehabilitation Department of the Children’s Hospital, Zhejiang University School of Medicine. Children were eligible if they received a clinical diagnosis of ASD based on DSM-5 criteria and if complete sensory processing and behavioural questionnaire data were available at intake. Exclusion criteria included known genetic syndromes (such as Rett syndrome, fragile X syndrome or trisomy 21), metabolic disorders, major neurological conditions, congenital hearing or vision impairment, degenerative or disintegrative disorders and incomplete clinical or questionnaire records. A total of 127 children met the inclusion criteria and were included in the study. Patient and public involvement Patients or members of the public were not involved in the design, conduct, reporting or dissemination plans of this research. This retrospective study used existing clinical records and parent-reported questionnaires, and it was therefore not appropriate or feasible to involve patients or the public in the research process. Measures SPDs were assessed using a parent-reported questionnaire adapted from the Sensory Experiences Questionnaire, V.2.1 (SEQ, V.2.1),14 which characterises sensory responses across multiple modalities in children and has demonstrated good internal consistency and excellent test–retest reliability for children with ASD or developmental delay ages 5–72 months.15 From the questionnaire, 15 binary SPD features were derived, encompassing hyper-responsivity, hyporesponsivity and sensory seeking across the visual, auditory, tactile, gustatory/olfactory and vestibular–proprioceptive modalities. Each feature was coded as present or absent based on parental endorsement, providing a multidimensional profile of each child’s sensory processing pattern. Behavioural problems were derived from parent-reported items included in the routine clinical intake questionnaire. The six parent-reported items capture commonly reported behavioural difficulties in daily life, including difficulty adapting to changes in routine, frequent tantrums, aggression, self-injury, crying during activities and elopement; each behaviour was coded as a binary variable based on parental endorsement to indicate its presence or absence. Cases with incomplete clinical or questionnaire records were excluded during participant selection. Statistical analysis All analyses were conducted in Python V.3.9 using standard scientific, statistical and ML libraries. Statistical analyses were conducted to examine both individual and combined associations between SPD and behavioural problems, which were conducted in three stages: Traditional statistical methods were used to examine direct associations between SPD and behavioural problems. Phi coefficients were calculated to assess pairwise relationships between each binary SPD feature and each behavioural outcome. To further evaluate the predictive contribution of individual sensory features, univariate logistic regression (ULR) models were fitted for each behavioural problem, followed by multivariable models including all SPD features and child age to identify independent predictors. Traditional statistical methods were used to examine direct associations between SPD and behavioural problems. Phi coefficients were calculated to assess pairwise relationships between each binary SPD feature and each behavioural outcome. To further evaluate the predictive contribution of individual sensory features, univariate logistic regression (ULR) models were fitted for each behavioural problem, followed by multivariable models including all SPD features and child age to identify independent predictors. A multi-output random forest classifier was constructed using all SPD indicators and age as predictors to assess the joint influence of multiple SPD features across all behavioural outcomes simultaneously. Model performance was evaluated using fivefold cross-validation, with Micro-F1, Macro-F1 and Hamming Loss used to quantify predictive accuracy. To aid interpretation of the model, Shapley Additive Explanations (SHAP) values were computed to estimate the relative contribution of each SPD feature to each behavioural label. A multi-output random forest classifier was constructed using all SPD indicators and age as predictors to assess the joint influence of multiple SPD features across all behavioural outcomes simultaneously. Model performance was evaluated using fivefold cross-validation, with Micro-F1, Macro-F1 and Hamming Loss used to quantify predictive accuracy. To aid interpretation of the model, Shapley Additive Explanations (SHAP) values were computed to estimate the relative contribution of each SPD feature to each behavioural label. Clustering analyses were conducted to explore whether combinations of SPD characteristics formed distinct sensory-based subgroups. K-means clustering was initially applied to examine general clustering tendencies. Given the binary nature of the SPD variables, hierarchical clustering using Jaccard distance (for binary variables) and average linkage was then performed to derive the final cluster solution. Differences in behavioural outcomes across the resulting clusters were assessed using χ2 tests. Clustering analyses were conducted to explore whether combinations of SPD characteristics formed distinct sensory-based subgroups. K-means clustering was initially applied to examine general clustering tendencies. Given the binary nature of the SPD variables, hierarchical clustering using Jaccard distance (for binary variables) and average linkage was then performed to derive the final cluster solution. Differences in behavioural outcomes across the resulting clusters were assessed using χ2 tests. Results Participants characteristics A total of 127 children were included in the study, of whom 104 (81.9%) were boys. The mean age was 3.4±1.1 years. 87.7% children were born at term and had no major perinatal complications. Parental education levels were generally distributed from junior high school to postgraduate degrees, with bachelor′s degrees (39.0% of father and 43.1% of mother) being the most common. The sociodemographic and clinical characteristics of the participants are presented in table 1. Table 1Sociodemographic and clinical characteristics of the participants (n=127)CharacteristicTotal sampleAge (years) (n=127), mean±SD3.4±1.1Sex (n=127), n (%) Male104 (81.9%) Female23 (18.1%)Preterm birth (n=106), n (%) Yes13 (12.3%) No93 (87.7%)Paternal education level (n=123), n (%) Junior high school or below9 (7.3%) Senior high school13 (10.6%) College diploma22 (17.9%) Bachelor’s degree48 (39.0%) Postgraduate degree (Master’s/PhD)31 (25.2%)Maternal education level (n=123), n (%) Junior high school or below10 (8.1%) Senior high school10 (8.1%) College diploma27 (22.0%) Bachelor’s degree53 (43.1%) Postgraduate degree (Master’s/PhD)23 (18.7%) Sociodemographic and clinical characteristics of the participants (n=127) Prevalence of SPD SPDs were common (115 cases, 90.6%) in these ASD children. The most prevalent features were vestibular/proprioceptive seeking (83 cases, 54.3%), tactile seeking (59 cases, 46.5%), visual seeking (46 cases, 42.5%) and auditory sensitivity (45 cases, 36.2%). Other sensory sensitivities or hyporesponsivity patterns occurred less frequently. Overall, children demonstrated heterogeneous sensory profiles, with many presenting combinations of seeking behaviours and heightened sensitivity (see figure 1a). Although 15 sensory processing features were initially derived from the questionnaire, four features had no positive cases in the present sample and therefore were not included in further analysis. Distribution of SPDs and behavioural problems. (a) Distribution of SPDs. Each cell indicates the number of children exhibiting a specific sensory feature. (b) Distribution of six behavioural problems. SPDs, sensory processing differences. Prevalence of behavioural problems Behavioural difficulties were also frequently reported (see figure 1b). The most common problems were frequent tantrums (55 cases, 43.3%) and difficulty adapting to routine changes (52 cases, 40.9%), followed by aggression (38 cases, 29.9%). Less frequent behaviours included crying during activities (18 cases, 14.2%), self-injury (10 cases, 7.9%) and elopement (6 cases, 4.7%). These patterns indicate that emotional reactivity and adaptability difficulties were widespread, whereas more severe externalising behaviours were comparatively less common. Associations between SPDs and behavioural problems The associations between SPDs and behavioural problems are summarised in table 2. Overall, most sensory features showed negligible or small associations with behavioural outcomes. Several statistically significant but modest correlations were observed. Visual seeking was associated with a lower likelihood of difficulty adapting to routine changes (φ –0.228, p=0.021). Tactile hyper-responsivity showed a positive association with aggression (φ=0.178, p=0.034). Within the gustatory/olfactory modality, hyper-responsivity was associated with a higher probability of self-injury (φ=0.211, p=0.017). The strongest association emerged between vestibular–proprioceptive seeking and tantrums (φ=0.302, p=0.002). Table 2Phi correlations between SPDs and behavioural problemsSPD featureDifficulty adaptingTantrumsAggressionSelf-injuryCryingElopementVisual Hyper-responsivity–0.018–0.108–0.028–0.084–0.0060.034 Seeking–0.228*–0.000–0.063–0.0190.114–0.088Auditory Hyper-responsivity0.019–0.0180.0190.0980.0490.029 Seeking–0.0430.082–0.026–0.0770.0610.048Tactile Hyper-responsivity0.1250.1060.178*0.096–0.107–0.025 Seeking0.0270.0330.012–0.0850.060.014Taste/smell Hyper-responsivity–0.169–0.1040.0450.211*0.1570.037 Hyporesponsivity0.0070.0420.036–0.0870.035–0.111 Seeking0.0450.090.0470.0360.066–0.013Vestibular/proprioception Hyporesponsivity–0.008–0.0060.0480.0820.040.126 Seeking0.101 0.302** 0.0420.1220.0540.137*p<0.05; **p<0.01.SPDs, sensory processing differences. Phi correlations between SPDs and behavioural problems *p<0.05; **p<0.01. SPDs, sensory processing differences. ULRs identified three significant sensory-behaviour associations (table 3). Visual seeking significantly predicted lower odds of difficulty adapting to change, vestibular–proprioceptive seeking was strongly associated with higher odds of tantrums and taste/smell hyper-responsivity significantly increased the odds of self-injury. These patterns align with the correlation analyses, indicating that only a limited subset of sensory features shows stable, behaviour-specific associations. Table 3Significant predictors from univariate logistic regressionSPD featureBehavioural problemOR95% CIP valueVisual: seekingDifficulty adapting0.3620.164 to 0.7960.012Vestibular/proprioception: seekingTantrums3.9171.745 to 8.793<0.001Taste/smell: hyperSelf-injury7.4671.114 to 50.0480.038 Significant predictors from univariate logistic regression Overall predictive performance and feature contributions To assess the predictive value of sensory features for behavioural problems, we built a multi-output random forest model using 11 sensory features and age as predictors. The model demonstrated moderate performance (Micro-F1=0.36; Macro-F1=0.30; Hamming loss=0.35), with ‘tantrums’ and ‘difficulty adapting to change’ showing the highest label-specific accuracy. SHAP analysis revealed distinct contribution patterns across behavioural outcomes (figure 2). Auditory hyper-responsivity showed consistent, moderate-to-high contributions across outcomes, whereas sensory seeking features displayed more domain-specific influences—for example, visual seeking for adaptability and vestibular–proprioceptive seeking for tantrums. Hyper-responsivity and hypo-responsivity features contributed less overall but remained relevant for particular behaviours. When aggregated across outcomes, auditory hyper-responsivity emerged as the most influential feature, followed by visual, vestibular–proprioceptive, tactile and taste/smell seeking (figure 3), suggesting that auditory sensitivity and multisensory seeking may represent core sensory characteristics linked to multiple behavioural problems in ASD. SHAP feature contributions to individual behavioural problems. SHAP, Shapley Additive Explanations; SPD, sensory processing difference. Overall feature importance aggregated across behavioural problems. RF, Random Forest. Sensory processing clusters and behavioural differences To explore whether combinations of sensory features formed meaningful subgroups associated with behavioural differences, unsupervised clustering analyses were conducted. We first applied K-means clustering to the 11 binary sensory features and identified a three-cluster solution. The clusters were characterised as follows: (1) a Low Atypicality group, generally low rates of sensory atypicalities across modalities; (2) a Multisensory Seeking group, defined by elevated visual, tactile and vestibular/proprioceptive seeking behaviours; and (3) a Sensory Sensitivity group, characterised primarily by auditory and visual hyper-responsivity. χ2 tests indicated significant group differences in the distribution of frequent tantrums (χ² = 7.38, p=0.025) and a trend-level difference for discomfort after routine change (χ² = 4.97, p=0.084). The Multisensory Seeking cluster showed the highest proportion of tantrums (approximately 53%), whereas the Low Atypicality group showed the lowest (approximately 25%). The distributions for aggression, crying during activities, elopement, and self-injury did not significantly differ across clusters. The behavioural characteristics of the three sensory clusters are summarised in table 4. Table 4Distribution of behavioural problems across sensory processing clustersBehavioural problemLow atypicality (%)Multisensory seeking (%)Sensory sensitivity (%)χ²P valueDiscomfort after routine change3858.366.74.970.084Frequent tantrums2552.966.77.380.025*Aggression2737.533.31.10.577Crying during activities1833.333.32.780.249Elopement1529.233.32.610.271Self-injury812.500.740.692*p < 0.05 Distribution of behavioural problems across sensory processing clusters p < 0.05 Discussion This study identified differentiated patterns of associations between SPD and behavioural problems in autistic children. Specific sensory features showed behaviour-specific associations (eg, vestibular–proprioceptive seeking with tantrums and visual seeking with reduced discomfort after routine change). Multi-output ML further revealed broader cross-domain contribution patterns, highlighting the central role of auditory hyper-responsivity and the consistent involvement of sensory-seeking behaviours. Unsupervised clustering identified a multisensory seeking subgroup with notably elevated behavioural risk. Together, these findings suggest that early auditory sensitivity and multisensory seeking may serve as key sensory profiles linked to behavioural problems, potentially reflecting shared underlying mechanisms. A major strength of this study is the use of complementary analytic approaches. Traditional correlation and regression analyses identified behaviour-specific sensory associations, while the multi-output ML model revealed cross-domain contribution patterns that cannot be detected by single-outcome analyses. The use of SHAP values added interpretable feature attribution,16 and unsupervised clustering highlighted meaningful combinations of sensory features that differentiated behavioural risk profiles. Key limitations include reliance on parent-reported measures, dichotomisation of behavioural outcomes, a modest sample size for multilabel modelling, low base rates for certain behaviours and the cross-sectional, single-centre design, which limits causal inference and generalisability. Previous research has demonstrated associations between SPDs and behavioural problems in autistic children.11 17 Among our major findings, the positive association between vestibular–proprioceptive seeking and tantrums aligns with recent reports linking heightened movement-based sensory seeking to externalising difficulties.11 In contrast, the negative association between visual seeking and difficulty adapting to routine changes has been less frequently described. One possible explanation is that active visual exploration may reduce uncertainty and enhance environmental predictability,18 thereby supporting adaptability in daily routines. In addition, few studies have systematically compared the differential contributions of sensory modalities. Our findings highlight the prominence of auditory hyper-responsivity and multisensory seeking across behavioural domains, consistent with evidence that auditory and motor-sensory processing atypicalities may be closely connected to behavioural dysregulation in autism.17 Building on earlier clustering studies that have shown potential of integrating sensory features to identify ASD subgroups19 and to reveal latent structures,20 the present study further highlights a multisensory high-seeking subgroup showing elevated behavioural risk. This pattern may reflect a risk-aggregation effect, whereby simultaneous high seeking across multiple modalities contributes to heightened arousal and lower regulatory thresholds, increasing the likelihood of explosive reactions. The sex distribution in the current study (81.9% male) was broadly consistent with epidemiological data (~82%),21 whereas the small number of females limited our ability to examine sex-specific effects.22 Accordingly, sex was considered only as a covariate in exploratory analyses and did not show a significant main effect. Overall, our findings extend prior work by emphasising the modality-specific and cross-modality differences in predicting behavioural difficulties, underscoring the value of analysing both individual sensory features and their aggregated patterns. Overall, the present analyses provided modality-specific understanding of the association between SPDs and behavioural problems. The differentiated sensory-behaviour patterns identified here highlight the value of assessing specific sensory features rather than global scores. Specifically, vestibular–proprioceptive and visual seeking, along with auditory hyper-responsivity, may be particularly informative for understanding tantrums and adaptability. ML results further suggest that multisensory seeking profiles can indicate elevated behavioural risk, supporting sensory-informed screening and intervention planning. Although the sample size is modest for modelling approaches involving multiple predictors, which may introduce potential overfitting and model instability, these analyses provide exploratory insights into modality-specific and multidimensional sensory-behaviour patterns. Future studies should adopt longitudinal designs to clarify temporal pathways between sensory processing and behavioural outcomes. Multimethod assessments beyond parents’ reports will help illuminate underlying mechanisms. Larger, multicentre samples are required to validate sensory subgroups and examine potential moderators. Integrating multimodal data into predictive models may enhance risk identification and support more precise intervention strategies. Conclusion Specific sensory features, particularly vestibular–proprioceptive seeking, visual seeking and auditory hyper-responsivity are differentially associated with behavioural problems in autistic children. Sensory feature combinations further distinguish behavioural risk profiles. Longitudinal and multi-method studies are needed to clarify underlying pathways and inform targeted interventions. Footnotes Funding: This work was supported in part by the National Natural Science Foundation of China (62306270 to JX and 32500925 to CY), in part by the Key Research and Development Program of Zhejiang Province (2023C03003 to HL) and in part by Autism Research Special Fund of Zhejiang Foundation For Disabled Persons (2025007 to CY). Patient consent for publication: Not applicable. Ethics approval: The study was approved by the Ethics Committee of the Children’s Hospital, Zhejiang University School of Medicine (2024-IRB-0025-P-01). Participants gave informed consent to participate in the study before taking part. Provenance and peer review: Part of a topic collection; not commissioned; externally peer reviewed. Patient and public involvement: Patients and/or the public were not involved in the design, or conduct, or reporting or dissemination plans of this research. Data availability free text: The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Data availability statement Data are available upon reasonable request. References Data Availability Statement Data are available upon reasonable request.