Protocol for the development of joint attention-based subclassification of autism spectrum disorder and validation using multi-modal data

Genetic factors are the primary risk contributors to ASD, encompassing a range of both rare and common genetic variants. This includes rare de novo mutations, copy number variants, protein-truncating single nucleotide polymorphisms (SNPs), indels, and even non-coding de novo mutations in chromatin interactions [17, 18]. Previous studies have linked these genetic elements to the wide-ranging clinical manifestations of ASD. The genetic origins of ASD are complex and multifactorial; however, our grasp of how these genetic elements functionally influence ASD is still limited. Additionally, the way these genetic factors interact with each other and non-genetic factors to shape risk remains uncertain. However, considering the enormous number of variants implicated in ASD, coupling these genetic data with other types of data obtained through different modalities may help us discern specific genetic vulnerabilities and their corresponding functional consequences in ASD [19].

A few studies have explored neuroanatomical differences between individuals with ASD and neurotypical individuals, as well as within the ASD group itself [11‐13, 20]. A significant portion of this research has been conducted using the Autism Brain Imaging Data Exchange (ABIDE) dataset, an open-source compilation of brain imaging data from ASD and neurotypical individuals [21, 22]. Using this resource, Hong et al. identified three distinct ASD subclasses with varying clinical outcomes, based on cortical thickness, intensity contrast, surface area, and geodesic distance [20]. Choi et al., applying an unsupervised clustering approach to ABIDE’s resting state fMRI (rsfMRI) data and utilizing ‘connectome-based gradient’ and ‘functional random forest’ methodologies, were able to differentiate various forms of ASD [12]. However, their results lacked complete replicability as the clinical scores of the subgroups identified through cluster analysis in a follow-up dataset deviated from those in the original dataset. Additionally, these ASD clusters failed to show significant differences in neural connectivity [12, 20]. These outcomes imply that purely unsupervised clustering using intricate data like fMRI might not yield clinically meaningful ASD subtypes.

A research team adopted a top-down approach in their pursuit of developing a new ASD subclassification method, using a clinical-score based clustering method and multimodality data [23]. While this approach allowed for the identification of distinct ASD subgroups based on balance between social-communicative and restricted repetitive behaviors – the two main symptom domains of ASD), these ASD subgroups showed no evident differences in neural circuitry and genetic components [23].

Previous attempts at cluster analysis using unsupervised learning on intricate data such as fMRI did not yield consistent or replicable subclassifications; even with identical experimental parameters, the replication dataset did not reproduce the same ASD subgroups as the original discovery dataset [11‐13]. Without any labels to guide the clustering, individuals were grouped into distinct subclusters based on non-linear data patterns that lacked clinical significance. Meanwhile, studies using supervised learning methods for cluster analysis also fell short, failing to identify subgroups with unique clinical phenotypes and neurobiological mechanisms [23]. It is worth noting that these studies used human-rated autism-related scale sub-scores as labels for supervised learning. Although these scores are clinically valid and frequently used, they may not adequately capture objective and qualitative differences in ASD phenotypes.

Joint attention, as indicated by prior research, may serve as an objective behavioral biomarker for ASD subclassification [24, 25]. This term refers to the sharing of attentional focus with others on objects or events and is believed to aid social learning [26, 27]. We have established a digitized method to assess joint attention, implementing a novel protocol to prompt three types of joint attention - initiation of joint attention (IJA), low-level and high-level response to joint attention (RJA_low and RJA_high) - as defined in the Early Social Communication Scales (ESCS) [28]. This method uses video-recorded task-related behaviors [26, 27], which are then utilized to train a deep learning model to identify ASD and evaluate its symptom severity [25].

Identifying the most affected level—biological, neural, or behavioral—in an individual with Autism Spectrum Disorder (ASD) to determine the suitable targeted intervention is a significant challenge. To address this, we have initiated the development of the Yonsei-Seoul Multi-modal Subclassification (YSMS) protocol, which was designed to streamline data collection from newly diagnosed ASD patients. Specifically, the focus is on the patterns of joint attention gaze demonstrated during behavioral tasks, with the aim to discern whether these patterns can differentiate various ASD subtypes, each unique in their biological and neurological attributes. We intend to implement a pilot version of this protocol in an exploratory study involving 60 individuals diagnosed with ASD from June 2023 to May 2024 to compare objective behavioral biomarker-based ASD subgroups using multimodality data. Such endeavor is expected to drive the creation of a novel ASD subclassification system, considering each individual’s biological, neural, and behavioral characteristics.

We have designed a single center, prospective, non-randomized experimental study, where individuals with ASD diagnosis will be recruited from June 2023 – May 2024. A group of 60 ASD individuals will be recruited from outpatient clinic at the Child and Adolescent Psychiatry Division of Seoul National University Hospital (SNUH). Enrollment criteria are: (1) age of 30 ~ 71 months, (2) children diagnosed with ASD by a psychiatrist and confirmed by Autism Diagnostic Observation Schedule II (ADOS2). Exclusion criteria are: (1) history of organic brain injury, (2) having visual or auditory deficit, (3) having neurological (motor/musculoskeletal) conditions, (4) history of adverse reaction to sedation anesthesia. A schematic showing the flow of patients is depicted in Fig. 1.

Standard gradient-echo echo planar imaging paradigm will be utilized for the acquisition of resting state fMRI scans [39]. We will carry out several procedures to find optimal parameters such as field of view, repetition time, echo time, flip angle, etc. During the resting state scan, participants will be sedated with their eyes closed. The Configurable Pipeline for the Analysis of Connectomes (CPAC, http://​fcp-indi.​github.​com) preprocessing pipeline will be employed for fMRI data processing to remove non-neural noise, such as motion artifacts, from the BOLD-signal [12, 20]. Multiple functional networks have been identified that are characterized by coherent patterns of intrinsic activity between ‘nodes’ that resemble patterns of activity that are engaged during specific cognitive functions [12]. We aim to identify whether social gaze-based subgroups differ in (hyper/hypo-) connectivity within and across these networks [10, 11, 13].

This study, adhering to the Helsinki Declaration and its subsequent amendments, mandates informed consent from participants and their legal guardians based on verbal and written details provided. Participants and their legal guardians reserve the right to withdraw at any point, with no obligation to justify their decision and no impact on future treatment. Approval for the study has been granted by the Seoul National University Hospital IRB Review Board (IRB No. H-2210-137-1374). The investigators anticipate no discomfort for the subjects from the tests and tasks involved, with no short- or long-term risks identified in relation to this study.

We posit that ASD individuals with decreased joint attention, characterized by atypical social gaze patterns, will exhibit unique genetic, structural, and functional neural patterns distinct from ASD individuals with mostly intact joint attention, who demonstrate appropriate social gaze patterns. The differentiation is likely rooted in the connectivity within the visual and social brain pathways. It is anticipated that ASD individuals with decreased joint attention will possess SNPs associated with the visual pathway. Furthermore, individuals with greater RRB than SC impairment are expected to exhibit less impairment in social gaze ability. Conversely, participants with a higher SC impairment relative to RRB are predicted to demonstrate a pronounced deficiency in social gaze ability. A layout of expected outcome table is depicted in Table 2.