Early diagnosis of autism spectrum is crucial in improving patients' quality of life. The use of machine learning techniques is accelerating and making diagnostic processes more accurate, while more inclusive research is essential to develop personalized and targeted treatments.
In recent years, research into Autism Spectrum Disorders and machine learning techniques have increasingly converged, with a growing number of international research teams focusing on two main areas. On the one hand, efforts are being made to refine early diagnosis, considered essential for initiating highly personalised treatment plans that can alleviate autism-related symptoms and provide better support to those affected.
On the other hand, studies are delving into the «molecular and cellular mechanisms involved in this condition which, from a biological perspective, is now widely recognised by the scientific community as a neurodevelopmental disorder of genetic origin».
In this article, we explore the directions taken by the latest studies and the advancements achieved so far, aiming to provide an overview of the outcomes and new research perspectives in this ever-evolving field.
TAKEAWAYS
What are Autism Spectrum Disorders?
The World Health Organization (WHO) defines Autism Spectrum Disorder (ASD) as a heterogeneous group of brain development conditions, characterised by difficulties in communication and social interaction, as well as repetitive or atypical behaviours and activities. These behaviours can include challenges in transitioning from one activity to another, an intense focus on details, and unusual reactions to sensory stimuli. Autism can be detected in early childhood, although diagnosis is often made later in life.
According to WHO estimates, approximately 1 in 100 children worldwide are affected by autism. However, this is an average figure, and prevalence rates can vary significantly between studies and geographic regions. In low- and middle-income countries, for instance, specific data on autism may be less accurate or simply not well-documented.
Individuals with ASD display a wide range of abilities and needs, which may change over time. Some people with autism are able to lead independent lives, while others experience severe disabilities and require lifelong support.
The WHO highlights an important point, one that underlines the value of ongoing research: Autism Spectrum Disorders can impact educational and employment opportunities, and families caring for individuals with autism often face significant challenges. Moreover, the quality of life for people with autism is influenced by societal attitudes and the level of support provided by local and national authorities.
To improve the health and well-being of individuals with autism, the WHO recommends evidence-based psychosocial interventions that enhance communication and social skills. It also emphasises the need for targeted community actions to promote greater accessibility, inclusion, and support.
The importance of early diagnosis
When it comes to Autism Spectrum Disorders, early diagnosis is crucial and in recent years, machine learning has begun to offer a promising path towards achieving quicker and more accurate diagnoses.
Despite advances in understanding autism, there is still no reliable biomarker for diagnosis, which remains primarily based on observing behavioural symptoms. Autism can often be detected by 14 months, and more reliably by the age of two. However, diagnosis typically occurs only after children are four and a half years old, when they are already in school. Diagnosing ASD before the age of three is vital, as early intervention can significantly enhance a child’s development.
The challenge, therefore, lies in identifying methods for early diagnosis and objectively quantifying behavioural atypicalities. In this context, artificial intelligence is proving to be a valuable tool, helping monitor movements and behaviours during standardised tests used to assess autism. Let us now explore the direction recent studies in this field are taking.
The use of machine learning techniques
A study titled “An evaluation of machine learning approaches for early diagnosis of autism spectrum disorder”, published this year, was conducted by a multidisciplinary team of researchers from universities in Kushtia, Khulna, and Savar in Bangladesh, and from the University of Missouri and Troy University in the United States. The study delved deeply into the use of various machine learning approaches to enhance early diagnosis of Autism Spectrum Disorders (ASD).
Traditional ASD diagnosis, which relies on interviews and behavioural observations, is often slow and resource-intensive. As a result, many patients receive their diagnosis later than ideal, delaying the start of necessary therapies.
To address this issue, the researchers applied eight classification models and five clustering methods to datasets containing information on both children and adults with ASD. Their analysis focused on metrics such as accuracy, precision, recall, and the Area Under the Curve (AUC) to assess the effectiveness of the models. (Note: The AUC evaluates a machine learning model’s ability to assign higher scores to positive examples than negative ones. Since it is independent of decision thresholds, the AUC provides a measure of a model’s accuracy without the need for a specific cut-off value.).
The researchers also acknowledged a critical point: while global rates of Autism Spectrum Disorders are rising, there is still a significant lack of publicly available datasets specifically tailored to autism research. Most existing databases focus on genetics, with a shortage of clinical datasets for autism screening. The datasets used in this study, “ASD in Children” and “ASD in Adults,” were sourced from the public UCI Repository. The children’s dataset contains 292 cases of individuals aged between 4 and 11 years, while the adult dataset has 704 cases, both with 21 attributes.
Both datasets include responses to a 10-question screening tool known as AQ-10 (Autism Spectrum Quotient), designed to determine whether an individual should undergo a full assessment for autism. Each question is scored 0 or 1, with a higher total score indicating a greater likelihood of ASD, suggesting the need for further evaluation.
The questions cover behavioural aspects such as communication, attention to detail, social interaction, and imagination. In addition to the questionnaire responses, the datasets contain information like age, gender, ethnicity, whether the participant had jaundice at birth, a family history of pervasive developmental disorders, and other participant characteristics. This approach highlights the potential of machine learning to streamline autism screening processes, offering more rapid and accurate results than traditional methods.
Data classification and analysis of the relationship between variables
Among all the models examined in the study, Support Vector Machine (SVM) and Logistic Regression (LR)delivered the best performance, achieving 100% accuracy for the children’s dataset and 97.14% for the adult dataset.
Support Vector Machine (SVM) is a machine learning method used for data classification. It works by finding an optimal boundary, known as the “hyperplane,” which separates data into two categories. In the case of multidimensional data, this hyperplane maximises the distance between the categories, ensuring the best possible separation.
Logistic Regression, on the other hand, is a machine learning algorithm that analyses the relationship between variables and predicts a discrete outcome, such as “yes” or “no”. Unlike linear regression, which predicts continuous values, logistic regression uses a logistic function to estimate the probability of a particular event occurring, allowing data to be categorised into defined groups.
Additionally, the researchers developed an Artificial Neural Network (ANN) model that achieved 94.24% accuracy on a combined dataset, following careful parameter optimisation.
One of the innovative aspects of the study was the use of clustering methods to analyse data where precise labels were not available. This approach allowed the researchers to evaluate algorithm performance in real-world scenarios, where information is not always fully defined. In particular, spectral clustering – a machine learning technique based on spectral graph theory and linear algebra – proved to be the most effective method. Its goal is to divide a graph into smaller groups by clustering similar or closely related values. Spectral clustering outperformed reference models like k-means in terms of Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI).
The study went beyond theoretical analysis, as the team developed a platform with an integrated graphical interface to allow for more practical and immediate clinical application of these models. The researchers expressed their intention to continue their work using larger datasets, with the ultimate goal of making ASD diagnosis faster, more accurate, and more accessible.
This would reduce the time and costs associated with traditional diagnostic methods, enabling doctors and specialists to intervene sooner, thereby improving the effectiveness of early interventions for children diagnosed at a young age.
Federated learning and machine learning: allies in autism research
Similar conclusions were drawn by a study published in November last year by a Pakistani research team led by Muhammad Shoaib Farooq, alongside Rabia Tehseen, Maidah Sabir, and Zabihullah Atal. This study combined Federated Learning (FL) and Machine Learning (ML) for the early diagnosis of Autism Spectrum Disorder (ASD) in children and adults.
A key feature of this study was the use of Federated Learning, which allows machine learning models to be trained without sharing sensitive data between organisations. The researchers employed the same two ML models identified as the best performers in the previous study: Logistic Regression (LR) and Support Vector Machine (SVM).
These models were trained locally on four different datasets containing information about children and adults with ASD. The results were then sent to a central server, where a meta-classifier was trained to determine the most accurate approach for detecting ASD. The proposed model achieved an accuracy of 98% for children and 81% for adults, demonstrating the effectiveness of FL in early autism diagnosis while maintaining data security and confidentiality.
Using statistical learning for Autism Spectrum Disorders diagnosis
Another study, published this year by Italian researchers Roberta Bettoni, Chiara Cantiani, Elena Maria Riboldi, Massimo Molteni, Hermann Bulf, and Valentina Riva in Plos One, also offers important insights into ASD diagnosis. Their findings suggest that autism spectrum disorders may begin to manifest as early as six months of age. The team examined Visual Statistical Learning (VSL), the brain’s ability to recognise patterns and regularities in the environment.
VSL refers to how the brain learns the probabilities with which visual forms or objects appear together in complex configurations, simply by observing them. For example, when viewing a scene, the brain can pick up on patterns, such as which objects frequently appear together or in specific sequences. Research shows that even infants are sensitive to these visual regularities and can learn them from the earliest months of life.
In this study, VSL was tested in a group of children at high risk for autism (siblings of children with ASD) and a control group. The high-risk children displayed greater difficulty recognising visual regularities compared to neurotypical children. These difficulties could impact their social and communication skills between the ages of 24 and 36 months.
The findings suggest that early statistical learning challenges may serve as a predictive marker for autism. Despite the limitations of the study, the researchers emphasise the importance of identifying these early signals to improve personalised interventions and prevent the development of social and communication difficulties.
Glimpses of Future
The research described has shown how machine learning techniques, applied through different approaches and methodologies, can be of great help in the early diagnosis of Autism Spectrum Disorders in children, from early infancy, and even in undiagnosed adults.
Now, with the aim of anticipating possible future scenarios, let’s use the STEPS framework to analyse what the impacts of these studies could be in the future – thanks also to the greater availability of data – and what their reflections might be in terms of social, technological, economic, political, and sustainability aspects.
S – SOCIAL: early diagnosis of autism can pave the way for timely interventions, such as therapies that improve children’s communication and social skills, helping to enhance their quality of life over time. These intervention programmes can begin as early as 2 or 3 years old, when the brain is still developing, making therapies more effective than when they start later in life. Some children diagnosed with autism can also access state services, such as regional centres that offer continuous assistance for developmental disabilities. Potential therapies include speech therapy, physical therapy, hearing services, nutritional support, and family training. These interventions help children develop basic skills such as communication, social interaction, physical strength, thinking, and emotional management, reducing frustration and problematic behaviours related to communication difficulties. In Italy, the NIDA Network, funded by the Ministry of Health to monitor the neurodevelopment of the general population and high-risk individuals, has achieved national coverage. In 2023, it was further strengthened with the BABY@NET project, thanks to the National Recovery and Resilience Plan (PNRR) and the collaboration of the National Centre for Disease Prevention and Control (CCM) and the Cassa Depositi e Prestiti. These new initiatives will allow for the integration of clinical and experimental data with neurophysiological data and for the extension of teleassistance to all Neonatal Intensive Care Units (NICUs) and child neuropsychiatry services (UONPIA) within the National Health System. This will facilitate quicker and more widespread access to specialised assessments and targeted interventions.
T – TECHNOLOGICAL: this is arguably the area where the journey is still long. The studies presented in this analysis demonstrate the importance of leveraging the latest advancements in technologies such as Artificial Intelligence (AI) and Machine Learning (ML). These could be used to develop new tools or adapt existing ones from other fields, such as movement monitoring in Parkinson’s patients. Brain Computer Interfaces could also be adapted to assist autistic individuals with verbal communication difficulties. Additionally, new approaches could be considered, such as implementing mobile labs, which would increase the geographical reach and accessibility of studies. Allowing autistic participants to undertake research tasks closer to their homes, rather than in sterile lab environments, could make research more inclusive and more closely aligned with the real-life experiences of autistic people. Above all, if significant progress is to be made in autism research, the scope of what is considered “researchable” must be expanded. This would unlock the true potential of autism research and lead to practical solutions. A key future challenge will be integrating technologies into everyday tools for early diagnosis, enabling large-scale data collection that can feed back into continuous AI learning systems, improving accuracy over time. AI and ML, along with wearable devices, could be instrumental in monitoring ASD-related behaviours, allowing for early intervention and tailored treatments.
E – ECONOMIC: from a purely economic standpoint, there have been numerous interventions, such as the one presented five years ago by Davide Moscone, President of Spazio Asperger ONLUS and spokesperson for the Progetto Autismo FVG association, at the Italian Parliament. These interventions highlight that investing in autism research could have a significant long-term economic impact, as it would improve the effectiveness of services and reduce future healthcare costs. Currently, autism training and services are unevenly distributed across the country and often not optimised. With targeted investments in prevention, training, and technological innovation, it would be possible to reduce long-term healthcare expenditure by making services more accessible and efficient. Key areas for investment include: early interventions, which, although costly in the initial years, reduce long-term healthcare costs by minimising the need for more intensive treatments later on; training for parents and teachers, which can improve autism management at school and home, enhancing the child’s development and reducing the need for external resources; establishing specialists within schools, to ensure continuity in education and reduce the costs of training new professionals. Other areas of investment include optimising diagnostic processes through more open and affordable technologies, preventing environmental risks, and utilising teleassistance to make services more accessible, particularly in remote areas, while reducing costs. Furthermore, integrating autistic individuals into the workforce could stimulate economic autonomy, reducing the social and economic burden of disability. Investing in these areas would not only improve the quality of life for autistic people and their families but also generate significant economic savings in the long term, as early intervention and tailored support reduce the need for expensive care later in life.
P – POLITICAL: the political challenge in managing individuals with Autism Spectrum Disorders (ASD) lies in the broad application of research, therapies, and educational strategies that do not account for the diversity among autistic people. This “one-size-fits-all” approach often leads to therapeutic failures, negative consequences such as placement in low-expectation classes, and limited opportunities, ultimately resulting in poorer life outcomes. Despite growing awareness of this issue, the practices of educators, service providers, and medical professionals often remain rooted in outdated information, with harmful consequences for those with ASD. It is therefore crucial, not only important, to expand the representation of autistic profiles in research, reviewing methodologies and assessment tools to include individuals with limited motor skills, communication difficulties, or slower reaction times. The European Parliament’s Resolution of 4 October 2023 on harmonising the rights of autistic people highlights this problem: «… in Member States, it can take years to obtain an autism diagnosis for children and adults, and as a result, there is a shortage of high-quality and affordable person-centred intervention and support services tailored to individual needs and provided by trained professionals… Currently, there are no EU guidelines on evidence-based and rights-based autism interventions. Families across Europe still receive proposals for unproven and potentially harmful therapies and interventions, including clearly illegal procedures involving serious physical abuse of minors, such as chlorine dioxide enemas, which are still widespread and insufficiently regulated in most Member States and should be banned. Delayed and inadequate diagnoses can have serious consequences, ranging from denial of services to premature deaths…». Political action must focus on addressing these gaps by fostering policies that promote inclusive research, evidence-based interventions, and the elimination of harmful practices. Expanding autism rights across Europe and standardising early diagnosis procedures would reduce disparities in access to care, especially for individuals with more complex needs.
S – SUSTAINABILITY: the issue of the digital divide in accessing technologies for autism diagnosis and treatment is critical, especially in low- and middle-income countries (LMICs), which the World Health Organization (WHO) identifies as having the highest concentration of ASD cases. Barriers in these regions are numerous: from unreliable electricity and unstable internet connections to limited access to digital devices. This impedes the use of emerging technologies that could potentially improve early diagnosis and treatments for autism. While digital technologies have the power to transform early diagnosis and therapy, the lack of access to these resources in less developed areas exacerbates the disparities between wealthy and poorer nations. Without access to tools like teleconsultation and monitoring apps, many families are deprived of timely interventions that could significantly improve the quality of life for individuals with autism. Without addressing this issue, the digital divide will continue to grow, further hindering the social inclusion and participation of autistic individuals in underserved regions.
