AI Analysis Challenges Traditional Autism Diagnostic Criteria, Emphasizing Repetitive Behaviors Over Social Skills

Repetitive behaviors and special interests are more indicative of an autism diagnosis than a lack of social skills

People with autism are typically diagnosed by clinical observation and assessment. To deconstruct the clinical decision process, which is often subjective and difficult to describe, researchers used a large language model (LLM) to synthesize the behaviors and observations that are most indicative of an autism diagnosis. Their results, publishing in the Cell Press journal Cell, show that repetitive behaviors, special interests, and perception-based behaviors are most associated with an autism diagnosis. These findings have potential to improve diagnostic guidelines for autism by decreasing the focus on social factors -- which the established guidelines in the DSM-5 focus on but the model did not classify among the most relevant in diagnosing autism. "Our goal was not to suggest that we could replace clinicians with AI tools for diagnosis," says senior author Danilo Bzdok of the Mila Québec Artificial Intelligence Institute and McGill University in Montreal. "Rather, we sought to quantitatively define exactly what aspects of observed behavior or patient history a clinician uses to reach a final diagnostic determination. In doing so, we hope to empower clinicians to work with diagnostic instruments that are more in line with their empirical realities." The scientists leveraged a transformer language model, which was pre-trained on about 489 million unique sentences. They then fine-tuned the LLM to predict the diagnostic outcome from a collection of more than 4,000 reports written by clinicians working with patients considered for autism diagnosis. The reports, which were often used by multiple clinicians, included accounts of observed behavior and relevant patient history but did not include a suggested diagnostic outcome. The team developed a bespoke LLM module that pinpointed specific sentences in the reports that were most relevant to a correct diagnosis prediction. They then extracted the numerical representation of these highly autism-relevant sentences and compared them directly with the established diagnostic criteria enumerated in the DSM-5. "Modern LLMs, with their advanced natural language processing capabilities, are natively suited to this textual analysis," Bzdok says. "The key challenge we faced was in designing sentence-level interpretability tools to pinpoint the exact sentences, expressed by the healthcare professional themselves, that were most essential to a correct diagnosis prediction by the LLM." The researchers were surprised by how clearly the LLM was able to distinguish between the most diagnostically relevant criteria. For example, their framework flagged that repetitive behaviors, special interests, and perception-based behavior were the criteria most relevant to autism. While these criteria are used in clinical settings, current criteria focus more on deficits in social interplay and lack of communication skills. The authors note that there are limitations to this study, including a lack of geographical diversity. Additionally, the researchers did not analyze their results based on demographic variables, with the goal of making the conclusions more broadly applicable. The team expects their framework will be helpful to researchers and medical professionals working with a range of psychiatric, mental health, and neurodevelopmental disorders in which clinical judgement forms the bulk of the diagnostic decision-making process. "We expect this paper to be highly relevant to the broader autism community," Bzdok says. "We hope that our paper motivates conversations about grounding diagnostic standards in more empirically derived criteria. We also hope it will establish common threads that link seemingly diverse clinical presentations of autism together."

Rethinking autism diagnosis with AI and clinical expertise

University of MontrealMar 26 2025 In diagnosing autism - the developmental variant that affects around 80 million people worldwide - medical practitioners today put too much emphasis on a child's lack of sociability and not enough on their interests and how they naturally behave spontaneously with objects. And so, to be more accurate in their assessments, health authorities should start tapping the vast analytic powers of artificial intelligence, combined with the experience of clinicians, and come up with better diagnostic criteria. That's what Canadian neuroscientists argue in a new study, published today in the journal Cell. A data-driven revision of autism criteria of the kind we're proposing, grounded on clnical certainty, would complement what has historically been done by expert panels and the judgment of humans, who can be wrong." Laurent Mottron, co-senior author, clinician-researcher in psychiatry at Université de Montréal Added co-first author Emmet Rabot, an UdeM clinical associate professor of psychiatry: "This project marks the successful outcome of a fruitful partnership between McGill University and UdeM. We hope our results will make a meaningful contribution to advancing diagnosis and support for the autistic community." The study involved Danilo Bzdok, Jack Stanley Siva Reddy and Eugene Belilovsky, all scientists at Mila - Quebec Artificial Intelligence Institute, which is affiliated with UdeM and McGill Stanley and Bzdok are also associated with The Neuro - Montreal Neurological Institute-Hospital, which is affiliated with McGill. DSM-5, the gold standard With no specific markers of autism yet available in a person's genes, blood or brain, diagnosis today still largely depends on clinical assessment by physicians and their assessment teams. The standard way of doing this is by observing how a child fits the criteria for autism listed in gold-standard manuals such as the American Psychiatric Association's Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), as well as in standardized instruments that are mapped on DSM. These criteria are divided into two categories: one for a child's differences in social communication and interaction and another for their restricted or repetitive behaviours, actions or activities. In the end, however, it is the clinicians, relying on years of experience, who decide whether a child gets diagnosed - and how much they fit the DSM-5 criteria can vary a lot. To empirically test which criteria clinicians most often observed in people diagnosed with autism, the McGill and UdeM scientists ran more than 4,200 observational clinical reports from a French-speaking cohort of children with suspected autism in Montreal through an AI program for analysis. They tailored and carried out large language modelling (LLM) approaches to predict the diagnosis decision in each case, based solely on these reports. In particular, the investigators came up with a way to identify key sentences in the reports that were most relevant in a positive diagnosis. That then allowed them to make a direct comparison with the U.S. diagnostic criteria also accepted worldwide - with surprising results. They found that criteria related to socialization - emotional reciprocity, nonverbal communication and developing relationships - were not highly specific to an autism diagnosis. In other words, they were not found much more in children diagnosed with autism than in those for whom a diagnosis was ruled out. Criteria related to repetitive behaviors, highly specific interests and perception-based behaviors, however, were strongly linked to an autism diagnosis. Reconsider and review the criteria These findings led the scientists to argue that the medical community may want to reconsider and review the established criteria used to diagnose autism - as the current criteria seem both inadequate and responsible for the over-diagnosis of autism that has been widely documented around the world. They should put much less weight on a child's lack of social skills, a weighting that's been emphasized for decades now, the authors argue. Challenges in socializing are common in autistic children but other atypical signs that are much easier to identify also characterize these children, they say. Increased focus should be put on children's repetitive and perception-based behaviours and special interests, they add, as those might be more specific to autism that previously thought. Receiving an autism diagnosis can take years, delaying interventions that improve outcomes and quality of life. Conversely, an unjustified diagnosis can lead to a whole host of bad decisions, the scientists say. Improving the assessment process, therefore, would provide vast benefits to autistic people and the public healthcare system. "In the future, large language model technologies may prove instrumental in reconsidering what we call autism today," said Bzdok, the study's other senior author. University of Montreal Journal reference: Stanley, J., et al. (2025). Large language models deconstruct the clinical intuition behind diagnosing autism. Cell. doi.org/10.1016/j.cell.2025.02.025.

AI analysis of health care records reveals key factors in autism diagnosis

Without clear and effective biological tests for autism based on genes, brain or blood measurements, diagnosis today still largely depends on clinical assessment. The standard way of doing this is by observing how the individual meets the criteria for autism listed in gold standard manuals like the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5). These criteria are divided into two categories: one for restricted or repetitive behaviors, actions, or activities, and another for differences in social communication and interaction. In the end, however, it is the clinician, relying on years of experience, who decides whether the individual is given an autism diagnosis. The degree to which an individual diagnosed with autism fits the DSM-5 criteria can vary considerably. To empirically test which criteria clinicians most often observed in people diagnosed with autism, scientists performed artificial intelligence (AI) analysis on more than 4,200 observational clinical reports from a French-speaking child cohort from Montreal, in Québec, Canada. They tailored and carried out large language modeling approaches to predict the diagnosis decision based solely on these reports. Their findings were published in the journal Cell. In particular, the investigators came up with a way to identify key sentences in the reports that were most relevant to a positive diagnosis, allowing direct comparison with the diagnostic criteria. The analysis found that criteria related to socialization, such as emotional reciprocity, nonverbal communication, and developing relationships, were not highly specific to an autism diagnosis, meaning they were not found much more in individuals diagnosed with autism than those in which a diagnosis was ruled out. Criteria related to repetitive movements, highly fixated interests, and perception-based behaviors, however, were strongly linked to an autism diagnosis. Their findings lead the scientists to argue that the medical community may want to reconsider and review the established criteria used to diagnose autism. Specifically, the heavy weighting of socialization -- for several decades now -- in assessing autism, which may also contribute to the increase in autism diagnosis in developed countries, may have to be reduced, with increased focus on certain repetitive behaviors and special interests. This would make diagnosis more effective and efficient, as social factors are relatively time-consuming, labor-intensive, and imprecise to assess compared to more obvious behavioral traits. Receiving an autism diagnosis can take years, delaying interventions that improve outcomes. Making the assessment process more focused and streamlined may provide vast benefits to autistic people and the health care system. "In the future, large language model technologies may prove instrumental in reconsidering what we call autism today," says Danilo Bzdok, a neuroscientist at The Neuro and Mila (the Quebec Artificial Intelligence Institute), and the study's co-senior author. "Such a data-driven revision of autism criteria is a complement to what has historically been done by expert panels and human judgment alone," says Laurent Mottron, a clinician-researcher at the Department of Psychiatry of Université de Montréal, and co-senior author of the study.

Using LLMs to understand how autism gets diagnosed | Newswise

In diagnosing autism - the developmental variant that affects around 80 million people worldwide - medical practitioners today put too much emphasis on a child's lack of sociability and not enough on their interests and how they naturally behave spontaneously with objects. And so, to be more accurate in their assessments, health authorities should start tapping the vast analytic powers of artificial intelligence, combined with the experience of clinicians, and come up with better diagnostic criteria. That's what Canadian neuroscientists argue in a new study, published today in the journal Cell. "A data-driven revision of autism criteria of the kind we're proposing, grounded on clnical certainty, would complement what has historically been done by expert panels and the judgment of humans, who can be wrong," said co-senior author Laurent Mottron, a clinician-researcher in psychiatry at Université de Montréal. Added co-first author Emmet Rabot, an UdeM clinical associate professor of psychiatry: "This project marks the successful outcome of a fruitful partnership between McGill University and UdeM. We hope our results will make a meaningful contribution to advancing diagnosis and support for the autistic community." The study involved Danilo Bzdok, Jack Stanley Siva Reddy and Eugene Belilovsky, all scientists at Mila - Quebec Artificial Intelligence Institute, which is affiliated with UdeM and McGill. Stanley and Bzdok are also associated with The Neuro - Montreal Neurological Institute-Hospital, which is affiliated with McGill. With no specific markers of autism yet available in a person's genes, blood or brain, diagnosis today still largely depends on clinical assessment by physicians and their assessment teams. The standard way of doing this is by observing how a child fits the criteria for autism listed in gold-standard manuals such as the American Psychiatric Association's Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), as well as in standardized instruments that are mapped on DSM. These criteria are divided into two categories: one for a child's differences in social communication and interaction and another for their restricted or repetitive behaviours, actions or activities. In the end, however, it is the clinicians, relying on years of experience, who decide whether a child gets diagnosed - and how much they fit the DSM-5 criteria can vary a lot. To empirically test which criteria clinicians most often observed in people diagnosed with autism, the McGill and UdeM scientists ran more than 4,200 observational clinical reports from a French-speaking cohort of children with suspected autism in Montreal through an AI program for analysis. They tailored and carried out large language modelling (LLM) approaches to predict the diagnosis decision in each case, based solely on these reports. In particular, the investigators came up with a way to identify key sentences in the reports that were most relevant in a positive diagnosis. That then allowed them to make a direct comparison with the U.S. diagnostic criteria also accepted worldwide - with surprising results. They found that criteria related to socialization - emotional reciprocity, nonverbal communication and developing relationships - were not highly specific to an autism diagnosis. In other words, they were not found much more in children diagnosed with autism than in those for whom a diagnosis was ruled out. Criteria related to repetitive behaviors, highly specific interests and perception-based behaviors, however, were strongly linked to an autism diagnosis. These findings led the scientists to argue that the medical community may want to reconsider and review the established criteria used to diagnose autism - as the current criteria seem both inadequate and responsible for the over-diagnosis of autism that has been widely documented around the world. They should put much less weight on a child's lack of social skills, a weighting that's been emphasized for decades now, the authors argue. Challenges in socializing are common in autistic children but other atypical signs that are much easier to identify also characterize these children, they say. Increased focus should be put on children's repetitive and perception-based behaviours and special interests, they add, as those might be more specific to autism that previously thought. Receiving an autism diagnosis can take years, delaying interventions that improve outcomes and quality of life. Conversely, an unjustified diagnosis can lead to a whole host of bad decisions, the scientists say. Improving the assessment process, therefore, would provide vast benefits to autistic people and the public healthcare system. "In the future, large language model technologies may prove instrumental in reconsidering what we call autism today," said Bzdok, the study's other senior author.