Another breakthrough in AI! Identifying children's "invisible killer" before the age of 1, with an accuracy rate of over 80%

Autism is becoming the "invisible killer" of children.

Data from the World Health Organization (WHO) shows that 1 in 100 children is autistic, and about 50% of them have intellectual disabilities . Early intervention can significantly improve the social and cognitive functions of autistic patients. However, early identification is extremely difficult before a child is 2 years old.

Now, artificial intelligence (AI) is expected to help human doctors "early lock" autism in children before they are 12 months old .

A multimodal data analysis AI model developed by a research team at the Karolinska Institutet can not only detect early signs of autism in children around 12 months old , but also has an accuracy rate of 80.5% in identifying children under two years old. More importantly, the entire process only requires relatively limited information .

The related research paper, titled “Machine Learning Prediction of Autism Spectrum Disorder From a Minimal Set of Medical and Background Information”, was published in the scientific journal Jama Network Open.

The research team said that this research result is of great significance for early diagnosis and intervention strategies. The successful application of this method will help reduce the burden on families and society and improve the efficiency of identifying patients with autism.

With an accuracy rate of over 80%, AI can achieve early screening for autism

Autism, also known as autism spectrum disorder (ASD), is a neurodevelopmental disease that begins in early development and is mainly manifested by "deficits in social interaction and social communication abilities" and "restricted, repetitive behavior patterns, interests or activities."

Early identification of autism is crucial to improving the long-term prognosis of patients, especially during the critical period of brain development in children. Due to the subjectivity and cultural differences of traditional screening, diagnosis is easily delayed, resulting in the miss of the best intervention period.

To address these challenges, the study used the Simons Foundation Powering Autism Research for Knowledge (SPARK), the largest ASD research database currently available, to develop a machine learning model based on minimal background and medical information in order to achieve earlier ASD screening.

In terms of research subject selection, the research team used data from 30,660 participants in the SPARK (version 8) database, including 15,330 participants diagnosed with ASD and 15,330 participants without ASD. The experimental sample size is not only large, but also covers individuals of different ages, races, and genders, ensuring the wide applicability of the research.

Figure｜SPARK (8th version) research subject demographic details

To achieve early screening, the research team selected 28 characteristics that can be obtained through basic medical screening and background history of young children before they are 24 months old, covering 11 basic medical screening indicators and 17 background history data.

These features were selected based on availability and non-invasiveness, and relied primarily on information reported by parents. Compared to other studies that rely on complex medical tests or genetic data, this simplified feature selection approach makes the model more practical and generalizable, especially for early screening.

During the model development process, the researchers developed four different machine learning models: Logistic Regression, Decision Tree, Random Forest, and XGBoost training models, and used independent data sets for verification. The data sets were divided into 60% training set, 20% validation set, and 20% test set.

Figure ｜Performance of machine learning algorithms on a combined dataset of medical screening and background historical information

The results of the study showed that the XGBoost model (named AutMedAI) had an AUROC of 0.895 in the test set, showing high prediction accuracy with an accuracy rate of 80.5%, and the model performed strongly in detecting ASD.

Figure | Performance of machine learning algorithms that combine medical screening and background history measures

The researchers also used SHAP values ​​to evaluate the contribution of each feature to ASD classification to ensure the interpretability and practicality of the model. The results showed that developmental delay or abnormality, picky eating or abnormal eating behavior played an important role in ASD prediction.

In addition, the study found that the model performed relatively consistently across age groups (AUROC for 0-2 years old was 0.868, AUROC for 2-4 years old was 0.920, and AUROC for 4-10 years old was 0.906), further verifying its robustness and generalization ability.

A notable highlight of this study is the breadth and diversity of its data, as well as simplified feature selection. The study used a large-scale database and easily accessible features to achieve efficient early screening. This approach not only reduced screening costs, but also improved the practicality and scalability of the model.

However, there are some shortcomings in the research. First, the model varies greatly among typically developing children, such as the timing of learning to speak and toilet training. In addition, the research has not fully verified the applicability of the model to a wider population, as well as its actual effect in clinical applications.

AI provides new ideas for autism treatment

With the continuous development of AI technology, the application of AI in the field of autism is also gradually expanding. In addition to early screening, AI is also used to formulate personalized intervention plans in the treatment and rehabilitation process of autism.

For example, the Chengdu Frontier Brain-like Artificial Intelligence Innovation Center has developed a precise neural intervention system based on brain imaging analysis technology for the treatment and rehabilitation of autism. At the same time, some research institutions are also using AI technology to provide personalized learning programs for autistic children to help them better integrate into society. AI can not only assist doctors in developing more accurate treatment plans, but also improve treatment outcomes.

The robot Emo, created by a research team at Columbia University, can predict the upcoming facial expressions of humans 0.9 seconds before they smile and respond accordingly. This technology not only demonstrates the potential of AI in emotional interaction, but also marks an important progress in AI's emotional understanding and the establishment of human-machine trust.

Another study found that AI may have acquired capabilities similar to "Theory of Mind", which means that AI can understand the psychological state of humans in specific situations, such as "discovering wrong ideas", "understanding indirect speech", "identifying impoliteness", etc. GPT-4, GPT-3.5 and Llama 2 have performed close to or even better than humans in these aspects.

These findings not only demonstrate the powerful capabilities of AI in emotional understanding and psychological reasoning, but also provide new ideas for the further application of AI in autism treatment. In the future, AI is expected to play a greater role in emotional understanding, social interaction, and other aspects of autistic children, thereby comprehensively improving treatment outcomes.