Fine-tuned AI model flags early autism from children’s medical records in Chinese study

A large language model fine-tuned specifically for autism detection has shown promising accuracy in identifying autism spectrum disorder in young children from their electronic medical records, according to a multicenter study conducted in China and published in Child and Adolescent Psychiatry and Mental Health.

The motivation is a well-known clinical problem: early identification of ASD and its differentiation from other neurodevelopmental disorders are crucial for timely interventions, yet overlapping symptoms among neurodevelopmental disorders complicate early diagnosis.

The study combined a multicenter cross-sectional design with a single-center retrospective cohort study, both conducted between January 2019 and September 2024. The cross-sectional component included 3,940 children from Chongqing, used for model training with 10-fold validation, and 1,657 children from nine other Chinese cities for external validation. Participants ranged in age from 1.5 to 6 years and had been diagnosed with ASD, global developmental delay (GDD), or language delay (LD). The retrospective cohort comprised 584 children who were initially diagnosed with GDD or LD but were later confirmed to have ASD.

This produced two specialized systems, named GLM4-ASD and LLaMA3-ASD. The team then compared the performance of these fine-tuned models against a baseline model, commercial general-purpose models (GPT-4 and ERNIE-4.0), and a healthcare-specific model, Huatuo2-7B.

In internal validation, GLM4-ASD delivered the best performance of all models tested, achieving an area under the curve (AUC) of 0.839, an accuracy of 83.7 percent, a sensitivity of 82.6 percent, a specificity of 85.2 percent, and an F1-score of 84.8 percent. It was followed, in descending order of performance, by LLaMA3-ASD, GPT-4, ERNIE-4.0, and Huatuo2-7B.

The external validation offered an important test of whether the model had merely learned patterns specific to one hospital system. GLM4-ASD maintained robust performance on data from nine other cities, reaching an AUC of 0.873 and an accuracy of 87.9 percent. Notably, the model performed particularly well in children under 3 years of age, with an AUC of 0.882 and an accuracy of 90.5 percent in that subgroup, the age range where early identification matters most.

The retrospective longitudinal cohort provided a first glimpse of how the model might compare with clinical practice over time. Among the 584 children initially diagnosed with GDD or LD who were later confirmed to have ASD, GLM4-ASD identified ASD earlier than pediatricians in 44.2 percent of cases, at a mean age of 3.2 years. The model’s timing aligned with the pediatricians in 40.1 percent of cases, and it delayed or missed the diagnosis in 15.8 percent of cases, at a mean age of 4.7 years.

The authors are careful about what these numbers do and do not demonstrate. In their conclusions, they state that the fine-tuned GLM4-ASD model shows promise as a decision-support system for assisting pediatricians in early ASD identification from medical records, especially in children under 3 years. However, they emphasize that, as a retrospective study, the current findings primarily reflect classification performance rather than real-world diagnostic benefit, and they call for future prospective studies to further validate the model’s clinical utility.

If prospective validation confirms these results, a fine-tuned model of this kind could eventually help flag children who warrant specialist assessment earlier than they might otherwise be referred, while leaving diagnostic decisions in the hands of clinicians.

Subject of Research: Fine-tuned large language models for early identification of autism spectrum disorder from electronic medical records in Chinese children.

Article Title: Fine-tuned large language models for early identification of autism spectrum disorder: a multicenter study in China

Article References: Wang, K., Wei, Q., Wu, Y., Zhang, Q., Dai, Y., Wei, H., Liu, X., Wang, L., Zhang, L., Hao, Y., Yang, B., Jia, F., Yi, M., Li, F., Zhang, Y., Chen, L., Cheng, Q., Ding, Y., Yang, T., … Li, T. (2026). Fine-tuned large language models for early identification of autism spectrum disorder: a multicenter study in China. Child and Adolescent Psychiatry and Mental Health. https://doi.org/10.1186/s13034-026-01184-4

Image Credits: AI Generated

DOI: 10.1186/s13034-026-01184-4

Keywords: autism spectrum disorder, large language models, early diagnosis, electronic medical records, fine-tuning, low-rank adaptation, GLM4-ASD, pediatrics, neurodevelopmental disorders, decision support, machine learning, China