Multimodal Machine Learning for Autism Detection

In this paper we propose a machine learning multimodal system approach to the diagnosis of autism of 600 subjects' studies, using a static - "ready-to-research" - simulated data set. The model was a supervised machine learning, multimodal model, that fuses the demographic, autism screening, eye-tracking, speech-and-language, motor data, features extracted from EEG connectivity and quality of the data. We found that the multimodal, gradient-boosting algorithm achieved the best score of 96.83%, the precision of 96.62%, the recall of 95.42%, the specificity of 97.78%, the F1-score of 96.02% and the ROC-AUC of 0.984 scores for the matched confusion matrix. The confusion matrix consisted of 229 true positives, 352 true negatives, 8 false positives and 11 false negatives. All these metrics define strong discrimination, but do not trans mutate the model itself. The conclusion was that machine learning of integrated digital markers was superior to using just features. The features with most distinctive information were the instability of the motor and the visual configuration, but variability and connectivity contributed to better performance for the borderline class. It demonstrates multimodal screening could be used as screening tool to detect acute cases to refer to a clinician and save work hours necessary for manual screening and initial risk assessment. It has to be applied with caution. Medicine, medical and family history, interview and observation are essential for autism screening. The trajectory has to be from experimental research from simulated data sets to ethical real data sets using different ages, languages, ability levels and service providers. External validation is necessary. The calibration, fairness, missing-modality robustness (in case some of the features are missing) explainable feature importance and temporal tracking of outcomes should also be strengthened. The new systems will also do modular inference and missing-modality estimates given missing data in the screening situation. The current system should therefore be considered the reasoner of modest multimodal autism screening system for testing hypothesis. The methodological system presented in this paper is only the first step towards systematic replication in the clinic with rigorous sampling, prospective data collection (recruitment) and review of all multimodal reasoning, multimodal fusion and multimodal classification - before claiming that this is "real screening" with real data.

[1]     S. Bae, J. Hong, S. Ha, et al., “Multimodal AI for risk stratification in autism spectrum disorder: Integrating voice and screening tools,” npj Digital Medicine, vol. 8, Art. no. 538, 2025, doi: 10.1038/s41746-025-01914-6. [2]     D. Y. Kim, R. Do, Y. Shin, et al., “Automated AI based identification of autism spectrum disorder from home videos,” npj Digital Medicine, vol. 8, Art. no. 607, 2025, doi: 10.1038/s41746-025-01993-5. [3]     H. Wang, H. Jing, J. Yang, et al., “Identifying autism spectrum disorder from multi-modal data with privacy-preserving,” npj Mental Health Research, vol. 3, Art. no. 15, 2024, doi: 10.1038/s44184-023-00050-x. [4]     S. Perochon, J. M. Di Martino, K. L. H. Carpenter, et al., “Early detection of autism using digital behavioral phenotyping,” Nature Medicine, vol. 29, pp. 2489–2497, 2023, doi: 10.1038/s41591-023-02574-3. [5]     J. T. Megerian, S. Dey, R. D. Melmed, et al., “Evaluation of an artificial intelligence-based medical device for diagnosis of autism spectrum disorder,” npj Digital Medicine, vol. 5, Art. no. 57, 2022, doi: 10.1038/s41746-022-00598-6. [6]     R. Assaf, Z. Shehabeddine, and V. Ramesh, “Screening autism spectrum disorder in children using machine learning on speech transcripts,” Scientific Reports, vol. 15, Art. no. 34134, 2025, doi: 10.1038/s41598-025-01500-6. [7]     U. J. Ganai, A. Ratne, B. Bhushan, et al., “Early detection of autism spectrum disorder: Gait deviations and machine learning,” Scientific Reports, vol. 15, Art. no. 873, 2025, doi: 10.1038/s41598-025-85348-w. [8]     M. S. Farooq, R. Tehseen, M. Sabir, et al., “Detection of autism spectrum disorder in children and adults using machine learning,” Scientific Reports, vol. 13, Art. no. 9605, 2023, doi: 10.1038/s41598-023-35910-1. [9]     M. T. Ali, A. Gebreil, Y. ElNakieb, et al., “A personalized classification of behavioral severity of autism spectrum disorder using a comprehensive machine learning framework,” Scientific Reports, vol. 13, Art. no. 17048, 2023, doi: 10.1038/s41598-023-43478-z. [10]   F.-L. Zhu, S.-H. Wang, W.-B. Liu, H.-L. Zhu, M. Li, and X.-B. Zou, “A multimodal machine learning system in early screening for toddlers with autism spectrum disorders based on the response to name,” Frontiers in Psychiatry, vol. 14, Art. no. 1039293, 2023, doi: 10.3389/fpsyt.2023.1039293.