In today’s world, ASD has gained significant attention. Diagnostic methods typically focus on behavioral symptoms in social, sensory, and motor skills. Recent studies indicate that ASD affects approximately 1 in 36 children in the United States, with prevalence rates rising globally [1]. This increase highlights the urgent need for accessible, efficient, and scalable diagnostic methods. Recent advances in technology, machine learning, and data analysis are improving quantitative and ecological validation methods. However, clinical screening tests remain expensive and time-consuming [2].

Machine learning has shown remarkable success across fields, providing techniques for learning, detection, data analysis, and pre-processing. Rapid developments in computer science have enabled prediction models that integrate multiple disciplines [3,4]. This study aims to develop a machine learning-based estimator for ASD detection across different age groups using limited data.

Early diagnosis is crucial, yet obtaining well-structured datasets remains a challenge. A dataset covering four age groups—Toddler, Children, Adolescent, and Adult—was selected. Six clustering techniques (k-means, agglomerative, DBSCAN, mean shift, spectral, and birch) and ten machine learning classifiers (logistic regression, support vector machines, k-nearest neighbor, multi-layered perceptron, extra tree classifier, Gaussian process classifier, passive aggressive classifier, ridge, stochastic gradient descent, and linear support vector machines) were employed. The prediction performance of these models was evaluated through clustering on ASD datasets.

Despite numerous studies, no research aligns precisely with the objectives of this study. Existing approaches primarily focus on popular machine learning and deep learning models with an emphasis on feature extraction rather than processing. Proper feature normalization significantly impacts model performance. This research aims to bridge this gap by designing a clustering-assisted classification model, contributing to cost-effective and efficient ASD screening solutions. Considering these aspects, the study aimed to address the following questions:

How do alternative machine learning algorithms, apart from the commonly used ones, affect the model’s performance?

The designed model consists of two phases: data preprocessing and classification. Numerical transformations of the four ASD datasets selected in the data preprocessing phase were performed, and clustering algorithms were used. Machine learning algorithms selected with a clustered dataset were trained in the classification phase, and the prediction results were observed. In the next part of the study, information about the related work is given in Section 2; in Section 3, the materials used in the development of the study and the proposed model are explained; in Section 4, experimental results are given; in Section 5, the experimental results are discussed, and in Section 6, the results are concluded.

ASD detection with Machine Learning (ML) has just begun to attract attention; not many journal studies on this subject have been found in the literature. Therefore, conference publications have been added to the relevant studies’ titles.

Abdelwahab and others explored the use of ML to improve ASD diagnosis [5]. Using publicly available datasets from Kaggle and UCI ML, the researchers tested several ML algorithms. Data preprocessing involved feature selection, encoding, and normalization. Among the algorithms, Random Forest achieved the highest accuracy at 99.75%, while Logistic Regression also performed well at 96.69%. The findings highlight ML’s potential to complement traditional ASD diagnosis, enabling earlier intervention and reducing costs.

In 2024, researchers compared two AutoML tools—TPOT and KNIME—for ASD detection using data from rehabilitation centers in Pakistan [6]. Both tools automated feature selection and model tuning using the Q-CHAT-10 questionnaire. TPOT achieved 85.23% accuracy, while KNIME reached 83.89%, with the Q-CHAT-10 score identified as the most important predictor. The study highlights AutoML’s potential to streamline ASD diagnosis, making ML more accessible to healthcare professionals while improving early detection and treatment.

Xu et al. developed a method to detect ASDs in EEG (Electroencephalogram) datasets without using data augmentation methods [7]. They collected data from 97 ASD and 92 typically developing individuals from publicly available datasets. The data was collected during rest and while performing a task. They designed and implemented a combined network based on convolutional neural network (CNN) and long short-term memory (LSTM) for ASD detection. The developed network achieved classification accuracies of 81.08% and 74.55% for resting state and task state data, respectively.

Dia et al. proposed a supervised learning method to classify Autism Spectrum Disorder and to assess emotion levels among autistic children [8]. To evaluate the performance of the proposed approach, they used YouTube video frames of autistic children exhibiting typical autistic behaviors in unconstrained environments and conditions, as well as images of neurotypical people. They also proposed an extended version of a dataset containing additional influence labels corresponding to the influence levels of autistic children. Experiments were conducted using different models to determine the optimal performance of their architecture.

In 2024, researchers explored how the use of AI (Artificial Intelligence), particularly ML and deep learning (DL), can improve ASD detection [9]. They used natural language processing (NLP) to analyze Twitter posts, aiming to identify linguistic patterns associated with ASD. Various models, including decision trees, XGBOOST (eXtreme Gradient Boosting), k-nearest neighbors (KNN), Recurrent neural network (RNN), long short-term memory (LSTM), bidirectional long short-term memory (Bi-LSTM), and BERT (Bidirectional Encoder Representations from Transformers)-based models, were tested on a dataset of 404,627 tweets. BERTweet achieved the highest accuracy of 87.7%, demonstrating AI’s potential in ASD diagnosis.

Researchers reviewed AI-based methodologies for ASD detection through computer vision techniques in 2024 [10]. They studied ML models such as SVM, decision trees, and gradient boosting, alongside deep learning models like CNNs, RNNs, LSTMs, and Transformer-based approaches. They proposed a binary image classifier using the Xception CNN model trained on facial images of children aged 2 to 8 years. With a dataset of 23,000 images, the model achieved an accuracy of 88.87%, highlighting the effectiveness of facial analysis in ASD detection.

Loganathan et al. developed a hybrid ensemble model combining ResNet101 and BiGRU networks optimized with the CHGSO algorithm for ASD detection using EEG signals [11]. The hybrid ensemble model shows superior performance in ASD detection compared to existing methods such as DNN (Deep Neural Networks), SVM (Support Vector Machine), KNN, and MGOA-RF. Their Hybrid ensemble model reaches Sensitivity of 98%, 99% higher Specificity, 98% F1-Score, MCC of 99%, Accuracy of 98%, and Precision of 99%.

ML for ASD detection faces challenges such as limited datasets, symptom variability, and model interpretability. Small sample sizes can lead to overfitting, while diverse symptom presentations make pattern recognition difficult. Additionally, selecting relevant features and ensuring model reliability remain key hurdles. Addressing these issues requires robust preprocessing and validation techniques.

In this section, the datasets, clustering techniques and machine algorithms used in the study are mentioned, and then detailed information about the proposed model is given.

In this section, the proposed approach for the study was implemented, and all experiments were conducted in the environment specified under the title “Environment and Development”.

The results were evaluated separately for the clustered dataset. In the subsequent section, the outcomes obtained for each performance metric will be presented and thoroughly analyzed. Various clustering and ML approaches were considered in the study, but only the most prominent ones were included. Algorithms such as random forest and decision tree, which are commonly used in the literature, were excluded from the study, as there are already sufficient studies available about these algorithms. Additionally, algorithms with low accuracy were not included in the study to focus on the most effective ones.

As in all ML studies, the accuracy rate was first calculated in this study. The accuracy rate results are given in Table 3.

Table 3 displays the datasets in the leftmost column, followed by the clustering methods and their accuracy rates. Initially, each dataset was classified without clustering. Notably, some values are written in bold to emphasize the increase in accuracy rate, which will also be applied in other tables. A careful examination of Table 3 reveals that most of the clustering methods enhance the performance of algorithms. Particularly, almost all algorithms with Spectral demonstrated an increase in accuracy rate across all datasets. Furthermore, there is a noticeable difference in accuracy improvement between the Toddler and Adult datasets, as indicated in the “Datasets” title. In the Adult dataset, clustering improved accuracy rates for all algorithms, with Extra Trees Classifier (ETC) showing particularly notable success. This is because the number of non-ASD samples is higher in the Adult dataset. Continuing the analysis of Table 3, the most successful algorithms were found to be MLP, PAC, and ETC. These algorithms either maintained or increased accuracy rates across all clustering methods. The ROC curve is a very important performance measure for classification problems. ROC is a probability curve, and the area under it, AUC, represents the degree or measure of separability. For this reason, these values are very important in solving classification problems. Fig. 2 shows the performance of classification algorithms in detecting ASD with and without clustering.

For the model, the ROC curve was also drawn for each execution. However, they are not given here because they seem too complex and are too numerous. Table 4 shows the F1-score of classification algorithms in detecting ASD with and without clustering.

Previously given metrics cannot give a complete result for imbalanced datasets. Basically, it is the MCC criterion that evaluates by looking at the correlation (phi-coefficient) relationship between the actual data and the predicted data. Since the Toddler and Adult dataset has an uneven distribution, the clustering algorithms and classification methods vary. However, the result still does not change. In ASD detection, ETC, MLP, PAC, and RC algorithms, together with the Spectral and DBSCAN algorithms, show great performance and make accurate detection.

When the literature and existing papers are examined, it is seen that many researchers tend to solve the issue of ASD detection. The main goal of the study is to design a model that will increase the detection performance without interfering with the number of features in a small sample size. This study aims to design an ASD detection system for people of different age groups. Since diagnosis is very important for ASD, studies in this area are very important. Detection of ASD is very difficult, especially in age groups with small data and limited number of features.

The study examined the effect of six clustering methods on ASD datasets and the rate of improvement in classification. For this, the selected datasets were clustered and then evaluated with the specified performance metrics. The classification results were analyzed, the changes in detection results after clustering were observed, and a hybrid model was proposed. Fig. 3 shows total time of clustering and classification for all algorithms.

It is seen that almost every clustering method works successfully in certain algorithms with every dataset. However, the Spectral method stands out in this sense. The values obtained as a result of clustering the ASD datasets with Spectral increased the detection of ASD in all datasets compared to the unclustered state. Spectral Clustering and DBSCAN have shown superior performance over other clustering algorithms, particularly in small datasets, due to their ability to capture complex data structures. Spectral leverages graph-based techniques, transforming data into a similarity matrix before applying clustering.

The results of the study are summarized in the following section:

Using clustering before classification, the dataset for ASD detection generally improves performance. The quality and size of the dataset are crucial factors for building an effective prediction model. Clustering has played a significant role in improving dataset quality, ultimately leading to the creation of more successful prediction models, through the increased availability of larger datasets.

Overall, the study highlights the importance of clustering techniques in improving the detection of ASD and identifies specific algorithms that perform exceptionally well in different dataset scenarios.

Table 5 has the accuracy rate comparison with similar ASD studies found in the literature. The values for [28] and [29]are based on the reference [29], and the values for [30] represent the overall average accuracy achieved by their proposed models.

The study has several limitations. First, while four different ASD datasets were used, the results may not fully generalize to other datasets or populations. The sample size in certain age groups remained relatively small, which could limit the model’s performance in real-world, larger datasets. Additionally, the datasets were imbalanced, with more non-ASD samples than ASD samples, potentially influencing the model’s ability to accurately detect ASD. Although Spectral Clustering and DBSCAN showed strong performance, their effectiveness may vary with different datasets, which limits the broader applicability of the findings. The preprocessing steps played a significant role in the model’s success, but these methods may not be equally effective for datasets with different distributions. Finally, while the classification algorithms demonstrated improved performance with certain clustering methods, the results may not be consistent across all algorithms or datasets.

Detection of ASD is a critical area of research in psychology, especially with the rise of technology and artificial learning approaches. This study aimed to improve ASD detection in different age groups, particularly focusing on performance enhancement with small sample sizes. A hybrid model was proposed that integrates clustering and classification techniques, evaluating various clustering methods on six different ASD datasets. The results show that clustering significantly improved the performance of the 10 ML algorithms tested, with Spectral Clustering and the ETC, MLP, PAC, and RC algorithms yielding the most prominent improvements. Importantly, the study demonstrated that clustering on limited data could enhance estimation performance without reducing any features.

This work makes several key contributions: the development of a hybrid model for ASD detection, the application of clustering methods to small datasets, and the identification of algorithms that perform particularly well in these conditions. Moreover, the proposed model outperformed previous studies for three of the four age groups (Toddler, Children, and Adult), indicating its potential for improved detection in these groups.

However, the study does have limitations, such as the reliance on small datasets, which may have influenced the results, particularly for the Adolescent group. Future work will focus on integrating larger and more diverse datasets to validate the model’s effectiveness. Collaboration with clinical psychologists will also be crucial to evaluate the clinical applicability and robustness of the model. Additionally, exploring other feature selection methods and testing the model on new datasets can help refine the approach and further enhance ASD detection performance.