Autism spectrum disorder is a neurological disorder that results in developmental setbacks affecting the children’s social, communication, and behavioral activities. Children affected by ASD have difficulty interacting with parents, teachers, and friends. They show restricted interest in communicating with others. Such children suffer from maintaining eye contact with the people that they interact with. They have poor attention-holding ability, making it difficult to listen to others. According to statistics, 16 to 18 percent of children diagnosed with Down syndrome have autism [1]. Autism-affected children are oversensitive to noise and insensitive to pain. They seem lost in their thoughts, show difficulty recognizing their emotions, and sometimes have unusual memory. ASD starts in childhood but persists even when the person reaches adulthood. In 2021, the Center for Disease Control and Prediction, United States of America, reported that approximately 1 in 44 children is diagnosed with ASD. Children are usually diagnosed with ASD at the age of 3 years [2]. Autism can be detected in children under three years, and language delays can be seen in children as early as 18 months [3]. Children detected with ASD below five years, when trained with occupational and speech therapies, find remarkable improvement in their communication skills. Repetitive and stereotyped behavior, lack of social and language skills, poor eye contact, and delayed speech are warning alarms of autism to the parents. In very young children, a social skills assessment is challenging. Therefore, the limited eye contact with the parents, inability to bring an object, and inability to imitate the parent are the critical factors in identifying a child affected with ASD. In ASD, children aged 15 to 24 months encounter a degradation in language development. The child should maintain joint attention with the caregiver in sharing a social interaction between 8 to 16 months. ASD children generally lack this joint attention skill, and it is a critical feature in identifying ASD at an early age. All these signs and symptoms aid the parents or guardians in detecting ASD at an early stage [4]. Early detection of ASD helps in early intervention, improving the children’s learning ability, communication, and social skills.

The study collected ASD children data studying in special schools under the age of 15 and was named as Questionnaire based ASD (QBASD) dataset. The QBASD dataset has been carefully selected to include questions about the vital features to detect autism in children. The study identifies the most discriminating features from the dataset by feature selection methods of LASSO and the ICA. The selected feature set improves efficiency in classifying the test data by diagnosing the child as ASD or non-ASD. LR, Decision tree, KNN, and RF classifiers classify ASD from non-ASD subjects. The novelty of this research work lies in the design of CADx, and the questionnaire-based dataset collected from parents of ASD and normal children. The proposed methodology with the ICA algorithm for feature selection in the QBASD dataset and LR as a classifier is a novel methodology for diagnosing autism with machine learning. This CADx aims to detect autism early by training the model with data samples of children under five.

This section discusses the related study on early diagnosis of ASD in children. Researchers have identified several biomarkers to detect ASD at an early age. Any neuroimaging modality that discovers abnormal patterns in brain activity could detect ASD. Children with autism have typical morphological patterns in the Electroencephalogram (EEG) signals. One-dimensional local binary pattern extracts the features from EEG signals. The spectrogram images are generated from feature-extracted images using a short-time Fourier transform. The Relief algorithm carries out feature ranking and selection. With the SVM classifier, the model achieved an accuracy of 96.44% [5]. Magnetic Resonance Imaging (MRI) and resting state functional Magnetic Resonance Imaging (fMRI) images are studied to represent anatomical and functional connectivity abnormalities. The classification accuracy is 75% and 79% for fMRI and MRI data. The fusion of both datasets gives a higher accuracy of 81% [6]. The shortcoming of the above research work is that children must undergo neuroimaging tests, which is an unpleasant experience for them. Sixty acoustic features were extracted from the audio recording of 712 ASD children, and the feature selection method selected twenty-one deterministic features. Convolution Neural Network (CNN) showed improved results compared to Support Vector Machine (SVM) and Linear Regression in this study [7]. Facial expressions and vocal characteristics are biomarkers and detect ASD with 73% accuracy. The signs like reduced social smiling, higher voice frequency, and harmonic-to-noise ratio are significant biomarkers for ASD detection [8]. The eye movement data of ASD children is analyzed to distinguish between autism and non-autism subjects. The eye gaze feature used in the supervised machine learning model aids in ASD detection. The model achieved an accuracy of 86% and a sensitivity of 91% [9]. This study investigated the gaze behavior with classifiers such as SVM, Decision tree, Linear discriminant analysis, and RF. Classification accuracy for the visual fixation feature is 82.64% [10]. The features extracted from acoustic, video, and handwriting time series classified ASD children and children with neurotypical development. Eigenvalues of this correlation presented the coordination of speech with handwriting as a potential biomarker for classifying subjects [11]. The ASD subjects have unusual facial expressions and gaze patterns when they look at complex scenes. The authors have leveraged this fact to classify healthy subjects from ASD subjects. Classification accuracy is 85.8% for studying facial expressions from photographs [12]. The research studies discussed so far have the shortcoming of being expensive to develop, time-consuming, and exhibiting unsatisfactory accuracy rates. This research study aims to collect data from parents regarding their children’s behavior to identify the most deterministic biomarkers for ASD detection. Selecting a highly deterministic feature set can improve the classification model performance to a great extent.

The feature selection method, a preprocessing phase to classification, escalates the model’s performance by avoiding the overfitting of the data [32]. In this research, an experimental study analyzes the model performance with the proposed feature selection algorithm. The proposed algorithms are implemented in Python using the dataset QBASD. The classification accuracy of common classifiers on the QBASD without feature reduction are tabulated in Table 3.

The receiver operating characteristics (ROC) curve is the graph showing the performance of the classification model. The graph plots the false positive rate along the x-axis and the true positive rate along the y-axis. The area under the curve (AUC) measures the area under the ROC curve. Fig. 3 shows the ROC curve results for the QBASD dataset. Fig. 3a shows the decision tree classifier has the highest performance metric compared to other classifiers. The decision tree performs best as the features at each node classify the dataset into two sets. LR performance could be better than the performance of the decision tree, as it is a more efficient prediction model than the classification model. KNN is a victim of the curse of dimensionality; hence, the performance is inferior for a complete QBASD dataset. High k values result in underfitting; low k values will result in overfitting. In RF, the recall rate is high as the true positives are high and false negatives are negligible for ASD class detection. The precision of the ASD detection model is low as false positives are high.

Table 4 shows the classification metrics for the proposed ASD detection model with the ICA feature reduction algorithm. The LR and RF classifiers outperform the other classifiers. RF has the advantage of being robust to outliers; hence this algorithm performs well with QBASD data. The data may have outliers as the data is based on the child’s behavior and social interactions.

Fig. 4 shows the ROC curves after implementing the LASSO feature reduction algorithm and building classifiers in Python. The decision tree classifier works on a greedy approach; the decision at each level affects the next level, so the features affect the tree’s shape. The classifier performance could be better for small datasets. KNN is sensitive to outliers so a single outlier can change the classification boundary. It performs poorly as there may be outliers in the reduced dataset. LR performs exceptionally well for linear separable datasets. QBASD is a simple, linearly separable dataset; hence, LR performs exceptionally well. The Random Forest can handle outliers by binning the variables. It performs feature selection and builds its decision trees; therefore, the accuracy is high.

Table 5 shows the classification metrics for the proposed ASD detection model with the LASSO feature reduction algorithm. LR shows improved performance compared to other classifiers. LR has the risk of overfitting in high-dimensional datasets. LR model shows increased accuracy for the QBASD dataset reduced by the LASSO method as the few features selected are highly correlated with the target class.

Fig. 5 shows the ROC curves after implementing the ICA feature reduction algorithm and building classifiers in Python. The ROC curve for the decision tree shows a poor recall rate as the false negatives are high in the dataset. The precision rate is affected by the false positive rate. The LR shows high precision and recall rate as the false positives and negative counts are less for the classification model.

The RF classifier gives a poor precision rate as the false positives are high. A simple KNN classifier is robust with noisy data and performs well compared to sophisticated classifiers [33]. KNN gives average accuracy for the ICA-based ASD detection algorithm. The optimum value of k also has an impact on the accuracy of the model. This study chooses the k value as five based on the trial-and-error method. The precision value is lower for non-ASD class detection than for ASD. In the medical field, precision in diagnosis is a significant factor. The LR gives improved accuracy, F1 score, and precision compared to other classifiers. The feature selection by LASSO and ICA modifies the dataset as a crisp dataset with few independent, uncorrelated features. Hence LR algorithm shows high performance for data with the feature set having independent variables [28]. The reduced feature set includes features that make the dataset linearly separable, and LR gives improved results with QBASD [34].

The data visualization of the QBASD dataset reveals some interesting facts regarding the questions in the QBASD questionnaire and ASD detection. Table 6 tabulates the essential questions related to the significant ASD signs in children under five. The CADx proposed in this study can detect ASD in children under five years as the symptoms analyzed are specifically for the age group of 3 months to five years. Fig. 6 shows the data visualizations of the correlation of features of the QBASD dataset. The figures show how strongly and weakly the features are correlated with the response variables. All the significant features of the QBASD dataset are listed in Table 6.

Fig. 6a represents the visualization for questions Q4 and Q6 in the QBASD dataset. If the child shows restricted interest in playing or interacting with another child, irrespective of the smile factor, the child is classified as belonging to the ASD class. In contrast, if the child shows interest in playing with other children, the child belongs to the non-ASD class. Fig. 6b is the visualization of questions Q8 and Q9.

A child with ASD often has difficulty pointing to objects with the index finger. Fig. 6b shows a child that cannot point the finger at anything interesting irrespective of being able to maintain eye contact, is diagnosed with ASD. A child capable of pointing fingers to indicate interest in objects and maintaining eye contact for a second or two falls under non-ASD class. Fig. 6c is a visual analysis of Q13 and Q14. Children who make unusual noises and have stereotyped movements are classified as ASD. However, all ASD children do not need to make an unusual noise. Some children detected with ASD are free of stereotyped repetitive movements.

Fig. 6d is the visualization of Q15 and Q16. The graph shows that the child’s hyperactivity and self-injurious behavior classifies it as ASD. The child with no symptoms of hyperactivity or self-injurious behavior belongs to the non-ASD class. Children with hyperactivity could have attention deficit hyperactivity disorder (ADHD), not mandatorily ASD. In the visualization, non-ASD children are found to have symptoms of hyperactivity. Fig. 6e is the visualization of Q21 and Q22. Most children having inconsistent attention are classified as having ASD. The child’s delay in responding to the call is a essential biomarker for ASD detection. The visualization of questions Q25 and Q26 is shown in Fig. 6f.

The visualization’s conclusions are unclear on whether nutrient deficiency during pregnancy can cause ASD. There are very few samples where mothers are on medication; hence, it is difficult to conclude whether the mother’s medications cause ASD in the child. Fig. 6g shows the most significant symptom of responding to social cues as a biomarker for ASD detection. Classification of a child as ASD or non-ASD depends on the ability to respond to the social cue. Fig. 6h shows the strong correlation between the biomarker Q18 and ASD detection. The swarm plot shows that the children unable to follow the gaze are diagnosed with ASD.

Fig. 6i shows inconsistent attention is found more in ASD children than the non-ASD children. Fig. 6j is the swarm plot between Q23 to Q29. Children with ASD have unusual memory, but children without ASD have a good memory. Hence, it cannot be a robust independent biomarker for ASD detection. The common signs of autism include not responding to social cues, not following the parent’s gaze, not pointing to objects, not following simple instructions, repetitive movements, unusual memory, inconsistent attention, and not maintaining eye contact [35]. The visualization shows the strong correlation between the signs of autism specified by experts and those automatically detected by the proposed ASD detection system.

In this research, we have experimentally studied the performance of the proposed automated detection tool for ASD detection. The proposed CADx selects the best features from the QBASD dataset using the ICA feature selection algorithm. The performance of the model built with LASSO as the feature selection method and LR classifier gives 95% accuracy for ASD detection. The evaluation of models with standard metrics shows that the ICA-based ASD detection algorithm provides 100% accuracy with LR as the classifier. The proposed CADx can detect ASD in children under five years, as the features included in the dataset are signs of ASD for children under five. The logistic regression as a classifier gives high accuracy as it can handle outliers. LR is suitable for linearly separable datasets. The model shows improved accuracy compared to the state-of-the-art methodologies. The exploratory data analysis phase shows the relations between vital symptoms of ASD identified in the study and collected as a dataset. The visualization of the dataset reveals that the features selected by the ICA algorithm are significant features for ASD detection at an early age. This research is novel as the dataset is a self-collected dataset from special schools for autism. The future direction of research is to study the neuroimages to detect autism.