Early diagnosis of autism using indian autism grading tool

1 Introduction

Autism Spectrum Disorder (ASD) is a neuro-development disorder considered by repetitive patterns of behaviors, complicatedness in social communication [1] and impaired social interaction [2, 3]. It can be identified by symptoms such as different social behaviors, expression of stereotyped [4] or repetitive activities and uncontrolled body movements, fascinated by moving, spinning objects and other sensory abnormalities. The most famous neurological debates for ASD states that 1 in 59 below 10 years is affected by ASD; the rate of ASD appears to be increasing year by year.

A child’s early signs of ASD can be identified by its parents. Nevertheless, it would be difficult for many parents to recognize the symptoms of autism because the child would have normal physical growth as of any typical child of that age, but would be abnormal in its activities. Though the symptoms of ASD different some symptoms appear to be similar to that of the other disorders. This makes it difficult for the autism to be identified at an early stage. Hence the parents have to often keenly observe their child for sometimes and consult their friends, relatives and doctor regarding their child’s abnormal behaviors impaired activities, developmental delays and other doubts, which lead to delay in the diagnosis of ASD [5] Medical practitioners tend to follow different schedules and various autism screening and diagnostic tools [6] that lead to a lack of unified diagnosis. There are some parental interview-based ASD diagnoses available; Autism Diagnostic Interview-Revised (ADI-R) [7], Diagnostic Interview for Social and Communication Disorders (DISCO), Developmental, Diagnostic and Dimensional interview (3Di), Childhood Autism Rating Scale (CARS/CARS-2) [8] and Indian Scale for Assessment of Autism (ISAA).

Experts and therapists suggest that ASD can be reduced by early intervention and early screening. However, the early identification becomes very complex and gets delayed [9, 10] due to the misunderstanding of disabilities, misguidance by elders, ignorance due to social status, feeling of inferiority and so on. The early autism identification by parents plays a vital role in the child’s early start of the therapy and disability constraints of ASD. However, in lower middle-income countries like India, the ASD screening and assessment process is more costly and a time-consuming one due to the dependence on patented western tools like ADI-R, DISCO and CARS, which may not be accessible to the people of remote villages.

A lot of studies are made in machine learning to improve ASD diagnosis. Most recently, for autism screening and diagnosis, a varsiety of methodologies are followed: Eye Tracking System, Personal Characteristic Data (PCD), Structural Magnetic Resonance Imaging (MRI) Image, Autism Brain Imaging Data Exchange (ABIDE) and diagnosing autism using home videos, a multi disease prediction using deep learning methodology. The prediction methodology created by the following steps like Data Acquisition, Optimal Feature selection, Statistical feature Extraction, and prediction, this system is verified by using different deep learning algorithms. [11].

This study makes possible to the development of machine learning model towards the automated diagnosis of ASD. From the above methods of ASD screening and diagnosis is impractical for the people who are living in India. The parental interview system is more suitable for the Indian people because the parent is the only person to know about the children who are all in abnormal activities and struggled in their daily regular activities.

In the present research, a novel ML powered mobile application called Indian Autism Grading Tool (IAGT) is developed for parents, caregivers and special school teachers for the initial screening and diagnosis of ASD. The parental questionnaire is designed in an easily understandable format in both Tamil and English language with the aim to provide accessibility to underprivileged people. This can help people to screen and identify ASD on their mobile devices. It also acts as the platform for giving first opinion to assessors. In addition, the IAGT is an intelligent tool-based screener which would give immediate results with a user-friendly environment. This ASD diagnostic tool simulates the expertise of the clinician with the intensive development of Artificial Intelligence (AI) based mobile application. Machine Learning (ML) is a part of an artificial intelligent technique to provide support for decision making in medical and healthcare sectors [12].

ML has a variety of classification algorithms to solve the supervised and unsupervised learning [13, 14]. Decision Tree Classifier, Gaussian Naive Bayes approach, K-Nearest Neighbors (KNN), Multinomial Logistic Regression (MLR) [14] and Support Vector Machine (SVM) are supervised machine learning techniques that can be implemented as automatic ASD screening and the results can be validated against the real-world autism data.

Due to the practical difficulties in data collection, an algorithm which performs well for the least number of data has to be identified. So, the MLR and DT models outperform in the training and testing compared to the other classifiers. At the same time the validation process for DT has less prediction accuracy from unseen dataset [15]. In these classifiers, MLR is known to be more robust and it supports least number of data and it can be used for the implementation of an ML powered mobile application.

The IAGT provides a great support to initial autism screening for the parents, teachers, social workers and experts who are working with autistic children [16]. It clears the parents’ doubts regarding their child’s developmental delays. Mobile based computing tool is compact and used as an automated evaluation with optimum data management system of storage and retrieval. The most important aspect of a mobile device is portability which helps to use the autism grading application at home, at school and at public places. Since it is proven to be consistent and has interactive dynamic applications, mobile technology reduces the time and cost of the screening process. The IAGT provides parents with an ability to assess and recognize their child’s behaviors, and identify whether the child has autistic symptoms or not. This application will further guide the parents if the child needs to be referred to any clinician for detailed diagnosis. The key contributions of the present work are as follows,

The IAGT application is capable of predicting ASD symptoms based on the medical records of a child with ASD.

The rest of this work is organized in the following sections. Section-2 briefly discussed about the related works, Section-3 includes the materials and method for developing IAGT, Section-4 comprise with results and discussion and finally section 5 encloses with conclusion.

2 Related works

There are various autism screening mobile applications available for early diagnosis. Most of the existing applications employ score-based evaluation [17, 18]. A few applications have implemented Artificial Intelligence (AI) techniques that are suitable for clinicians or experts in autism assessment, but are not useful to parents for the initial assessment of their child. The following are the list of applications used for autism assessment: Indian Scale for Assessment of Autism (ISAA), ASD Tests app, Eye-Tracking App, Autism & Beyond, Mobile Video Rater platform for autism Risk Classification, Naturalistic Observation Diagnostic Assessment (NODA) and Autism AI App [19, 20].

System Architecture and Design: The present research is carried out in four phases, namely Parental Interview Questionnaire Development, Data Collection, Development of ML models and implementation of the best performing ML model for Development of IAGT Mobile Application. The architecture and design of ASD screening is shown in Fig. 1.

OPEN IN VIEWER

Fig. 1

Architecture and design of ASD screening.

3.5 Mobile application tool for autism identification

An outperforming machine leaning approach is further combined to a mobile application which assists as a tool for the autism parental interview and screening based on IAPQ questionnaire.

This questionnaire is answered by the parents, caregivers and teachers. IAGT parent ASD screening tool is shown in Fig. 2(a), (b), (c) and (d). The IAGT mobile application encourages home based autism screening in the smartphones and tablets and in all the other mobile devices. It is a robust and upgradable application. Parents can install the application in their mobile phones and can perform autism screening easily. They can choose either Tamil or English in which they wish to respond. Then the mobile tool would display the questions in the language chosen. Initially, the personal and medical information about the child like the child’s behavior, birth history, problems and the present treatment and child’s personal information are to be entered. Then the user has to fill answers to 37 questions with 5 prescribed options which reflect the child’s behavior. Figure 3 illustrates the steps involved in the use of ML powered IAGT mobile tool.

OPEN IN VIEWER

Fig. 2

Autism screening mobile application; (a) home screen; (b) parent questionnaire in tamil; (c) parent questionnaire in english; (d) ML predicted result.

OPEN IN VIEWER

Fig. 3

Flow of IAGT.

Finally, the answers are fed into the machine learning model to predict the autism grade automatically for the child. Then the autism grade given in Fig. 2(d) is displayed. This outcome is obtained from a highly trained ML model rather than a score-based prediction.

4 Experiment and results

The machine learning models are trained and tested using 37 features. Hyper parameter tuning is an essential aspect of machine learning process. A hyperparameter is an input into a model that is determined before the learning process starts. The model is validated by partitioning the collected samples into training set and test set to evaluate the model performance. Increasing the complexity or learning rate of a model demonstrates the importance of hyperparameters. The most optimal combination of hyperparameters can be analyzed as a search problem for models with many parameters. There are also some hyperparameters for each classifier, for example SVM has some parameters like C and gamma and by tuning these parameters SVM yields higher weighted average precision of 0.98. DT has max_depth, max_features and max_leaf_nodes parameters, by tuning these parameters DT achieves the weighted precision of 0.98, KNN yields the weighted precision of 0.96 by tuning n_neighbors and leaf_size parameters. The regression parameters (β coefficients) is the hyperparameters of MLR that can be tuned to yield the weighted precision of 1. Although, GNB does not have hyperparameters and it obtains the weighted precision of 0.97 as depicted in Table 4. The parameters that are learned in GNB include the prior probability of various classes, as well as the probability of various attributes for each class. Due to the assumption of feature independence, the GNB may produce erroneous class probabilities.

Furthermore, the performances of the models are analyzed by two different ways. In the first case, the test result of the individual class’s prediction accuracy, sensitivity, specificity and Kappa of all five models is obtained as shown in Table 7. Then the machine learning models are verified by previously trained data and also unseen data. The Validation results of DT, GNB, KNN, MLR and SVM for both Module 1 and Module 2 are shown in Fig. 4.

OPEN IN VIEWER

Fig. 4

Validation results for DT, GNB, KNN, MLR and SVM for both module 1 and module.

Table 5 shows the estimation of AUC and ROC values of the model. The macro one-vs-one ROC-AUC value is 0.967 and 0.979 is weighted prevalence. Similarly, the one-vs-rest ROC-AUC yields the same value for macro and weighted prevalence.

Figure 5 shows the ROC for autism and Fig. 6 shows the multiclass classification of autism based on the TPR and FPR. Figure 7, show that for each of the ML classifiers was tested, normal (class 2) has the high classification value, with only one misclassification. For the MLR classifier, the ASD prognosis was effectively classified 100% of the time successfully classifying 75% of the testing data.

OPEN IN VIEWER

Fig. 5

ROC curve of ASD.

OPEN IN VIEWER

Fig. 6

ROC curve for multiclass of ASD.

OPEN IN VIEWER

Fig. 7

Confusion matrices of (a) MLR (b) SVM (c) DT (d) KNN and (e) GNB classifiers.

The ML powered autism diagnostic mobile application allows people to screen and detect autism spectrum disorder in both offline and online modes. This tool has the capability of changing languages according to users native. The robustness of the proposed ML techniques was evaluated using performance metrics such as accuracy, precision, recall and f1 score. The model achieves better robustness from the experimental results illustrated in Table 6. The scalability of this model is depicted in Fig. 8. Similarly Table 6 reflects accuracy, precision, recall, f1-score and support values for the machine learning models. In this, DT and MLR classifiers’ accuracy, precision, recall and f1-score are high with 100% for all four classes. The accuracy of mild, no autism and severe has 100% and then moderate class accuracy is 97%. The GNB and SVM classifiers’ accuracy of mild, No Autism and severe is 100%, and the accuracy of moderate class is 97%. In precision accuracy in Moderate, no autism and severe class have 100%, and then the accuracy of mild cases is 96%. In the recall the accuracy of mild, no autism and severe is 100%, and it is 97% in moderate cases. In f1 score the accuracy of both mild and moderate class has 98%, and then no autism and severe accuracy have 100%. The KNN classifier is accuracy, precision, recall, f1-score accuracy for the no autism and severe class have 100%, mild class has 97% and moderate class accuracy has 96%. Finally, support counts have 26 records in mild, 29 records in moderate, 31 records in normal and 23 records in severe autism.

OPEN IN VIEWER

Fig. 8

Scalability of the ML classifiers.

In the validation process of modules 1 and 2 in Fig. 4, KNN has the smallest difference from trained data that accuracy has 98.63% and new data accuracy has 93.33%. GNB, MLR and SVM have 100% and 86.67%. DT has 100% and 76.67% in shown in Table 7.

The predictive model is cross validated by 10-fold cross validation technique by partitioning the collected samples into training set and test set to evaluate the model performance. In the 10-fold cross validation, the dataset randomly partitioned into 10 equal size groups and validate the predictive model performance. In this validation accuracy of the DT has 95.53, GNB has 96.34, KNN has 97.02, MLR has 97.85 and SVM has 97.14 is shown in the Fig. 6 and the training and testing loss function curve is depicted in Fig. 9.

OPEN IN VIEWER

Fig. 9

Fold cross validation result for autism predictive model.

From Fig. 9 MLR achieves the overall accuracy of 97.85%, which comparatively 0.72%, 2.37%, 0.84% and 1.54% better than SVM,DT,KNN and GNB respectively. Figure 10 depicts the number of iterations and loss value, which shows that as the number of iterations increases, the model loss decreases. So, the model predicted result highly reliable.

OPEN IN VIEWER

Fig. 10

Loss function curve for (a) MLR, (b) SVM, (c) DT, (d) KNN and (e) GNB classifiers.

Mobile applications are being increasingly used in the medical healthcare sector. The majority of healthcare applications can only be handled by clinicians. Automated patient assessment applications handled by people are very few; parents need easy to use applications to motivate them for the first screening. The proposed mobile healthcare IAGT application allows people to screen autism at anytime and anywhere, and thus it provides a valuable first opinion at an affordable cost. Despite the rapid growth of ASD, diagnostic centers are busy and doctors spend less time on screening, which leads to misdiagnosis. Moreover, the initial autism screening assessment takes a longer time to complete the parental interview and autism administration process. Therefore, screening and diagnosis should be administered at an earlier stage and procedures should be understood by everyone. The present study proposes a multiple-choice IAGT tool consisting of 37 questions with a five-point Likert scale. The five machine learning models DT, GNB, KNN, MLR and SVM are identified and trained using the 182 samples collected. MLR is selected as the best suitable model from the above models because it has 100% accuracy in both training and testing, and 100% accuracy in module 1 and 86.67% in module 2 validation. The ML model was validated by 10 fold cross validation to prevent model overfitting. MLR is known to be more robust and to support fewer data sets, so it can be used for the implementation of ML-powered mobile applications. This machine learning model is incorporated with a mobile application. The ML powered autism diagnostic mobile application allows people to screen and detect ASD in both offline and online modes. Furthermore, Tamil language-based applications make it easier for parents from the Tamil region to assess and screen for autism, while English-based applications can be used by all parents.