Deep Neural Network-based Handheld Diagnosis System for Autism Spectrum Disorder
Keywords: Advanced artificial intelligence, ASD, deep learning, diagnosis, expert system
In the area of psychosis, autism spectrum disorder (ASD) is a complex developmental disorder that includes deficits associated with social interaction, communication, and repetitive or stereotypical behavior. Initially, the prevalence associated with autism was in the range of 3.3 to 16 per 10,000 where intermediate prevalence studies identified 60 autism cases per 10000. It was evidenced that the ASD cases could be constantly increasing day-by-day,, and the dramatic increase (from 4–5 in 10000 to 1 in 59) was recorded in autistic cases in various geographical locations across the world.[5–11] According to the recent report presented by Autism and Developmental Disabilities Monitoring Network in America, the percentage of ASD affected children was 1 in 59. However, as per the statistics provided by the Rehabilitation Council of India, the prevalence rate in India can be 1 in 500 people or more than 2,160,000 people. The short awareness of autism diagnosis in India is a significant cause of the low prevalence rate in India. The early and correct detection of autism is participating crucially in the treatment.
ASD included some significant deficits associated with social interaction, communication, and repetitive or stereotypical behavior. Typical or typically developing children usually behaved as socially interactive, gazing at faces, respond to voices, replying gestures, the immediate reaction of events, attached to caregivers, involved normal in physical activities, attracted to social activities, etc.,, However, in the case of ASD, children could have atypical behavioral characteristics such as socially inactive, slow motor skills, the deficit in formal or informal language, and repetitively/stereotypical actions, or communications. The detailed symptoms of autism included lack of eye contact, lack of social responses, no attachment to caregivers, delayed motor skills, inadequate response to calling, failure to respond gestures, language delay, unusual physical control, unusual coordination, unusual attachment for objects, lack of imaginative, or pretend play, not interested in playing with other children, absence of nonverbal communication, anxious behavior, repetitive responses and so on.,,,,,,
The early and accurate diagnosis of autism could help in converting the abnormal emotional behavior into survivable emotions of individuals with ASD. The aim of this work was to propose the deep neural network (DNN) algorithms based handheld diagnosis system (HDS) for the correct and accurate diagnosis of ASD and identification of its severity level based on detailed DSM-V criteria. The present work could be treated as an extension of the work presented by Khullar et al. as the comparative analysis of the various DNN algorithms for HDS was performed on the basis of performance parameters such as accuracy, loss, mean squared error (MSE), precision, recall, and AUC.,,,,
The significance of initial screening could be critical to conduct the detailed diagnostic procedure to ascertain the concrete and concerned information. Without screening tools and diagnostic criteria, it could be complicated to differentiate between different psychoses issues by nonprofessionals. The diagnosis of ASD using manual tools could depend on the capability of a physician or clinician, where accuracy could be challenged. The requirement of professional ASD screening and diagnostic tools could be a high and basic need for identifying the signs and symptoms through behavioral observations and assessment questionnaires or tests. Various questionnaires and interview tests for diagnosis of ASD were developed based on diagnostic and statistical manual of mental disorders, fourth and fifth edition., The various descriptive diagnosis tools were developed and considered in practice such as the childhood autism rating scale (CARS), modified checklist for autism in toddlers, revised with follow-up (MCHAT-R) and so on.,,,,
In ASD, the correct diagnosis and identification of the disorder related symptoms could be the responsibility of the practitioner to initialize proper treatment. Now, pre-tested artificial intelligent computing techniques could perform diagnosis-related activities without human-initiated errors. With the everyday enhancement in technologies (such as virtual reality, responsive displays, wearable devices, artificial intelligence, internet of things, robotics, etc), automated behavior recognition of exceptional children and identification of causes opened several cards in the area of ASD. To achieve more accurate results in the diagnosis of certain mental disorders, advanced AI-based Computer-Aided Diagnosis (CAD) systems were used for the assistance of clinicians to diagnose different mental disorders which could be able to provide more consistent and accurate results than direct human evaluation with less time.,
Sunsirikul et al. classified ASD individuals on the basis of behavioral factors and patterns using predefined classification based association (CBA) rules. Further, multimodel behavior-related information of ASD was gathered using a socially assistive robotic system designed by Petric et al. and improved objective diagnosis for ASD was developed for a repetitive, consistent, and unbiased evaluation.
Cohen et al., proposed the backpropagation ANN algorithm for the prediction of autism with an accuracy of 90%. Yekta et al. designed an expert system for screening autism on the basis of support vector machine (SVM) and random forest (RF) algorithms and achieved accuracy of 92.4% and 93.1%, respectively. Khullar et al. worked for diagnosis and identification of severity level of ASD individuals using basic artificial neural network (ANN) in which approximately optimum accuracy was achieved at 125 epochs of training with DSM-V based dataset of 128 entries. The state-of-the-art results in the area of natural language processing, handwriting identification, digital signal processing, and so on were identified by using CNN and LSTM algorithms.,,,,
Alhagry et al. worked for EEG signal analysis for emotion recognition using LSTM, and with an advanced artificial LSTM algorithm, 87.99% of average accuracy was achieved. AI algorithms such as a convolutional neural network (CNN) were also used for identification of subject independent smile and satisfactory accuracy of 93% was obtained., With the advancement in deep learning algorithms, more accurate and robust facial emotion detection systems were designed, which satisfied the user's need successfully., LSTM algorithm was also applied to detect suicidal ideation on different data sources and achieved accuracy in 95% during training and 88% during testing.
During the design and implementation of this study, a number of ethical considerations were incorporated likewise informed written consent, right to withdraw for the participants from the study at any time; none of the participants had received any psychoactive medication and anonymity of participants.
Data acquisition and participants
In this study, DSM-V Diagnostic criteria for Autism Spectrum Disorder-299.00 were included for the collection and creation of the ASD dataset for the training of the proposed DNN based system using the method provided by Khullar et al. Initially, binary represented data of 128 entries was collected according to the format mentioned in [Table 1] where false value was represented by zero (0) and one (1) representing the true value. The outcome was also represented in the last column as ASD or Non-ASD identification remarks.
Further, the validation of the proposed system was conducted on the basis of collected data from 40 voluntarily participated caregivers of individuals. Of the two groups of individuals who participated, 20 children were diagnosed as ASD according to DSM-V criteria and another 20 were TD children. All the participants had normal or corrected to normal vision. [Table 2] presented the demographic characteristics of participated groups.
Severity level in ASD
DSM-V for ASD explained the impairments and behavior severity in three different levels as discussed below:
Level 1: The patient could have a milder level of the deficit with noticeable impairments, required support for social interaction and communication and faces difficulty in the interaction, activities switching and behavioral flexibility.
Level 2: Patients required substantial support, having inflexible behavior, change nonacceptance and frequent restricted/repetitive behavior. Reduced and abnormal responses of social interactions, narrow special interests, and odd nonverbal communication reflected with limited initiation.
Level 3: The patient could have been functioning severe impairments, very low social initiation, and lack of responses caused by deficits in verbal and nonverbal communication skills. Level 3 ASD individuals could have extreme difficulty in change acceptance along with restricted/repeated behavior.
The DNN was able to overcome the hurdle in medical diagnosis with its powerful statistical paradigm in recognition of complex behavioral patterns and the ability to maintain high levels of accuracy. The various DNN algorithms such as MLP, 1-dimensional CNN, and LSTM were utilized for the analysis of acquired data and detection of ASD and further, the results were compared for the optimum performance.
The basic multilayer perceptron (MLP) was a traditional and common type of artificial neural network with an input layer, an output layer, and one hidden layer as shown in [Figure 1]., In MLP, backpropagation based supervised learning and a linear sigmoid activation function were applied on each artificial neuron as.
1-Dimensional convolutional neural network
One-dimensional CNN (1D CNN) was one of the most common types of deep-multilayer perceptron (deep-MLP) models with a special structure. Every single neuron in the deep-MLP could have an activation function with more than one hidden layer in the network and an activation function mapped the weighted inputs to the output that created a special structure on CNN. The CNN architecture consisted of three basic layers such as the convolutional layer, pooling layer, and fully-connected layer, which was further connected by a rectified linear activation function for generating output as shown in [Figure 2].,,,,,, Each convolution layer was convoluted with their respective kernel (filter) size as
Where y is signal, f is a filter, N number of the signals and x is output vectors respectively.
The function of max-pooling operation was to apply the reduced size feature maps after every convolution layer and also called down-sampling layer. The leaky rectified linear unit (Leaky ReLU) was used as an activation function for initial activation layers. It added nonlinearity and sparsity in the structure of the network. Softmax function was used to separate output and provided a prediction for each class. The fully connected layer contains the full connections for all activations.
The Leaky ReLU function was presented as:
The Softmax function was presented as:
Long Short-Term Memory (LSTM)
Recurrent neural networks (RNN) worked with loops in them and allowed information to persist. LSTM was the part of deep learning algorithms and a special type of RNN with essential memory cells., LSTM model was capable to remember information over a long period and also deal with the vanishing/exploding gradient problem that was raised during backpropagation in deep artificial neural networks. The chain-like structure of LSTM along with memory cell was shown in [Figure 3]. To add or delete the information in LSTM, the sigmoid activation function (σ) and pointwise multiplication operation (*) gates were formed.
The standard symbols and equations for LSTM are as follows:
Xt is input vector, Ht-1 is previous (cell) output, Ct-1 is previous (cell) memory, Ht is present (cell) output, Ctis present (cell) memory, Ht + 1is next (cell) output, Ct + 1 is next (cell) memory, W and V are weight vector, σ is sigmoid function, tanh is tangent function. * is element wise multiplication function, + is represents element wise addition function.
Raspberry Pi with a 7-inch touch screen as shown in [Figure 4] was utilized to implement the proposed DNN-based handheld ASD diagnosis system using efficient Python programming language., In addition, the designed interactive graphical user interface for easy access by caregivers or clinicians was presented in [Figure 5]. The severity level of ASD individuals could be decided by the system based on integer entered such as low as integer 1, medium as integer 2, and high as integer 3.
HDS implementation detail
The overall implementation procedure for the proposed HDS for ASD diagnosis performed through three stages such as data acquisition, training-testing of optimum algorithm among MLP, CNN, and LSTM, and further implementation and analysis of optimum DNN algorithm on HDS for individuals with ASD. The overall HDS implementation was demonstrated through the flow chart presented in [Figure 6].
Further Stepwise detail of the overall implementation of DNN based proposed HDS is as follows:
Step 1: Initially DSM-V based dataset has acquired
Step 2: Training, testing, and analysis of DNN algorithms such as MLP, CNN, and LSTM for ASD diagnosis based on parameters viz. accuracy, MSE, and loss to find an optimum and best suitable algorithm for implementation on HDS.
Step 3: Implementation of the trained model of optimum algorithm found in Step 2 on Raspberry Pi based HDS.
Step 4: Testing and analysis of optimum DNN based proposed HDS for ASD and TD subjects
Step 6: Statistical validation of the accuracy of cases with t-test
Step 7: Identification of ASD severity level.
This section provided a detailed analysis of the obtained results in three cases regarding training and testing of DNN (MLP, CNN, and LSTM) algorithms, 5-fold validation and analysis of optimum DNN-based handheld diagnosis system and statistical analysis of the implemented results of case 1 and case 2.
Case 1: Training and testing of DNN algorithms - MLP, CNN, and LSTM
In this case, the MLP, CNN, and LSTM algorithm was trained and tested to find the most appropriate and accurate model with least MSE and loss. The testing was conducted by providing the different number of epochs (1 to 150) on the same training and testing data set as shown in [Table 3], [Table 4], [Table 5]. Dataset of 128 binary (0/1) entries was utilized for training and testing. The analysis was conducted based on parameters such as viz., accuracy, mean squared error, and loss as presented in [Figure 7]. The results predicted that the LSTM provided 100% accuracy at a very less number of epochs (50 epochs) as compared to CNN and MLP. It was observed from [Figure 7] that LSTM had achieved 0.02 mean square error and loss at near 50 epochs, which was very low in comparison to MLP and could not affect the results. It was observed that the LSTM algorithm provided best results among all the tested algorithms and best suitable for implementation on proposed Raspberry Pi-based HDS system.
Case 2: 5-fold validation and analysis of optimum DNN (LSTM) based HDS
The trained LSTM model was implemented on Raspberry Pi-based proposed HDS and the proposed handheld diagnosis system was utilized for individuals with ASD. The proposed AI-based HDS was tested with data collected from 20 ASD and 20 TD individuals according to the format presented in [Table 6]. The performance of pretrained HDS for detection of ASD was evaluated through K-fold (5-fold) analysis based on the parameters such as viz., accuracy, mean squared error, loss, precision, recall, and AUC as presented in [Figure 8], [Figure 9], [Figure 10] and [Table 6].
Case 3: Statistical analysis of the implemented results
The analysis of Case 1 and Case 2 was performed with the same training dataset but the testing was performed with different datasets and accuracies Acc-1 and Acc-2 were achieved for Case 1 and Case 2 respectively with different numbers of epochs. The various results for Acc-1 and Acc-2 for Case 1 and Case 2 were presented in [Table 7] with a different number of epochs. The statistical validation of the implementation results of both cases Case 1 and Case 2 was analyzed with a two-tailed t-test as represented in [Table 8]. The hypothesis of the study was considered as follows.
Null Hypothesis- “No significant difference in accuracies achieved from Case-1 and Case-2”
Alternate Hypothesis- “Significant difference in the accuracy percentage achieved in Case-1 and Case-2”
The mean scores of the Case 1 accuracy were found 89.06 which was very much similar to the accuracy of the Case 2 and no significant difference was found at significant value P < 0.05
It was observed from the statistical analysis that Acc-1 of LSTM in Case 1 was similar to Acc-2 of LSTM based HDS in Case 2 and the null hypothesis was accepted.
It was observed from the results that the LSTM algorithm provided the best training results among the tested AI algorithms with maximum accuracy. The validation of LSTM based proposed HDS was analyzed in Case-2 and it was observed from [Table 6], that the proposed system achieved 100% in 50 epochs with 0.02 loss value and almost negligible MSE. Further detailed parametric analysis using precision, recall, and AUC scores resulted as highest on the scale of zero to one. The proposed system provided maximum stability since the statistical analysis of accuracies of LSTM in Case-1 and proposed LSTM based HDS system in Case -2 were similar without significant difference. On the basis of the results found in the analysis, it was suggested that the proposed system was also successfully able to identify the DSM-V based identification of ASD level of severity as high, medium, or low.
The accurate diagnosis of autism plays very crucially in deciding the required correct treatment for individuals with ASD and the role of an artificially intelligent diagnosis system could fulfill the need for accurate diagnosis with negligible errors in spite of manual traditional diagnosis systems.
DNN-based HDS was designed and analyzed in the present work and it was observed that the proposed HDS system resulted in 100% accurate diagnosis of ASD along with the identification of ASD severity level in reference to DSM-V, 299.0 criteria in terms of accuracy, mean square error, and loss. It was suggested that the proposed system could be the best assistant to clinicians in the diagnosis of ASD as compared to the manual/traditional diagnosis systems. In the future, advanced artificial intelligence and evolutionary algorithms would be implemented to design HDS for medical diagnosis.
We are thankful for the IKG Punjab Technical University, India and CT Group of Institutions, India for made this research possible.
Declaration of patient consent
The authors certify that they have obtained all appropriate patient consent forms. In the form, the patient (s) has/have given his/her/their consent for his/her/their images and other clinical information to be reported in the journal. The patients understand that their names and initials will not be published and due efforts will be made to conceal their identity, but anonymity cannot be guaranteed.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5], [Figure 6], [Figure 7], [Figure 8], [Figure 9], [Figure 10]
[Table 1], [Table 2], [Table 3], [Table 4], [Table 5], [Table 6], [Table 7], [Table 8]