Development and Validation of a System for the Anticipation of Challenging Behaviors of People With Autism Spectrum Disorder Based on a Smart Shirt: a Mixed-method Design
Development and Validation of a System for the Anticipation of Challenging Behaviors of People With Autism Spectrum Disorder Based on a Smart Shirt: a Mixed-method Design
Most people with autism spectrum disorder (ASD) present at least one form of challenging behavior (CB). Self-injurious, aggressive, and disruptive CBs linked with social interaction, community-based service exclusion, and a life quality reduction for people with ASD, their caregivers, and health professionals. The current study has three objectives: 1) to assess the differences in the physiological reaction of high-functioning adults with ASD and typically developed peers, using bio-signal measurements such as heart rate derived from wearable Smart Shirt (SS), 2) to learn which physiological parameters can best predict the imminent onset of a CB, and 3) to develop a system able to predict the incoming occurrence of a CB in real-time and inform the caregiver through an alert notification sent on a smartphone application.
Methods and analysis: comparison between physiological parameters will carry out with two groups of 20 participants with and without ASD. Each participant will be asked to watch two five-minute videos while wearing the SS: one showing relaxing images and the other impressive human body deformities. To identify the matching between the physiological parameters variation collected by the SS and the CBs, ten participants with ASD and aggressive or disruptive CBs will be recruited. Each of these participants will wear the SS for seven consecutive days during waking hours, performing their usual daily activities. During the same seven days, the caregivers who care for the participant will fill a behavioral diary with the participant's status, reporting the times of the day in which he is quiet, agitated and the occurrence of CBs. A learning algorithm capable of predicting immediate CBs occurrence based on physiological parameter variations will be developed together with an ad hoc smartphone application. If the algorithm detects the possibility of an incoming CB, a notification will be sent to the caregiver's smartphone to inform of the possible advent of a CB, therefore enabling the implementation of the selected intervention strategy. After developing the algorithm and related smartphone application, a system efficiency proof of concept (POC) will be carried out with one participant with ASD and CB for seven days in a special school setting with healthcare professionals and teachers. A focus group including health professionals will be conducted after the POC to identify the strengths and weaknesses of the developed system.
Study goals
The current protocol consists of three phases, each with a specific goal. The first aim is to assess the differences in physiological reaction measured with a Smart Wearable Shirt (SWS) between adults with high-functioning Autism Spectrum Disorder (ASD) and typically developed peers. The second goal is to learn which physiological parameters can best predict the imminent onset of a Challenging Behavior (CB). The third aim is to develop a system able to predict the incoming occurrence of a CB in real-time and inform the caregiver through an alert notification sent on a smartphone application.
Study design
An observational study design will be implemented in the first two phases of the current research. In phase one, participants' physiological reactions to two visual stimuli will be collected and analyzed. In phase two, the physiological characteristics of CBs presented by people with ASD will be collected. A single case study with a mixed-method design will be implemented in phase three, where the system validity proof of concept will be performed.
Ethics and safety issues
The research proposal was approved by the Ariel University institutional review board (AU-HEA-ML-20201203). The study will be carried out following the Declaration of Helsinki principles. At the recruitment, written informed consent will be collected from all participants or their legal guardians. The SWS planned to use in the current protocol is a non-invasive medical device. However, if a participant does not tolerate SWS, he can withdraw from the study at any time without any repercussions.
Participants
According to a sample size calculation analysis, a group of 20 subjects diagnosed with high functioning ASD aged between 20 and 40 years residing at home (observation group - OG) along with an age and sex-matched control group (CG) of 20 typical developed peers will be enrolled in the first protocol phase. For the study's second phase, 10 people with ASD presenting aggressive or disruptive CBs aged between 20 and 40 years and their caregivers will be recruited. Finally, one participant with ASD, aged between 20 and 40 years attending a special school and living at home, which exhibits aggressive or disruptive CBs, will participate in the third phase of the research.
Procedure
Phase one - comparison between people with and without ASD physiological outcomes.
For the first phase of the current protocol, the physiological parameters of the people in the OG and CG will be acquired using the SWS while participants watch two five-minute videos. One video will show relaxing images and will emit relaxing music (relaxing video). The second video will present impressive human body deformities accompanied by anxious music (disturbing video). Both videos will be presented to the participant in a sitting position. Before starting the relaxing video, the participant will be invited to relax and lean back onto the chair's backrest. The participant can close his eyes or keep them open at his discretion to promote relaxation. Participants will be asked not to lean against the chair backrest and keep their eyes open for the duration of the disturbing video watching. The first video watched by each participant will be chosen randomly between the two videos. The duration of the entire session will be approximately 15 minutes.
Phase two - Classify the physiological parameters' variations in people with ASD.
Each participant enrolled in phase two of the protocol will wear the SWS for seven consecutive days during waking hours, performing his usual daily activities. During the same seven days, the caregivers who care for the participant will report the participant's status in the behavioral diary. Each evening, data collected by the SWS will be uploaded to an online cloud along with the behavioral diary of the day. Once the data from all the 10 participants have been collected, a Deep Learning (DL) algorithm will be developed to learn the variations in the individual's physiological parameters that occur before a CB and predict future CB. Moreover, a smartphone application will be developed to receive the SWS data in real-time and send it to a remote server where it is analyzed through the developed algorithm and the classified CB events will be extracted and presented on the applications' notification screen. In other words: if the algorithm detects the possibility of an incoming CB, a notification is sent to the caregiver's smartphone to inform of the possible advent of a CB, therefore enabling the implementation of the selected intervention strategy.
Phase three - System proof of concept.
For one week, the developed system prototype and its efficacy will be tested on one participant with ASD for seven days with healthcare professionals and teachers in a special school setting. The participant will wear the SWS during all hours of attendance at the special school. At the end of the seven days, the QUEST 2.0 will be administered to each professional and teacher who will interact with the system, and a focus group will be carried out with them to address the research questions mentioned above. The focus group will aim to answer the following research questions:
Information obtained from the QUEST 2.0 administration will be discussed within the focus group. In the last part of the focus group, a summation of solutions to each research question will be proposed to the group, and the number of participants that agree or disagree with the proposed summation answers will be collected.
Data analyses
Phase one
Data collected by the SWS from participants in the OG and CG will be analyzed and compared. From ECG received data, heart rate (HR) will be calculated between two consecutive QRS complexes. Considering the time interval between two QRS complexes as "t", the corresponding temporal HR will be 60/t. A percentage threshold value will be set using the sliding window method, and the minimum allowed peak width will be identified to remove unwanted artifacts from HR. The removal process will be performed for positive and negative peaks in two rounds. A window will be slid over the HR signal, and its median value will be calculated. The maximum (positive and negative) allowed peak amplitude will be determined for every window by multiplying the window's median value for a threshold value. The threshold value for positive peaks was set at 30% (for the first removal round) and 25% (for the second removal round) of the window's mean value. The threshold value was set for negative peaks at 50% (for the first removal round) and 30% (for the second removal round) of the window's mean value. Then, all peaks with amplitude bigger than the allowed value will be identified from every window. If one of these peaks is narrower than the minimum allowed peak width, it will be replaced with the reference window median value. Otherwise, if an identified peak width is bigger than the allowed peak width, its value will be replaced with the maximal allowed HR (for positive peaks) or the minimal allowed HR (for negative peaks). The maximal allowed HR will be calculated with the following formula: "209-(0.7×(Participant age))". The minimal allowed HR will be 60 bpm. After removing abnormal peaks, the signal will be filtered with a Gaussian filter with a sigma equal to 1. After the HR signal filtering process, the obtained cleaned HR signal will be used to classify the participant stress within the following levels: "no stress", "mild stress", "moderate stress", and "high stress". Each stress level will refer to an HR signal positioned within a specific range of values. The "no stress" level will include the HR values below 90% of the cleaned HR signal's lowest peak. If this HR value is lower than The "high stress" level will comprise values above 90% of the cleaned HR signal's highest peak. If this value is above the maximal allowed HR, it is substituted with 90% of the maximal allowed HR value. The range left between these two thresholds will be divided into two equal parts (lower and upper half). The HR data positioned in the lower half of this range will be classified as "mild stress" and those positioned in the upper half as "moderate stress". Each HR value will be classified and assigned with a numerical value corresponding to a stress level ("no stress" = 0, "mild stress" = 1, "moderate stress" = 2, and "high stress" = 3). After acquiring the sequences of the stress levels of all participants of Phase one, the sequences of the subjects in the OG and CG will be compared using a version of the Smith-Waterman algorithm adapted for the analysis of the obtained data.
Phase two
The data gathered by the SWS from participants enrolled in phase two will be analyzed as described above, and a DL algorithm will be developed in order to predict the incoming participants' stress levels. To find CBs patterns among subjects, a classifier based on supervised learning to find anomalies in the subject's data that might indicate on CB that is about to occur will be constructed. A long-short term memory (LSTM) algorithm will be taught to recognize data patterns corresponding to CBs occurrence using the data collected by the participants' caregivers through the behavioral diary and the information collected in Phase one. LSTM is an extension of the recurrent neural network (RNN). In contrast to other applications of Machine Learning (ML) and DL, in the process of analyzing and predicting time-series information, each data point is based on previous information, which must be examined as well. RNN is the most used network for time series applications since it can form the target vector by observing the current input data history, using shared wights among the hiding units of the network across each time step of the data. The authors choose LSTM, not the RNN, because RNN has a significant vanishing gradient problem. The gradient of output error is based on previous inputs vanishes when time lags between inputs and errors increase. To overcome this problem, the LSTM is introduced. LSTM has a memory, which comes to practice by replacing the nonlinear units of RNN in the hidden layers with memory blocks. The network propagates errors throughout the entire network, and as a result, it can learn long-term dependencies and forget unnecessary information based on the data at hand. The accuracy of the prediction model will be calculated according to standard estimation methods such as the confusion matrix and the area under the curve (AUC) Values. These values range from 0.5 to 1, with 1 being perfect classification and 0.5 being no better than luck.
Phase three
The themes that will emerge from the focus group will be extracted from the discussion transcription. Axial coding strategy will be applied to calculate the extensiveness of each theme discussion. This qualitative data analysis consists of assigning a reference number to each theme and then marking any sentence related to that theme with that number. A reliability check for the code-to-sentence matches will be applied by giving the list of codes to an independent researcher experienced in qualitative analysis and asking him to identify the sentence that matches each code. The level of agreement to summation answers to the research questions will be obtained by calculating the percentage of participants that agree with the proposed statement. The authors will discuss the developed answers in the light of the relevant emerged themes and the level of agreement of the discussion group.
Inclusion Criteria for Observation Group One:
Exclusion Criteria for Observation Group One:
Inclusion Criteria for Observation Group Two:
Exclusion Criteria for Observation Group Two:
Inclusion Criteria for Proof of Concept:
Exclusion Criteria for Proof of Concept: