A systematic review of telehealth screening, assessment, and diagnosis of autism spectrum disorder

Study characteristics

The search yielded 627 articles; 8 additional articles were obtained from other sources, for a total of 635 articles. After removing 91 duplicated articles, 544 were chosen for title and abstract screening. A total of 55 articles were chosen for full-text review. Following a thorough examination, 29 articles were excluded for the following reasons: treatment or intervention studies only (n = 5); review articles (n = 9); focus on algorithms (n = 14), and studies for purposes other than screening and diagnostics (n = 1). Thus, 26 articles remained in the systematic review and underwent subsequent data extraction (see Fig. 1 and Table 1). The quality of the included articles was assessed using the SMRS, which revealed that the overall quality was good, with a mean score distribution of 3 or more; only two studies received a score of 2, indicating the need for more scientific rigor [40, 45]. Only English-language articles were considered for this study; thus, out of the 26 studies, most were conducted in the US. Two studies were conducted in the United Kingdom, 2 in New Zealand, and 1 in Australia. In addition, of the 17 applications reported in this review, 4 were developed by companies, and 10 were based on university-funded projects. The website for each application is shown in Table 1.

Table 1 Summary of study characteristicsTelehealth method

Telemedicine provides access to expert medical care and support without the need to meet in person [4]. When using a video-based approach to screen for ASD, two main technological approaches apply: real-time and store-and-forward methods. Real-time interactions allow patients to communicate with healthcare providers in real time, primarily through video conferencing, whereas store-and-forward interactions do not necessitate the simultaneous presence of patients and healthcare providers [27]. Scales and questionnaires used to remotely screen for ASD can be divided into two approaches: static scoring and machine learning (ML). Static scoring is the online use of conventional questionnaire scoring, while machine learning is the use of online ML algorithms to score questionnaires more individually and flexibly [40]. This review elaborates on telemedicine studies in ASD focusing on three purposes: screening, assessment, and diagnosis. Some applications can satisfy more than one purpose.

Screening

Eleven ASD screening applications were summarized in this study. Overall, most of the applications are well-established. Cognoa, Autism&Beyond, ASDtests, ASDetect, AutismAI, and VIRSA are all mobile-based applications that can be downloaded directly to a smartphone or tablet [5, 15, 20, 40, 45, 49]. Although the TEDI (Telehealth Evaluation of Development for Infants), CHICA (Child Health Improvement through Computer Automation), Lena and SORF (the Systematic Observation of Red Flags) applications are not web-based, empirical studies have demonstrated the feasibility and validity of these tools [7, 13, 14, 44]. TEDI is a novel tool for remotely assessing the behavioral characteristics of children with ASD, and CHICA is a computer-automated decision support system that combines screening questions and medical history data displayed on a tablet to conduct remote screening for ASD. In addition, one more application was recommended but has not been adequately developed [48]. These tools involved four different types of processing, namely, (a) static scoring, (b) machine learning, (c) storage and uploading, and (d) real-time methods. Table 1 illustrates the methods used for each screening application.

Four applications used the store-and-forward approach (Cognoa, Autism&Beyond, Lena, and SORF), and one application used the real-time approach (TEDI). Cognoa is a machine learning-based application that analyzes two short (1 min and 30 s to 5 min) home videos recorded by parents, along with caregiver questionnaires and clinical question sets from primary care physicians, to generate “ASD positive,” “ASD negative,” or “no result” reports [28, 29]. The Autism&Beyond application mainly uses the camera to record the child’s facial features while watching a short video and combines surveys to screen for ASD [6]. Lena is a wearable device that allows teachers or parents to record a child’s language environment remotely and analyzes recorded language information features to screen for ASD [32]. SORF is an observational coding system for identifying red flags (RFs) for ASD in young children, also allowing remote screening for ASD via naturalistic video-recorded home observation [13]. Parents receive verbal and written instructions for interacting with their children and engage in suggested activities for approximately 5–10 min per day for no more than 1 h. Behaviors in the videos are coded by trained coders using SORF to predict ASD risk. Only one tool, TEDI, employed the real-time approach. TEDI uses video conferencing to screen ASD by directly observing children’s interactions during 10 activities [44].

The other six screening tools used scales or questionnaires to screen for ASD, with ASDTests, ASDetect, and CHICA scoring questionnaires statically only and AutismAI, VIRSA, and Wingfield’s application all using machine learning to score [5, 14, 40, 45, 48, 49]. Compared to static scoring, the machine learning approaches yielded slightly higher sensitivity and specificity. For example, AutismAI reported a sensitivity of 0.955 and a specificity of 0.986, and ASDTests reported a sensitivity ranging from 0.922 to 0.98 and a specificity ranging from 0.85 to 0.997. The machine learning application proposed by Wingfield et al. [48] used the PAAS checklist (Picture Autism Assessment Scale) to screen ASD in infants aged 2–4 months and reported a sensitivity of 0.88 and a specificity of 0.96. VIRSA, on the other hand, uses videos instead of questions to allow caregivers to select the most relevant videos of the child’s behavior; thus, is easier for the user to understand and make more accurate judgments. VIRSA was reported to be more sensitive for ASD screening (sensitivity = 1).

Diagnosis

One application and three other studies refer to remote ASD diagnosis. NODA is a well-established application that can be downloaded and used directly. Other studies only investigated the feasibility of telemedicine applications from different perspectives. Among them, the BOSA (The Brief Observation of Symptoms of Autism) approach proposed in the study by Dow et al. [13] is very promising, but further research is still needed. NODA (Naturalistic Observation Diagnostic Assessment) is a store-and-forward application that consists of 2 main components [41]. (1) NODA Capture enables parents to use their mobile phones to upload short videos of their concerns and their child’s behavior in specific situations, such as family meals, playtime with others and playtime alone, for sharing remotely with a clinician. The child’s play-based behaviors and social communication abilities are primarily shown in the first three scenarios, while the parent’s concern about the child’s behavioral symptoms is primarily shown in the final scenario. Parents set up the environment according to the provided scenario descriptions and sample videos. All recordings (including the child’s developmental history) are uploaded to the portal. (2) NODA Connect is a portal that enables physicians to diagnose the child based on the home video and developmental history and to link the marked behavioral characteristics in the video (e.g., no eye contact) with DSM-5 criteria. It is possible to inform parents and other medical professionals about the diagnosis findings through this portal [41].

For ASD diagnosis, Reese’s, Juárez’s and Dow’s teams all used real-time video conferencing [19, 37]. The Reese et al. study was assessed primarily by combining (a) 10 min of observation of the child’s performance in an unstructured setting, (b) activities from the revised ADOS-2, (c) a structured interview using the ADI-R (Autism Diagnostic Interview-Revised), (d) a medical history review, (e) feedback and advice to the family, and (f) a satisfaction survey. Juarez et al. assessed ASD using (a) a brief interview, (b) the implementation of STAT (Screening Tool for Autism in Toddlers & Young Children), and (c) a focused diagnostic ASD interview. Clinicians interviewed caregivers directly via video conferencing and observed the child’s behavior at STAT for ASD diagnosis. The diagnostic procedure of the Dow et al. study included (a) the ADOS–Toddler Module (ADOS-T), (b) a 1-h home observation coded by the BOSA (The Brief Observation of Symptoms of Autism) manual, (c) assessment of developmental level using the Early Screening for Autism and Communication Disorders (ESAC) and Mullen Scales of Early Learning (MSEL), and (d) adaptive behavior assessed by the Vineland Adaptive Behavior Scale-Second Edition (VABS-2) from a caregiver interview.

The main difference between the two studies was that Reese et al.’s study focused on comparing real-time video conferencing with an in-person assessment to show the feasibility and reliability of the online ASD assessment, while Juárez’s study provided a detailed description of the procedure for telemedicine diagnosis of ASD. Reese et al.’s study found that there was still excellent diagnostic agreement between clinicians and other teams in the video conferencing setting. Real-time videoconferencing achieved the same results (sensitivity = 0.84, accuracy = 0.88) as the psychometric properties of in-person assessments (sensitivity = 0.88, accuracy = 0.78), consistent with a previous review [37, 42, 43]. In contrast, Juárez et al. studied the preliminary feasibility, accuracy, and clinical utility of diagnosing ASD via televideo conferencing. This study showed that telemedicine procedures were as capable as in-person assessments at identifying children diagnosed with ASD. Although 13% of the participants could not provide a diagnosis [19], it is undeniable that video conferencing could improve the efficiency of ASD diagnostics.

Another application that also uses real-time video conferencing is BOSA, which was proposed by Dow et al. [13]. BOSA begins with a 12–14-min interaction between the individual and a clinician or caregiver, followed by observation by the trained clinician via videoconference and objective coding according to the BOSA manual, leading to a diagnosis of ASD. This application has the advantages of ADOS-2 in that it allows a relatively natural and consistent context to observe social-communicative behaviors and perform standardized coding while avoiding the drawbacks of ADOS-2, such as the long diagnostic time and the inability to perform remote assessments [13].

Screening and diagnosis

One application, TeleNP, performs both screening and diagnostic functions. TeleNP is a real-time practice model encompassing almost the entire patient assessment process from the initial screening visit to the subsequent diagnostic visit and is currently being evaluated in two studies [36, 39]. This complete model includes the process of online clinic intake, history collection, appointment visits, and diagnostic assessment. First, patients can upload information directly to the online platform to generate an electronic medical record. After receiving the medical file, a questionnaire for an ASD screening appropriate for the child’s age and area of concern is sent via email. Caregivers then complete the questionnaire and are given verbal instructions over the phone regarding how to create a distraction-free environment to maximize the child’s participation in the testing procedure. The ideal room or seating area is determined based on the child’s behavior. For example, sitting on the parent’s lap in the living room or setting up the bedroom with preferred toys on the floor. The examiner asks the parents about the child’s preferred toys and any fixations during the appointment to help promote a positive relationship and reduce the likelihood of emotional outbursts. Finally, after a 20-to 30-min child interview and observation period, the examiner conducts a 20-to 30-min rapid diagnostic interview with the caregiver based on DSM-5 criteria.

Assessment

Only one instrument, which is neither a screening nor a diagnostic tool, is used to assess ASD core behaviors. TELE-ASD-PEDS is a core behavioral remote assessment tool developed for young children at high risk for ASD and has been discussed in three studies [10]. TELE-ASD-PEDS assesses ASD via real-time video conferencing and focuses on young children aged 1–3 years who are not yet verbal. Trained examiners assess children based on their behaviors during different activities via real-time video conferencing. A machine learning algorithm selects the 12 most predictive ADOS-2 activities and applies them to a parent-led interactive task. During implementation, clinicians rate the child’s behavior on seven items via video conferencing (e.g., eye contact and unusual repetitive restricted interest/stereotyped play) and use a dichotomous (yes/no) and a Likert scale to rate the severity of symptoms (3 = behavioral features of ASD present; 2 = probably atypical behavior; 1 = behavioral features of ASD not present). Although the TELE-ASD-PEDS is not designed for screening and diagnosing ASD, preliminary data suggest that it is a useful and valid ASD diagnostic tool [42, 43].

Observed behavioral indicators

ASD screening, diagnosis, or assessment applications rely on ASD behavioral indicators to allow parents to report a symptom or for a clinician to give a score. Although the 17 applications had different purposes, the behavioral indicators they observed fell into two broad categories: comprehensive and noncomprehensive. Comprehensive behavioral indicators usually cover the full dimension of ASD characteristics stated in DSM-5, including social communication and interaction deficits, restricted and repetitive patterns in behavior and interest, developmental histories and related information. Noncomprehensive observed indicators only focus on some core behavioral characteristics of ASD.

Most applications reviewed here used comprehensive observational indicators, and three studies used noncomprehensive observational indicators in the screening tools. It is worth noting that most of the screening tools employed extensive observational indicators. Cognoa, a screening tool, not only collected information from the caregiver’s perspective and flagged core behavioral characteristics from videos but also received additional information from clinicians for screening. Moreover, TEDI used a real-time approach to observe different core behavioral characteristics in 10 scenarios, and SORF used the store-and-forward method to collect videos to code 22 items, with 11 items in each of the two DSM-5 symptom domains—Social Communication (SC) and Restricted Repetitive Behaviors (RRB). ASDTests, ASDetect, AutismAI, CHICA, and the studies by Wingfield et al. used established screening tools or combined them with their developed questionnaires, with the most commonly used screening tools being The Modified Checklist for Autism in Toddlers (M-CHAT) and the 10-item Autism-Spectrum Quotient (AQ-10).

All diagnostic tools used a comprehensive set of observables. The DSM-5 diagnostic criteria were used by NODA, TeleNP, and the tool proposed by Juárez et al. to diagnose ASD. Although BOSA uses a standardized coding manual developed in-house, the score still corresponds to the DSM-5 checklist and the ADOS-2 score. Moreover, TeleNP also used the childhood autism rating scale (CARS-2), the NEPSY second edition (NEPSY-II), Delis-Kaplan executive function system (DKEFS), vineland adaptive behavior scales, third edition (VABS-3), and autism diagnostic observation schedule, second edition (ADOS-2). Juárez et al. used STAT, clinical best estimate (CBE), Mullen scales of early learning (MSEL), VABS-2, and ADOS-2 as other observables. The most commonly used diagnostic tool is the ADOS-2. Assessment tools such as the TELE-ASD-PEDS and the study by Reese et al. also used a full range of observables, both of which used the ADOS-2; Reese et al. also used the Autism Diagnostic Interview-Revised (ADI-R) for assessments.

Only three screening applications used noncomprehensive observables. Autism&Beyond, Lena, and The Video-referenced Infant Rating System for Autism (VIRSA) examine only one core behavioral characteristic of ASD [8, 32, 49]. Autism&Beyond examines the child’s facial features and uses an algorithm to code the child’s social emotions while watching the video. Lena focuses on child’s language environment and screens ASD from a language development perspective. VIRSA, on the other hand, assesses ASD from the standpoint of social skills. A total of 268 videos of social activity were collected so that caregivers could view them in a mobile app and choose the one that best matched the child’s social communication behavior.

Psychometric properties

We identified 26 studies describing the telemedicine application they used and their observed indicators for screening, assessing, or diagnosing ASD. Table 2 shows the demographic information (e.g., sample size, age and sex) of the individuals studied and key information regarding the instruments’ psychometric properties. The age range of the participants in the studies varied widely. Fourteen studies included children between the ages of 18 and 72 months, 4 included children between the ages of 3 and 18 months, 6 included children with ASD aged 6 or older, and 2 studies did not provide sufficient information. Seventeen studies had > 50% male representation, and sex information was missing in 4 studies (see Table 2). Four studies reported all four metrics considered relevant to the instrument (i.e., sensitivity, specificity, positive predictive values and negative predictive values), and 15 papers provided data on at least one psychometric property. Among the studies reporting psychometric information (n = 15), 8 reported sensitivity and specificity values equal to or greater than 75%. However, it should be noted that sensitivity values below this threshold may not indicate poor psychometric properties as the tool may be reliable for detecting specific subgroups of ASD patients [11, 34].

Table 2 Analysis of the studies included in the review

Eight screening tools and four diagnostic tools existed among the 15 studies that reported psychometric attributes. Of the eight studies with screening tools, only the ASDTests study had a small sample size (n = 20), while the others were above 100. VIRSA had the highest sensitivity (sensitivity = 1), AutismAI, ASDTests, Autism&Beyond, and Cognoa had great sensitivity (exceeding 0.9), Wingfield’s proposed application’s sensitivity was 0.88, and Lena’s was 0.75. SORF had the lowest sensitivity (0.70). The proposed applications from AutismAI, ASDTests Lena, and Wingfield all had good specificity (> 0.9); Cognoa, SORF, VIRSA, and Autism&Beyond had specificities of 0.79, 0.67, 0.53, and 0.38, respectively.

Of the four studies of diagnostic tools, only the study by Dow et al. [12] had a sample size greater than 100, while the other three studies had sample sizes less than 100, with Reese et al. [37] including only 17 participants. In addition, the study by Dow et al. had a better sensitivity (0.86–0.96) than that by Reese et al. [37] (0.88), while the application studied by NODA and Juárez et al. [19] presented a sensitivity between 0.79 and 0.85. However, both NODA and Juárez et al. [19] reported specificities greater than 0.94, while the specificity of BOSA fluctuated between 0.70 and 1.

留言 (0)

沒有登入
gif