Copyright: 2020 Billing et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The complete dataset is available for download through the Swedish National Data Service, -6k13. Sample data and usage instructions can be accessed at In addition, source code for the DREAM RET System, with which the present dataset was gathered, is made available at
Autism Dataset Download
Download File 🔥 https://tiurll.com/2yG9YU 🔥
Competing interests: The stuy from which the manuscript data was collected made use of the Nao robot, which is developed and sold by SoftBank Robotics with whom author A.M. is affiliated. This does not alter our adherence to PLOS ONE policies on sharing data and materials.
Children diagnosed with Autism Spectrum Disorder (ASD) typically suffer from widespread difficulties in social interactions and communication, and they exhibit restricted interests and repetitive behavior [1]. ASD is referred to as a spectrum disease because the type and the severity of the symptoms vary significantly between individuals. At one pole mild difficulties in social interaction and communication, such as problems in the initiation and maintenance of a conversation, the integration of verbal and nonverbal communication and the behavior adaptation to various contexts, together with some behavior rigidity can be seen. The opposite pole is characterized by severe deficits in verbal and nonverbal communication, low level of social initiation, absence of peer interest, strong behavior inflexibility, and restricting/repetitive behaviors [1].
The therapeutic intervention that currently has the most consistent empirical support is Applied Behavior Analysis (ABA) [5]. ABA is a structured intervention following behavioral learning principles, in which reinforcements are manipulated in order to increase the frequency of desired behaviors and decrease the frequency for those that are maladaptive. The discrete trial training (DTT) is a common method employed by ABA treatments, in which the child is presented with a discriminative stimulus for a specific behavior (e.g., an instruction from the therapist), and the child receives a reward if he or she performs the expected behavior. If he or she does not, the therapist might correct the behavior by offering a demonstration or by offering a prompt [6]. In order to be effective, ABA therapies need to be both intensive and extensive and are thus associated with significant efforts from both patients and therapists providing the treatments.
In RET interventions developed on ABA principles, the robot guides the child through a game-like activity in order to develop a behavior that is relevant for social communication, while the therapist supervises the interaction. The robot acts as a model by performing the desired behavior, or as a discriminant stimulus, by giving verbal or non-verbal instructions. The robot also acts as a source of social reinforcement, by providing positive or negative feedback on the performance of the child. The justification for using a robot in this form of treatment relies on the empirical findings indicating that ASD children are learning social behaviors from these interactions and might be more motivated to participate in the intervention as a result of the presence of the robot [17, 18].
Robots have also been proposed as a means for screening, diagnosis, and improved understanding of ASD [19, 20], the potential of which are still not fully exploited due to a majority of research on RAT and RET taking the form of small scale or single-case studies, without the methodological rigor required to make the data applicable in clinical domains [12, 21].
While the clinical results from the evaluation are in the process of being published elsewhere, we here present a public release of the DREAM dataset, made available for download by the Swedish National Data Service, -6k13. Following the ethical approval and agreements with caregivers, this public release does not comprise any primary data from the study. Primary data refers to direct measurements, e.g., video and audio recordings, of children in therapy. Instead, this public release comprises secondary data not revealing the identity of the children. Secondary data refers to processed measurements from primary data, including 3D skeleton reconstructions and eye-gaze vectors.
Further background on data-driven studies of autism and relevant datasets is presented in Section 2, followed by a presentation of the clinical evaluation from which this dataset was gathered (Section 3). Details of the DREAM dataset are provided in Section 4. Finally, the paper concludes with a discussion in Section 5.
Considering the large effort involved in its diagnosis and treatment of ASD, there is an urgent need to better understand the autism spectrum and to develop new methods and tools to support patients, caregivers, and therapists [25]. One initiative was made by Thabtah [26], who developed a mobile application for screening of ASD, based on DSM-5 [1] and two questionnaire based AQ and Q-CHAT screening methods [27, 28]. While this is far from the only mobile application for screening of ASD, we believe this initiative stands out by, in contrast to several other applications, being supported by published research and by sharing parts of the underlying databases publicly [29]. Such datasets, covering for example traits, characteristics, diagnoses and prognoses of individuals diagnosed with ASD could be important assets, and are still very rare.
One example that clearly demonstrates the value of data-driven analysis of ASD is the work by Anzulewicz et al. [30]. The authors report a computational analysis of movement kinematics and gesture forces recorded from 82 children between 3 and 6 years old. 37 of these children were diagnosed with autism. The analysis revealed systematic differences in force dynamics within the ASD group, compared to the typically developed children included in the study. Unfortunately, this dataset has not been released publicly.
Another dataset covering social interaction is the Multimodal Dyadic Behavior Dataset (MMDB) [32, 33]. This dataset comprises audio and video recordings from semi-structured play between one adult and one child in the age of 1 to 2 years. To date, 160 sessions of 5-minute interaction from 121 children have been released. Videos are annotated automatically for gaze shifts, smiling, play gestures, and engagement. An attractive aspect of this dataset is that the raw data streams are provided, including a rich set of 13 RGB cameras, one Kinect (RGBD) camera, 3 microphones, and 4 Affectiva Q-sensors for electrodermal activity and accelerometry, worn by both the adult and the child.
Focusing instead on human-robot interaction, the UE-HRI dataset [34] is a recent example. It includes audio and video recordings of 54 adult participants engaged in spontaneous dialogue with the social robot Pepper. The interactions took place in a public space, and include both one-to-one and multi-party interactions.
In sum, several datasets related to the study of social interaction, human-robot interaction, and autism can be found in the literature. Some are also released publicly, but none of them reach the same size as the dataset we present here. While there are other datasets with a similar, or even richer, set of features, none of these cover children diagnosed with ASD. Under the label Behavior Imaging, Rehg et al. [36, p. 87] explicitly argue for the need for such a dataset:
We believe this approach can lead to novel, technology-based methods for screening children for developmental conditions such as autism, techniques for monitoring changes in behavior and assessing the effectiveness of intervention, and the real-time measurement of health-related behaviors from on-body sensors that can enable just-in-time interventions.
Since children diagnosed with ASD are often sensitive to new clothing and wearable equipment, we consciously avoided on-body sensors. However, in other respects, we hope that the present work constitutes one important step towards a data-driven study of autism outlined by Rehg et al. [36].
The clinical evaluation of RET, from which the present dataset is gathered, was conducted between March 2017 and August 2018 at three different locations in Romania. 76 children, age 3 to 6 years, were recruited to the study, out of which 70 met the inclusion criteria and were randomly assigned one of two conditions, RET or SHT. Participants in both groups went through a protocol of initial assessment, eight interventions, and a final assessment. The effect of the treatment was assessed using the Autism Diagnostic Observation Schedule (ADOS), in terms of the difference between the initial and final assessments [37]. Nine children did not continue the treatment beyond initial diagnosis, e.g., as a result of high skill performance, leaving 61 children with an initial ADOS score between 7 and 20 in the study (RET n = 30, SHT n = 31). A letter of consent was signed by at least one parent before initiating the study, expressing their consent to record the assessment and the intervention sessions and to use the data and recordings for scientific purposes in an anonymous fashion. The clinical study where this data was collected received prior ethical approval from the Scientific Council of Babes-Bolyai University in Cluj-Napoca, Romania, where the trial was conducted (record no. 30664/February 10th, 2017). The clinical trial was pre-registered to the U.S. National Library of Medicine database (ClinicalTrials.gov) under the number NCT03323931.
The therapy environment followed two configurations illustrated in Fig 1. The two configurations (RET and SHT) were designed to be as similar as possible, with the interaction partner constituting the primary difference. A therapist was present during both conditions, seated at the side of the table. A picture from the RET condition is shown in Fig 2. The exact setup varied slightly between different tasks. Some tasks made use of a touch screen placed between the child and the interaction partner, referred to as a sandtray [38]. Other tasks had a table as illustrated in Fig 1. 152ee80cbc
concrete mix design spreadsheet free download
creeper aww man sound effect download