This study was based on data collected in a laboratory-based setting while children performed a variety of tasks with and without technology.

The study gained ethical approval from the Curtin University Human Ethics Committee (HRE2022-0157). Written informed consent was obtained from caregivers and age-appropriate assent (written or verbal) was obtained from children.

Children with typical development aged 3 to 14 years were recruited through university networks, word of mouth, flyers, and social media posts. Children with parent-reported psychological or physical clinical diagnoses that could affect their ability to follow instructions or perform tasks were excluded.

Recruitment was sex- and age-stratified for balanced representation across girls and boys and the following three age brackets: aged 3 to 5 years to represent a preschool population, aged 6 to 10 years to represent a primary school population, and aged 11 to 14 years to represent secondary school or early adolescent population. Age categories were defined to reflect transitions at school leading to different day-to-day structures in their environment [31,32] and acknowledge technology engagement differences across childhood [33,34].

Data were collected between July 2022 and October 2022. Children attended a single approximately 1-hour data collection session at a Curtin University laboratory with their caregivers either alone (n=10) or with a friend or sibling (n=38). At each data collection session, 2 to 3 researchers with experience working with children were present. The laboratory was set up with different activities aimed at capturing different tasks with and without technology. Caregivers provided informed consent and completed a brief sociodemographic questionnaire. Depending on their age, children were then asked for verbal or written assent and subsequently fitted with the wearable camera.

One Brinno TLC120 Automated Wearable Camera operating at a 1-second frequency was fitted to the child’s chest using an adjustable chest-mounted harness. The Brinno camera weighs 101 g, is 60×60×35 mm in size, captures a 112° field of view. The Brinno camera was used to capture technology exposure during the data collection session and has previously been used in adolescent [21,22] and adult populations [35].

After becoming familiar with the laboratory and equipment, children performed a range of tasks for 2 to 3 minutes each. Tasks were generally completed in a standard order, but adjustments were made to maintain participant enjoyment. Participants engaged in tasks such as watching shows on a television or laptop, gaming, reading, drawing, and using a smartphone for taking photographs. Children received specific instructions for the postures or movements required for each task, allowing the Brinno camera to capture images from various positions and movement patterns. Table S1 in Multimedia Appendix 1 provides a full overview of the tasks completed by the children and the corresponding technological devices.

Among the 48 included participants, 21 were boys and 27 were girls (overall mean age 8.5, SD 3.3 y; boys: mean age 8.2, SD 3.2 y; girls: mean age 8.7, SD 3.5 y). A total of 221,226 images were coded by humans and processed through the Square Eyes model.

Humans detected technology in 92,745 (41.9%) images, of which Television (n=40,231; 43.4%) and Laptop (n=12,717; 13.7%) were the most frequently coded as Technology 1 (Table S3 in Multimedia Appendix 3). The Square Eyes model detected technology in 101,075 (45.7%) images, with Television (n=23,590; 23.3%) and Computer monitor (n=17,766; 17.6%) being the most commonly detected as Device 1 (Table S4 in Multimedia Appendix 3).

Both humans and the Square Eyes model detected a technology in 72,470 images (ie, true positive rate=32.8%) and no technology in 99,876 images (ie, true negative rate=45.1%). A technology was detected by humans only in 20,275 images (ie, false negative rate=9.2%) and by the Square Eyes model only in 28,605 images (ie, false positive rate=12.9%). Thus, overall sensitivity was 78.1%, specificity was 77.7%, and overall accuracy for detecting technologies was 78.0%.

Table 2 shows the frequency distribution of the 6 “any technology combination” categories for images where both the humans and the Square Eyes model had detected a technology (n=74,470). The most frequently coded technologies based on the human coding were Any television (n=32,662) and Any laptop (n=17,880). The Square Eyes model most frequently detected Any television (n=21,762) and Any computer (n=34,530). Overall, there was a below chance agreement between the human coding and the model. Mean percentage agreement was 32.8% (14.0%‐54.3%) and the PABAK score ranged between –0.719 and 0.086. The highest level of agreement was observed for Any television (n=19,148, 54.3%; PABAK=0.086), while the lowest level of agreement was found for Any gaming (n=1013, 14.0%; PABAK=−0.719).

The N-back algorithm (with n=2) flagged 148,546 (67.2%) images that would require human review and 72,680 (32.8%) images were not flagged.

Among the flagged images, a technology was coded by humans only in 13,131 (8.8%) of the flagged images, and a technology was detected by the model only in 28,605 (19.3%) images. Neither the humans nor the model detected a technology in 34,340 (23.1%) of the flagged images. In 72,470 (48.8%) of the flagged images, both humans and the model detected a technology. When considering the images not flagged for human review, 7144 (9.8%) images had a technology coded by humans only, while 63,536 (90.2%) images did not have a technology coded by humans. None of the images had a technology detected by the model only.

Table 3 shows the proportion of human-coded technologies (based on Technology 1) that were not flagged for human review by the N-back algorithm. Among the images that were not flagged, the most frequently missed technologies (based on prorata percentages) were Smartwatch (n=115, 46.7%), Single controller (n=62, 19.8%), and Tablet (n=1518, 15.1%).

Figure 1 illustrates the trade-off between sensitivity, specificity, and the proportion of images flagged by the N-back algorithm as the N-back buffer size increases. Specifically, the figure illustrates that as the N-back buffer size increases, sensitivity increases, and the proportion of images flagged for human review increases while specificity decreases.

Detection performance by the Square Eyes model on blurred and nonblurred versions of 5667 images from a single participant was examined. The model detected technology on slightly more blurred images than on their nonblurred counterparts. Specifically, the model detected a technology on 173 blurred images, which was not detected when the images were nonblurred. In contrast, 60 images were classified as having a technology when nonblurred but not when they were blurred.

Figure 2 represents an example where the model detected a technology only in the blurred version. The bounding box for the model predictions is shown as a blue square, illustrating that the model detected a Mobile phone. Figures 3 and 4 show examples of the alternative blur styles. When the blur was applied either as a square without border (Figure 3) or in a circular shape without border (Figure 4), the model no longer detected technologies in these images. Thus, this suggests that the specific characteristics of the original blur effect influenced the model’s misclassification rather than the presence of blurring itself.

This study aimed to evaluate the performance of the Square Eyes model as a screening tool for identifying technologies in wearable camera images among children for further review by humans. The Square Eyes model demonstrated good agreement with human coders in detecting technologies with an overall accuracy of 78.0% and a false negative rate of only 9.2%. However, when considering specific technologies, lower levels of accuracy were found for some devices (eg, Smartphone and Computer). The model’s N-back algorithm effectively flagged images that required further human review. Specifically, of the complete sample of 221,226 images, only 7144 were not flagged for screening and contained a human-coded technology, thereby reducing the false negative rate from 9.2% to 3.2%. Our explorative examination of the potential influence of the image blurring method on the model performance revealed that using a square blurring with border could have reduced the model’s ability to accurately detect technologies.

When assessing the Square Eyes model’s ability to detect technologies, it should be noted that the model was intended to be used as a screening tool for reducing researchers’ burden of manually annotating wearable camera images rather than as a replacement for human coding. Nevertheless, the model showed good agreement with human coding, with sensitivity of 78.1%, specificity of 77.7%, and overall accuracy of 78.0%. Thus, the model failed to detect a technology on 9.2% (n=20,275) of all images. Given that each image was taken at a 1-second frequency, this corresponds to approximately 338 minutes of missed technology exposure out of a total of 1546 minutes of technology exposure as identified by humans. Moreover, differences were observed in the performance for detecting specific technologies. We found that the best agreement between the Square Eyes model and human coders was for Television (54.3%, n=19,148) and Laptop (44.5%, n=8492). In contrast, lower levels of agreement were found for Smartphone (31.3%, n=2600), Computer (27.5%, n=9611), Tablet (25.1%, n=3685), and Gaming devices (14.0%, n=1013). Thus, a relatively high proportion of handheld devices was not identified by the model. This finding aligns with known limitations of object recognition models that tend to show lower accuracy in detecting smaller objects compared to larger objects as these are often partially occluded and with limited discriminative visual features [38]. The poorer performance for detecting Computer by the Square Eyes model may have been a result of overlap with related categories being present in the image (eg, Laptop, Monitor, or Keyboard), which complicates correct classification. The poor detection of particularly handheld devices should be acknowledged as a key limitation of using the Square Eyes model as a standalone tool given that such devices are commonly used among children and adolescents [41].

The observed accuracies of the Square Eyes model are generally lower than what has been reported in studies evaluating the use of similar object recognition models for identifying technologies in wearable camera images. One prior study based on 1029 images from an adult sample reported detection accuracies ranging between 81% and 84% for phones, laptops, and televisions [25]. Another adult-based study found a mean accuracy of 85% for identifying phones, laptops, tablets, desktops, computers, and television screens based on 215 images [26]. A study based on 1800 images from cameras worn by children aged 3 to 5 years observed a mean accuracy of 93% for identifying computers, smartphones, and televisions [19].

Several methodological factors may explain the lower accuracy of the Square Eyes model found in the present study. First, we evaluated the accuracy of the Square Eyes model based on an independent set of images. In contrast, in all the previous studies, model performance was evaluated based on a subset of images drawn from a sample used for training the models. While this approach is commonly used for model evaluation, it is likely that performance estimates were increased due to similarity between the training and testing datasets. Moreover, the subsets were substantially smaller (n=215‐1800 images) than the number of images used for evaluating the model in the current study (n=221,226). The smaller samples may have resulted in higher performance estimates as a result of issues such as overfitting [42]. The studies by Hou et al [19] and Li et al [25] excluded images with multiple screens in the same image, whereas multiple devices were present on the images used in the current study. We elected to keep multiple screens as this is more likely what will be found in free-living environments. However, the presence of multiple screens will increase the challenge of object recognition for the model, particularly when screens partially occlude one another. The studies by Li et al [25] and Su et al [26] were based on adult samples. It can be expected that images from cameras worn by adults are less influenced by motion blur than those from cameras worn by children who typically have more rapid movements than adults. Moreover, the lower height of children may influence the ability of wearable cameras to capture stationary technologies, such as televisions, which may be placed above the children’s eye line. Finally, low-quality images affected by motion blur were excluded prior to analysis in the child-based study by Hou et al [19], which likely improved model performance. Taken together, differences in training–testing dataset overlap, sample size, image complexity, and image quality likely contribute to the lower agreement observed in the present study.

The model’s N-back algorithm effectively excluded 32.8% (n=72,680) of the images from requiring manual annotation, indicating a substantial reduction in workload. Importantly, of the images not flagged for human review, 90.2% (n=63,536) did not contain a human-coded technology, suggesting that most images were correctly identified as not requiring human review. Across the images not flagged for human review but containing a human-coded technology (n=7144, 3.2%), these images primarily contained smaller devices (eg, smartwatches and smaller gaming devices). Thus, while the algorithm reduced human coding burden by approximately one-third, some smaller technologies were more frequently missed. Nevertheless, our findings suggest that the Square Eyes model could be used as an efficient tool for reducing human coding burden.

To our knowledge, no prior study has reported an evaluation of the ability of object recognition models to reduce overall researcher burden for processing wearable camera images. Decreasing reliance on intensive manual coding may contribute to greater research workforce sustainability by reducing a highly time-consuming task. Thus, we encourage future research to explore the use of similar hybrid approaches. Such approaches could facilitate the use of object recognition models to identify images likely to contain technologies, followed by human coders to confirm the presence of technologies. Moreover, human coding would remain essential for obtaining information on the content and context of technology use. Additionally, engaging with parents and children in study designs may further facilitate the inclusion of larger and more diverse study populations by supporting feasibility and acceptability of wearable camera research.

An additional consideration for future research relates to the rapid evolution of wearable camera devices, which include higher quality images and continuous video recording. The Square Eyes model was trained and developed for still-image object recognition and thus, we encourage future work to test the feasibility of model development for video-based analysis. This would likely involve further development of the current YOLOv8-based framework to incorporate temporal information across continuous frames and to optimize performance for small objects such as smartphones and handheld devices. Future advances in AI are likely to enhance model performance and should be evaluated as advanced models become available, building on the knowledge from this study.

The primary purpose of obscuring faces using a blurring technique is to ensure third-party privacy [29]. However, a potential consequence may be reduced model performance [30]. In the current study, we observed that the use of a square-bordered blur was sometimes mistakenly identified as a technology by the Square Eyes model. In contrast, this did not appear to be the case when using either a nonbordered square or nonbordered circular face blur. Thus, the performance of the Square Eyes model for detecting technologies reported above may have been reduced by the facial blur technique applied to the images in this study. To our knowledge, the potential influence of facial blur techniques on the performance of object recognition models used for technology detection in wearable camera images has not been previously examined. However, our findings contradict those of a study assessing the effect of face blurring on the performance of a deep-learning model used for human pose estimation [43]. The authors concluded that face blurring had minimal impact on model performance when the model was trained using face-blurred images. However, this model was designed for whole-body posture recognition, not for the detection of technologies. As ensuring privacy is important in wearable camera research, we encourage future work to evaluate how different face-blurring techniques may influence the performance of machine learning models used for detecting technologies. This may involve training models directly on face-blurred images and comparing different blurring features (eg, circular vs square) across diverse real-world conditions to support the development of standardized ethical pipelines for analyzing wearable camera images.

A strength of this study was the large data sample of images captured by wearable cameras worn by children aged 3 to 14 years. Few studies have evaluated the use of wearable cameras and machine learning models for technology identification among children, and thus our study substantially adds to an underinvestigated research area. Moreover, images captured among children are likely to have lower visual quality due to increased motion blur compared with images captured among adults. Thus, evaluating the Square Eyes model using a visually challenging dataset provided a rigorous test of model robustness. This was further supported by using images that were not part of the dataset used for training the model. Specifically, models often demonstrate reduced performance when applied to datasets from other samples and settings. Our results, therefore, support the external validity of the Square Eyes model. Nevertheless, some limitations to the dataset should be acknowledged. This study was based on images obtained in a laboratory, which limits generalizability to free-living settings. Specifically, in free-living settings, the quality of images obtained with wearable cameras is likely influenced by substantial variability in lighting, postures, and movements, and thus image quality. Moreover, patterns of technology use in a laboratory setting may not reflect real-world settings where children are likely to engage with multiple technologies across varied physical environments and social settings. Such factors are likely to reduce detection performance of object recognition models. Thus, the findings should be interpreted as a proof-of-concept under controlled conditions rather than estimates of real-world applicability. Given the high burden of human coding, only one coder was used. Although coder training and initial review procedures were implemented to ensure high-quality coding, reliability was not assessed across a larger subset of the dataset. Therefore, we cannot assume that the human-coded reference was free from misclassifications. The potential influence of the facial blurring technique on model performance was evaluated using a small sample of images from a single participant. This analysis was explorative, and it was beyond the scope of the study to estimate performance differences between facial blur methods for the whole sample. Therefore, given the limited sample size and lack of systematic comparison, these findings regarding the impact of facial blur methods should be interpreted with caution.

We found that the Square Eyes model could detect technologies with overall satisfying accuracy and successfully flag images that required further review by humans. These findings suggest that the model can be used as an efficient screening tool to reduce human coding burden for researchers processing wearable camera images of children. However, the model could be improved for accurately detecting smaller devices, such as smartphones. Moreover, we found that the method used for facial blurring might influence model performance. We encourage further research to evaluate the use of similar hybrid approaches that use machine learning models to reduce human coding burden for both technology identification and other important aspects of technology use, such as context and content.