## Main

Visual impairment is one of the most important causes of long-term disability in children worldwide and has a detrimental impact on education and socioeconomic achievements1,2. Infancy and toddlerhood (early childhood) are critical periods for visual development3, during which early detection and prompt treatment of ocular pathology can prevent irreversible visual loss4,5. Young children are unable to complain of visual difficulties, and since they are unwilling or find it difficult to cooperate with standard vision tests (for example, optotype tests), age-appropriate tests such as grating acuity cards are commonly used to observe their reactions to visual stimuli6,7. However, evaluating the vision of young children using these tests requires highly trained operators, which greatly hinders their wider adoption, especially in low-income and middle-income countries with the highest prevalence of visual impairment but poor medical resources8. In addition, these tests, even when performed by experienced pediatric ophthalmologists, have been shown to have low repeatability in large-scale population screening studies9,10,11. Therefore, it is imperative to develop an easy-to-use and effective detection tool to enable the timely diagnosis of visual impairment in young children and prompt intervention.

Ocular abnormalities causing visual impairment in children often manifest with typical phenotypic features, such as leukocoria (white eye) in cataract12 and retinoblastoma13, eyelid drooping in congenital ptosis14, and a cloudy and enlarged cornea in congenital glaucoma15. In addition, previous studies have found that dynamic aberrant behavioral features such as abnormal ocular movement, fixation patterns or visual preference can also point toward an underlying ocular pathology in children16,17. These phenotypic manifestations are frequently seen in ocular diseases, such as amblyopia and strabismus, and they can provide valuable clues for diagnosing visual impairment in young children18,19,20. However, systematically recording and applying these features to real ophthalmic practice are still in their infancy due to the lack of practical and effective tools.

Given the rapid development of mobile health (mHealth) and artificial intelligence (AI) algorithms in identifying or monitoring disease states21,22, the use of mobile devices, such as smartphones, to record and analyze phenotypic features to help identify visual impairment in young children presents great opportunities. However, developing such a system for large-scale ophthalmic application is hindered by three main challenges: (1) collecting phenotypic data that reliably reflect the visual status of the children in complex environments, (2) generalizing the system for large-scale applications and (3) providing evidence of its feasibility. The major bottleneck that impedes the widespread adoption of many medical AI systems is the limited feasibility and reliability when applied to settings with various data distributions in the real world23,24. A lack of cooperation is very common in pediatric ophthalmic practice, with constant head movement during examinations introducing test noise that poses several challenges to the stability of the system25. For the nascent technology of mHealth, rigorous evidence of clinical application is necessary but generally lacking21. These major difficulties explain the current lack of an effective and practical tool for detecting visual impairment in young children.

In this prospective, multicenter, observational study, we developed and validated a smartphone-based system, the Apollo Infant Sight (AIS), to identify visual impairment in young children in real-world settings. AIS was designed to induce a steady gaze in children by using cartoon-like video stimuli and collect videos that capture phenotypic features (facial appearance and ocular movements) for further analysis using deep learning (DL) models with robust quality control design against test noises. We collected more than 25,000,000 frames of videos from 3,652 children using AIS for DL model training and testing. We evaluated the system for detecting visual impairment caused by any of 16 ophthalmic disorders in five clinics at different institutions. Furthermore, we validated this system under different conditions with various test noise levels or ambient interference presented in real-world settings. We also evaluated AIS used by untrained parents or caregivers at home to test its wider applicability. This preliminary study indicates that AIS shows potential for early detection of visual impairment in young children in both clinical and community settings.

## Results

### Overview of the study

We conducted this prospective, multicenter and observational study (identifier: NCT04237350) in three stages from 14 January 2020 to 30 January 2022 and collected a total of 3,865 videos with 25,972,800 frames of images from 3,652 Chinese children (aged ≤48 months) to develop and validate the AIS system in clinical and at-home settings (Fig. 1). The AIS system was developed and comprehensively tested (internal validation and reliability analyses under different testing conditions) at the clinic of Zhongshan Ophthalmic Center (ZOC) in the first stage, and was further tested in four other centers (external validation) and community settings (at-home implementation) in the second and third stages, respectively.

### Development of the mHealth AIS system

We developed AIS for detecting visual impairment in young children tailored to the present study (Fig. 1a and Supplementary Video 1). A child-friendly app was designed to attract children to maintain their gaze using cartoon-like stimuli (Extended Data Fig. 1). The inbuilt front camera of the smartphone recorded 3.5-min videos that captured phenotypic features of the facial appearance and ocular movements during gazing. In this process, the mHealth app interactively guided users (healthcare professionals, volunteers, parents and caregivers) to familiarize themselves with the system and complete standardized preparations, including choosing and maintaining a suitable testing setting (Extended Data Fig. 2). After data collection was completed, DL models were applied to analyze the collected features and identify visually impaired children. To ensure the system’s performance in chaotic settings (environments with various interference factors or biases that can impact the system’s performance), a series of algorithm-based quality checking operations, including face detection (max-margin object detection (MMOD) convolutional neural network (CNN)); facial key point localization (ensemble of regression trees); and crying, occlusion and interference factor detections (the EfficientNet-B2 backbone shown in Extended Data Fig. 3a,b), was first automatically performed by a quality control module to extract consecutive frames of high quality from the original video as short clips. Facial areas were cropped out to further eliminate environmental interference before the qualified clips were sent to a DL-based detection model for identifying visually impaired children and a diagnostic model for discriminating multiple ocular disorders (the EfficientNet-B4 backbone shown in Extended Data Fig. 3c). The final results were returned to the mHealth app to alert users to promptly refer children at high risk of visual impairment to experienced pediatric ophthalmologists for timely diagnosis and intervention.

We first developed the data quality control module. Two facial detection and key point localization models were pretrained on publicly available datasets and adopted from an open-source library26. Additionally, we developed three CNNs for crying, interference and occlusion detection using images sampled from raw videos collected at the ZOC clinic (Extended Data Fig. 3d and Supplementary Table 1). Then, we trained and validated the detection/diagnostic models on the development dataset collected by trained volunteers using iPhone-7/8 smartphones at the clinic of ZOC (Extended Data Fig. 3e). A total of 2,632 raw videos from 2,632 children were collected, and after automatic quality control, videos of 2,344 children (89.1%) were reserved as the development dataset (Fig. 1b), including 871 (37.2%) for children in the ‘nonimpairment’ group, 861 (36.7%) in the ‘mild impairment’ group and 612 (26.1%) in the ‘severe impairment’ group. Detailed information on the qualified dataset is provided in Table 1. Before model training, the development dataset was randomly split into training, tuning and validation sets stratified on sex, age and the ophthalmic condition (Supplementary Table 2). The videos utilized for quality control module development were excluded from the detection/diagnostic model validation.

### Performance of the detection model in real clinical settings with trained volunteers

The detection model was trained to discriminate visually impaired children from nonimpaired children based on the high-quality clips extracted from the phenotypic videos. At the clip level, the detection model achieved an area under the receiver operating curve (AUC) of 0.925 (95% confidence interval (95% CI), 0.914–0.936) in the internal validation (Extended Data Fig. 4a). Furthermore, we evaluated the performance of the detection model via an independent external validation performed by trained volunteers using iPhone-7/iPhone-8 smartphones at the routine clinics of four other centers. In this stage, quality checking was embedded in the data acquisition process, and the quality control module automatically reminded volunteers to recollect data when the videos were of low quality (Fig. 1b). Qualified videos for 298 children undergoing ophthalmic examinations were utilized for final validation, including 188 (63.1%) nonimpaired children, 67 (22.5%) mildly impaired children and 43 (14.4%) severely impaired children (Table 1). At the clip level, the detection model achieved an AUC of 0.814 (95% CI, 0.790–0.838) in the external validation (Extended Data Fig. 4b).

The performance of the detection model to identify visually impaired children was evaluated by averaging the clip-level predictions. Figure 2a shows distinguished clip predicted probability patterns for children with various visual conditions. At the child level, the detection model achieved an AUC of 0.940 (95% CI, 0.920–0.959), an accuracy of 86.5% (95% CI, 83.4%–89.0%), a sensitivity of 84.1% (95% CI, 80.2%–87.4%) and a specificity of 91.9% (95% CI, 86.9%–95.1%) in the internal validation (Fig. 2b and Supplementary Table 3). It achieved a child-level AUC of 0.843 (95% CI, 0.794–0.893), an accuracy of 82.6% (95% CI, 77.8%–86.4%), a sensitivity of 80.9% (95% CI, 72.6%–87.2%) and a specificity of 83.5% (95% CI, 77.6%–88.1%) in the external validation (Fig. 2c and Supplementary Table 3).

Furthermore, we investigated whether our system could identify visual impairment with any of 16 common ophthalmic disorders at the child level (Table 2 and Supplementary Table 4). For different ophthalmic disorders, the predicted probabilities of the detection model were all significantly higher than those for nonimpairment (Fig. 2d). AIS achieved AUCs of over 0.800 in 15 of 16 binary classification tasks to distinguish visual impairment with various causes from nonimpairment (Fig. 2e,f and Supplementary Table 5), except for limbal dermoid with an AUC of 0.747 (95% CI, 0.646–0.849). Even for diseases not present in the training set, our system showed effective discriminative capabilities, revealing wider extendibility and generalizability to other conditions (Fig. 2f). In addition, we initially recruited children with aphakia (including iatrogenic aphakia cases with common features of visual impairment, accounting for 10.2% of the visually impaired participants enrolled) to increase diversity of training samples for the robustness of the system. Therefore, to evaluate the performance of AIS in the natural population without iatrogenic cases or cases with medical interventions, the children with aphakia were removed from the validation datasets for further analysis and AIS remained reliable (Supplementary Table 6). These results indicate the advanced classification of AIS to detect common causes of visual impairment in young children.

Additionally, the performance of AIS in discriminating mild or severe impairment from nonimpairment was assessed at the child level (Fig. 2g–j and Supplementary Table 3). Significantly lower predicted probabilities of AIS were obtained for the nonimpaired group than for the mild or severe impairment groups. For discriminating mild impairment from nonimpairment, an AUC of 0.936 (95% CI, 0.912–0.960) and an AUC of 0.833 (95% CI, 0.774–0.892) were obtained for the internal validation and the external validation, respectively. For discriminating severe impairment from nonimpairment, an AUC of 0.944 (95% CI, 0.919–0.969) and an AUC of 0.859 (95% CI, 0.779–0.939) were obtained for the internal validation and the external validation, respectively.

To further evaluate the performance of AIS when applied to a population with a rare-case prevalence of visual impairment, we conducted a ‘finding a needle in a haystack’ test based on the internal validation dataset, with the simulated prevalences ranging from 0.1% to 9%. AIS successfully identified visually impaired children at different simulated prevalences, with AUCs stabilized around 0.940 (Supplementary Table 7).

### Performance of the detection model in at-home settings with untrained parents or caregivers

After validation in real clinical settings, we further implemented a more challenging application in at-home settings by parents or caregivers using their smartphones according to the system’s instructions (Fig. 1b). Of the 125 children recruited online from the Guangdong area, 122 children (97.6%) successfully completed qualified video collection, among whom 120 children undergoing ophthalmic examinations were enrolled. Other detailed information on the qualified data is summarized in Table 1. Given the great difference in data distributions for the home environments compared with the clinics, we fine-tuned the detection model using qualified videos from 32 children and then tested it by the subsequently collected validation set from another 88 children. On the validation set, 31 (35.2%) children were classified as nonimpaired and 57 (64.8%) children were classified as visually impaired. AIS achieved effective performance in the at-home implementation, with an AUC of 0.817 (95% CI, 0.756–0.881) for discriminating clips of visually impaired children from those of nonimpaired children (Extended Data Fig. 4c). At the child level, significantly lower predicted probability patterns were obtained for the nonimpaired children compared with mildly or severely impaired children (Fig. 2k). An AUC of 0.859 (95% CI, 0.767–0.950), an accuracy of 77.3% (95% CI, 67.5%–84.8%), a sensitivity of 77.2% (95% CI, 64.8%–86.2%) and a specificity of 77.4% (95% CI, 60.2%–88.6%) were attained for discriminating visual impairment from nonimpairment (Fig. 2l and Supplementary Table 3).

### Model visualization and explanation

We improved the interpretability of the detection model outputs by visualizing the model results in the internal validation set. After being projected into a two-dimensional space, the feature information extracted by the detection model exhibited distinct patterns between the visually impaired and nonimpaired clips (Fig. 3a). The attention patterns of the detection model presented by the average heat maps varied with the children’s visual functions and underlying ophthalmic disorders (Fig. 3b,c). Among the visually impaired children, the detection model focused more on the eyes and areas around the neck (Fig. 3c). In particular, for the clips extracted from visually impaired samples, those classified by human experts as having abnormal patterns were more likely to be predicted by our system as ‘visual impairment’ than those that were randomly extracted (Fig. 3d,e and Supplementary Table 8), indicating that the detection model might pay more attention to the morphological appearance or behavioral patterns of the eye and head regions, as we previously reported16.

Additionally, the clips misidentified by the system exhibited different clustering characteristics from the correctly recognized clips (true visually impaired or true nonimpaired clips), and more of the misidentified clips fell in the intermediate zone of the two clusters for the correctly recognized clips (Extended Data Fig. 5). Moreover, for the 20% of samples with the lowest predicted confidence values, the false identification rate was significantly higher than that of other groups and the system was equivocal. We aimed to find a solution when the system was unreliable by filtering out equivocal samples for manual review by ophthalmologists. The results show that the system performance was substantially improved with the increasing ratios for manual review. For instance, when selecting cases with confidence values less than 0.071 for manual review, accounting for 3% of the total cases, the sensitivity improved from 84.1% to 85.1% and the specificity improved from 91.9% to 93.1%; when selecting cases with confidence values less than 0.193 for manual review, accounting for 7% of the total cases, the sensitivity and specificity improved to 85.4% and 94.2%, respectively (Extended Data Fig. 6).

### Multiple-category classification of ophthalmic disorders

Considering that our system exhibited different attention patterns for visual impairment caused by specific ophthalmic disorders (Fig. 3c), we further developed a DL-based diagnostic model to differentiate ophthalmic disorders with characteristic attention patterns by the detection model (aphakia, congenital glaucoma, congenital ptosis and strabismus) and nonimpairment at the child level. In the diagnostic validation, our system effectively discriminated multiple ophthalmic disorders, achieving AUCs ranging from 0.918 for strabismus to 0.996 for congenital ptosis (Fig. 2m,n).

Stable performance is critical for real-world applications of mHealth and medical AI systems. Thus, we investigated the reliability of AIS at the clinic of ZOC. We first evaluated the influences of patient-related factors, including sex, age, laterality of the eye disorder and the apparency of the phenotypic features, on the performance of AIS. For the reliability stratified by sex, AIS achieved an AUC of 0.948 (95% CI, 0.921–0.971) in the boys group and an AUC of 0.931 (95% CI, 0.899–0.961) in the girls group (Fig. 4a). The predicted probability pattern of AIS remained stable under various age conditions (Fig. 4b), and the system achieved AUCs ranging from 0.909 for age group 4 to 0.954 for age group 3 (Fig. 4c). Additionally, AIS effectively identified visually impaired children with bilateral or unilateral eye disorders, with an AUC of 0.921 (95% CI, 0.891–0.952) in the unilateral group and an AUC of 0.952 (95% CI, 0.932–0.973) in the bilateral group (Fig. 4d). In addition, AIS achieved satisfactory performance with an AUC of 0.939 (95% CI, 0.918–0.960) in identifying hard-to-spot visually impaired children, who could have insidious phenotypic features and were easily neglected by community ophthalmologists (Supplementary Table 9).

Furthermore, we investigated the reliability of AIS under different data capture conditions, including testing distance, room illuminance, repeated testing and duration of the video recording. Similarly, AIS obtained stable detection performance among groups of different testing distances, with the lowest AUC of 0.935 (95% CI, 0.912–0.958) in the medium-distance group (Fig. 4e). Additionally, the AIS predicted probability pattern remained stable under different room illuminance conditions (Fig. 4f). Our system achieved the lowest AUC of 0.932 (95% CI, 0.901–0.963) in the medium illuminance group (Fig. 4g). In the retest analysis, the system remained robust with an intraclass correlation coefficient for predicted probabilities of 0.880 (95% CI, 0.843–0.908) and a Cohen’s κ for predicted categories of 0.837 (95% CI, 0.758–0.916) in another independent validation population recruited at ZOC (Fig. 4h and Table 1). In addition, as the duration of the video recording increased, AIS remained stable and achieved a maximal AUC of 0.931 (95% CI, 0.914–0.956) with a video duration longer than 30 s (Fig. 4i).

To further verify that the detecting results of our system were reliable and not solely mediated by baseline characteristics as confounders, we examined the odds ratios (ORs) of the AIS predictions adjusted for baseline characteristics at the child level. Even after controlling for potential baseline confounders, the AIS predictions had statistically significant adjusted ORs for detecting visual impairment in the internal and external validations and the at-home implementation (P < 0.001). The adjusted ORs ranged from 3.034 to 3.248 for tasks in the internal validation (Supplementary Table 10) and from 2.307 to 2.761 for tasks in the external validation (Supplementary Table 11). For the at-home implementation, the AIS predictions had a statistically significant adjusted OR of 2.496 (95% CI, 1.748–3.565, P = 4.815 × 10−7) for detecting visual impairment (Supplementary Table 12).

### Performance of the AIS across different smartphone platforms

To test the stability of our system in more complex settings, we performed adjustments to a dataset randomly sampled from the ZOC validation set with various blurring, brightness, color or Gaussian noise adjustment gradients to simulate the diversity of data quality collected by different smartphone cameras. Our system remained reliable and achieved AUCs of over 0.800 with blurring factors no more than 25 or brightness factors no more than 0.7, and it achieved AUCs of over 0.930 under different color adjustments and over 0.820 under various Gaussian noise adjustments (Extended Data Fig. 7).

Furthermore, an independent validation set from 389 children was collected at ZOC using the Huawei Honor-6 Plus and Redmi Note-7 smartphones with the Android operation system to evaluate the performance of AIS (Fig. 1b and Supplementary Table 13). After data quality checking, videos of 361 children were reserved (92.8%), including 87 (24.1%) children without visual impairment, 169 (46.8%) children with mild visual impairment and 105 (29.1%) children with severe visual impairment (Table 1). AIS showed significantly higher predicted probabilities for mild or severe impairment than for nonimpairment and achieved an AUC of 0.932 (95% CI, 0.902–0.963) for identifying visual impairment for the Android system at the child level (Extended Data Fig. 8).

## Discussion

With the high incidence of visual problems during the first few years of life, timely intervention to counter pathological visual deprivation mechanisms during this critical development period can prevent or minimize long-term visual loss3. However, early detection of visual impairment in young children is challenging due to the lack of accurate and easy-to-use tools applicable to both clinical and community environments. To overcome these challenges, we developed and validated a smartphone-based system (AIS) that provides a holistic and quantitative technique to identify visual impairment in young children in real-world settings. We comprehensively evaluated this system for 16 important causes of childhood vision loss. Our system achieved an AUC of 0.940 in the internal validation and an AUC of 0.843 in the external validation at the clinics of four different hospitals. Furthermore, our system proved reliable when used by parents or caregivers at home, achieving an AUC of 0.859 under these specific testing conditions.

One of the merits of AIS is in its applicability to different ocular diseases. Previous studies have utilized photographs to detect ocular and visual abnormalities in childhood27,28. These technologies, which focus on a single static image, are not suitable for large-scale applications due to their limited effectiveness and inability to handle multiple abnormalities with variable patterns. Given the complexity of ocular pathologies in children, the concept of accurately assessing a broad range of ocular conditions is attractive. In our prospective multicenter study, we analyzed more than 25,000,000 frames of information-rich phenotypic videos and accurately identified visual impairment caused by a wide range of sight-threatening eye diseases. Strikingly, AIS was able to detect most of the common causes of visual impairment in childhood, including anterior and posterior segment disorders, strabismus, ocular neoplasms, developmental abnormalities and ocular manifestations of systemic and genetic diseases29. Although cases like congenital cataracts tend to be easily diagnosed in specialist settings by experienced doctors, they are still frequently missed in the community, especially in areas with pediatric ophthalmic resource shortfall28. To apply AIS to various scenarios, we recruited cases of a broad range of eye disorders with variable severity in terms of their impact on vision. Our system was reasonably accurate in identifying mildly impaired children who could have subtle phenotypic features, making them easy to miss. Furthermore, our results indicate that AIS can be extended to diseases that have not been previously encountered in the training process, demonstrating its broader applicability.

The use of smartphones to detect visual impairment caused by extraocular diseases or systematic diseases is an important application in the future, but the feasibility remained to be further verified. Some systemic diseases, such as cardiovascular, hepatobiliary and renal diseases, can exhibit ocular manifestations that are recognizable by algorithms, which is also indicated by our findings in small samples30,31,32. Furthermore, disorders of neurological system can impact vision and cause cerebral visual impairment with pathology outside the eye, which is a common type of visual impairment in developed countries but lacking in this study33,34. Therefore, future work is needed to evaluate the merit of AIS in detecting visual impairment caused by a broad range of diseases, such as cerebral visual impairment, and in reducing the extraocular morbidity associated with systemic diseases in a larger population: for example, the cardiovascular complications linked with Marfan syndrome.

A major strength of AIS is its reliability in real-world practice. Although a large number of medical AI systems have been evaluated with high performance in the laboratory setting, only a few systems have demonstrated real-world medical feasibility23,25. Bias from training data and low stability of the model design greatly limit the generalizability of these AI systems. Previously, we evaluated the feasibility of identifying visual impairment in children by analyzing their phenotypic characteristics using DL algorithms16. For that study, the evaluation was conducted by experienced experts under a tightly controlled, standardized laboratory setting to strictly control for interference factors, which is not possible in routine ophthalmic practice. In this study, we prospectively collected a large amount of phenotypic data (facial features and ocular movements) to develop a DL system with a highly reliable design. Our results show that AIS exhibited high stability and prediction effectiveness under various testing conditions. Importantly, AIS remained effective in multicenter external validation and crucially, when rolled out in the community and used by parents or caregivers at home. When transferred to at-home settings, factors such as environmental interference, blurring, brightness, pixels of different cameras and the influence of untrained operators may impact the system’s performance. Therefore, we used a pilot dataset to fine-tune our system for its generalizability to various home environments and broader applications. AIS achieved an acceptable AUC of 0.859 in the subsequent implementation, which indicates that it can benefit from further model updating on larger-scale datasets for broader applications. Importantly, AIS kept stable in 88 different types of home environments after one round of fine-tuning, demonstrating its potential to be used generally in a variety of complex environments with no requirement of regular adaptations or fine-tuning in the future application.

Our findings demonstrate that sensory states, especially vision, can be derived from phenotypic video data recorded using consumer-grade smartphones. Two types of underlying features seemed to be captured by smartphones. First, changes in facial appearance caused by ocular pathologies can be directly recorded by mobile devices, especially those of the ocular surface or adnexa: for example, eyelid drooping in congenital ptosis. Second and more importantly, individuals may display aberrant behaviors to adapt to changes in their sensory modality, a process conserved from arthropods to mammals35,36 and confirmed in human children16. Our results show that the model can focus on behavioral features replicated in various eye diseases, such as abnormal ocular movement or alignment/fixation patterns. These common behavioral patterns may broaden the applicability of AIS to multiple ocular diseases, including posterior segment abnormalities that are more challenging to diagnose based on phenotypic video data.

A smartphone-based system to detect ocular pathology in children has obvious clinical implications. Early identification by parents or caregivers of ocular abnormalities facilitates timely referral to pediatric ophthalmologists and prompt intervention. AIS does not require professional medical equipment; smartphones and simple stabilization are sufficient. This low-barrier system is a promising tool for the timely testing of children in the community, which is a major advantage given the rapidly changing nature of the ocular pathology encountered in children. This could have a major impact by improving vision-related outcomes and even survival rates in cases such as retinoblastoma37,38. Furthermore, AIS is a promising tool to screen young children for ocular abnormalities remotely, which can reduce ophthalmologists’ exposure risk to infectious agents, as exemplified by the impact of the coronavirus disease 2019 (COVID-19) pandemic, in the so-called ‘new normal’ period39.

This study has several limitations. First, although we may miss the recruitment of some patients with conditions causing slight visual impairment in specialist clinical settings, our system was satisfactorily accurate in identifying mildly impaired children with subtle phenotypic features. Importantly, the versatile AIS system kept reliable performance to detect visually impaired children who were hard to spot even for community ophthalmologists, which sheds light on its significant application prospect of expanding our future work to the general population and groups of children with mild or early-stage ocular pathology. Second, to develop the quality control module and analyze the influencing factors, only a single video was collected for each child at ZOC, accounting for the relatively high rate of unsuccessful cases in this stage. However, our system allowed users to repeat video recordings until the qualified videos were acquired. As a result, the successful rate of identification greatly improved. Although a proportion of uncooperative children may not be appropriate for our tool, our AIS system has greatly lowered the minimal operating threshold for untrained users, indicating the potential for the general applications. Third, our cohorts recruited in clinical settings may not represent the real-world population. Although AIS effectively identified visually impaired children in the finding a needle in a haystack test with a prevalence simulated to a general population, a large-scale screening trial is needed in the future to validate the utility of the AIS system in the real-world applications. Fourth, AIS requires collecting facial information from children, which may pose a risk of privacy exposure. To avoid potential privacy risks, future techniques such as lightweight model backbones40 and model pruning41 could be applied to deploy the DL system in individual smartphones with no requirement for additional computing resources. In addition, digital fingerprint technology, such as blockchain42, can also be applied to monitor data usage and mitigate abuse effectively. Additionally, we developed a real-time three-dimensional facial reconstruction technology to irreversibly erase biometric attributes while retaining gaze patterns and eye movements43, which can be used in the future to safeguard children’s privacy when using AIS.

In conclusion, we developed and validated an innovative smartphone-based technique to detect visual impairment in young children affected with a broad range of eye diseases. Given the ubiquity of smartphones, AIS is a promising tool that can be applied in real-world settings for secondary prevention of visual loss in this particularly vulnerable age group.

## Methods

### Ethics approval

The predefined protocol of the clinical study was approved by the Institutional Review Board/Ethics Committee of ZOC and prospectively registered at ClinicalTrials.gov (identifier: NCT04237350), and it is shown in Supplementary Note. Consent was obtained from all individuals whose eyes or faces are shown in the figures or video for publication. Before data collection, informed written consent was obtained from at least one parent or guardian of each child. The investigators followed the requirements of the Declaration of Helsinki throughout the study.

### Study design and study population

This prospective, multicenter and observational study was conducted between 14 January 2020 and 30 January 2022 to recruit children for the development and validation of the mHealth system in three stages (Fig. 1b). Major eligibility criteria included an age of 48 months or younger and informed written consent obtained from at least one parent or guardian of each child. We did not include children having central nervous system diseases, mental illnesses or other known illnesses that could affect their behavioral patterns, in the absence of ocular manifestations. Children who could not cooperate to complete the ophthalmic examinations or the detection test using AIS were excluded. We also excluded children who had received ocular interventions and treatments in the month immediately preceding data collection.

In the first stage completed from 14 January 2020 to 15 September 2021, children were enrolled at the clinic of ZOC (Guangdong Province) to develop and comprehensively validate (internal validation and reliability analyses) the system. In the second stage, which occurred from 22 September 2021 to 19 November 2021, children were enrolled at the clinics of the Second Affiliated Hospital of Fujian Medical University (Fujian Province), Shenzhen Eye Hospital (Guangdong Province), Liuzhou Maternity and Child Healthcare Hospital (Guangxi Province) and Beijing Children’s Hospital of Capital Medical University (Beijing) to additionally evaluate the system (external validation). We selected these sites from three provinces across northern and southern China, representing the variations in clinical settings. In the first two stages, recruited children underwent ophthalmic examinations by clinical staff, and phenotypic videos were collected by trained volunteers using mHealth apps installed on iPhone-7 or iPhone-8 smartphones at each center. In the third stage conducted from 24 November 2021 to 30 January 2022, we advertised our study through the online platform of the Pediatric Department of ZOC and the social media of WeChat. We recruited children and their parents or caregivers online from the Guangdong area for at-home implementation. The investigators recruited the children following the same eligibility criteria as the previous two stages by collecting their basic information and medical history online. In addition, children who could not come to ZOC for an ophthalmic assessment or who had been included in other stages of this study were excluded. Untrained parents or caregivers recorded the phenotypic videos with their smartphones according to the instructions of the AIS app at home (Extended Data Figs. 1 and 2). The quality control module automatically reminded parents or caregivers to repeat data collection when the video recordings were unqualified. In this stage, all the children who completed successful video recordings underwent ophthalmic examinations at ZOC. A total of 3,652 children were finally enrolled, recording more than 25,000,000 frames of videos for development and validation of the system.

### Definition of visual impairment

Comprehensive functional and structural examinations were performed to stratify children’s visual conditions for developing and validating the DL-based AIS. For unified examination, a teller vision card (Stereo Optical Company) was utilized to measure children’s monocular visual acuity44. In addition, high-resolution slit lamp examinations, fundoscopy examinations and cycloplegic refraction were used to detect abnormalities in the eyes. Additional examinations, such as intraocular pressure, ultrasound, computerized tomography scans and genetic tests, were determined by experienced pediatric ophthalmologists when necessary.

According to the results of the abovementioned examinations and a referenced distribution of monocular visual acuity45, experienced pediatric ophthalmologists comprehensively stratified children’s visual conditions into three groups. Children with the best-corrected visual acuity (BCVA) of both eyes in the 95% referenced range with no abnormalities of structure or other examination results were assigned to the nonimpaired group. Children with the BCVA in the 99% referenced range in both eyes with abnormalities of structure or other examination results were assigned to the mildly impaired group. Children with the BCVA of at least one eye outside the 99% referenced range or worse than light perception with structural abnormalities or other examination results were assigned to the severely impaired group16. We recruited visually impaired children with primary diagnoses of the following 16 ocular disorders: aphakia, congenital cataract, congenital glaucoma, high ametropia, Peters’ anomaly, nystagmus, congenital ptosis, strabismus, persistent fetal vasculature, retinoblastoma, other fundus diseases, limbal dermoid, microphthalmia, pupillary membranes, systemic syndromes with ocular manifestations and other ocular conditions (Table 2 and Supplementary Table 4). A tiered panel consisting of two groups of experts assigned and confirmed the primary diagnosis as the most significant diagnostic label for each child. The first group of experts consisted of two pediatric ophthalmologists with over 10 years of experience in each recruiting ophthalmic center who separately provided the preliminary labeling information. If a consensus was not reached at this stage, a second group of more senior pediatric ophthalmologists with over 20 years of experience at ZOC verified the diagnostic labels as the ground truth. The diagnoses of children recruited online for the at-home implementation were made by experts at ZOC following the same criteria.

### Concept of the AIS system

The AIS system consisted of a smartphone app (available for iPhone and Android operating systems) for data collection and a DL back end for data analysis (Fig. 1a and Extended Data Fig. 1). To ensure the quality of data collected in real-world settings, AIS interactively instructed users to follow a standardized preparation sequence for data collection (Extended Data Fig. 2). Before data collection, a short demo video was displayed to instruct users on the standard operation and how to choose an appropriate environment to minimize testing biases (for example, room illuminance, background, testing distance and interference). Once the smartphone was firmly in place, a face-positioning frame was shown on the screen to help adjust the distance and position of the child in relation to the smartphone. After all preparations were completed properly, AIS played a cartoon-like video stimulus lasting approximately 3.5 min to attract children’s attention, and the inbuilt front camera recorded the children’s phenotypic features (ocular movements and facial appearance) in video format.

Then, the collected data were transferred to the DL-based back end, where the quality control module automatically performed quality checking on each frame first. To eliminate background interference, the children’s facial regions were then cropped out of consecutive frames of sufficient quality to form short video clips as inputs of the subsequent DL models for final decision-making (a detection model to distinguish visually impaired children from nonimpaired individuals and a diagnostic model to discriminate multiple ocular disorders). The DL models produced classification probabilities for short video clips, which were eventually merged into the video-level classification probability as the final outcome by averaging. The final results were returned to the mHealth app to alert users to promptly refer children at high risk of visual impairment to experienced pediatric ophthalmologists for further diagnosis and intervention.

### Deep quality control module

To ensure prediction reliability, we adopted a strict data quality control strategy to ensure that the input clips of the detection/diagnostic models satisfied certain quality criteria (Fig. 1a). First, for each frame, the child’s facial area was detected, and frames without successful face detection were rejected. If two or more faces were detected in a given frame, it suggested that the child’s parents or other persons were inside the scene, and such a frame was also rejected. The facial region detection algorithm was based on MMOD CNN46, which consisted of a series of convolutional layers for feature learning and max-margin operation during model training. In this study, the MMOD CNN face detector pretrained on publicly available datasets was adopted from the Dlib Python Library, which has been proven to be effective and robust in facial detection tasks26.

Second, a facial key point localization algorithm was applied to the detected facial area to extract the landmarks of facial regions, including the left eye, right eye, nose tip, chin and mouth corners, which served as the reference coordinates for the cropping of facial regions. The facial key point localization algorithm was realized based on a pretrained ensemble of regression trees, which was also provided by the Dlib Python Library47,48. We adopted a cascade of regressors to take the facial region of the frame as the input. The network was able to learn coarse-to-fine feature representations of the child’s face, especially details of the facial patterns. The output of this model was then fitted to the coordinates representing facial structures to generate 68 key target points. The coordinates of the key points then served as the reference for facial region cropping. All video data and image data processing were performed using the FFmpeg toolkit and OpenCV Python Library49.

Then, a combination of crying, interference and occlusion classification models based on EfficientNet-B2 networks (Extended Data Fig. 3a,b) was applied to each frame, which was trained based on the data collected at ZOC (Extended Data Fig. 3d and Supplementary Table 1)50. During model training and inference, the input frame was first rescaled to 384 × 384 resolution and then sent into the models for deep feature representation learning (Supplementary Table 14). Positive outputs by the models indicated that the child was crying, was interfered with or had its facial region blocked by objects such as toys or other persons’ hands, and the corresponding frames were also discarded. In practice, we fine-tuned the models pretrained on the ImageNet public dataset51.

Eventually, the remaining frames were considered high-quality candidates, and consecutive high-quality frames were selected to form short video clips. Each clip lasted at least 1.5 s and at most 5 s. The child’s facial region within each clip was then cropped out to serve as the final input of the subsequent detection/diagnostic models based on the facial key point coordinates to eliminate the interference of the background region. A qualified video should contain more than ten clips; otherwise, the video was treated as a low-quality sample and discarded.

### DL framework of the detection/diagnostic models

Two models with various clinical purposes were developed in this study: a detection model to detect visually impaired children from nonimpaired children and a five-category diagnostic model to discriminate specific ophthalmic disorders (aphakia, congenital glaucoma, congenital ptosis and strabismus) and nonimpairment. The backbone of each DL model was built on a deep convolutional network known as EfficientNet-B4 (Extended Data Fig. 3c and Supplementary Table 14)50. The models made predictions on the children’s cropped facial regions. Specifically, spatial cues of the input clips were learned by cascaded convolutional layers, while temporal cues were integrated by temporal average pooling layers, which was inspired by successful applications in gait recognition52. The temporal average pooling operator was given by $$\frac{1}{{{{{n}}}}}\mathop {\sum}\nolimits_{{{{{i}}}} = 1}^{{{{n}}}} {\mathop{x}\limits^{\rightharpoonup}}_{\!i}$$, where n was the number of frames in the input clip and $${\mathop{x}\limits^{\rightharpoonup}}_{\!i}$$ was the feature map of each frame output by the last convolutional layer of the network. Before training, all convolutional blocks were initialized by the parameters of the models pretrained on the ImageNet dataset51. At the inference stage, class scores given by the models were treated as the final clip-level probability outcomes. For the detection model, the output of the last classification layer, indicated by xi, was normalized to the range between 0.00 and 1.00 for each clip using the sigmoid function $${{{{p}}}}_{{{{i}}}} = \frac{1}{{1 + \exp ( - {{{{x}}}}_{{{{i}}}})}}$$, representing the final probability of the ith clip being classified as a visually impaired candidate. To train the detection model, the cost function was given by the classic binary cross-entropy loss $${{{{L}}}} = - \frac{{{{{1}}}}}{{{{{N}}}}}\mathop {\sum}\nolimits_{{{{{i}}}} = 1}^{{{{N}}}} {\left( {\widehat {{{{{y}}}}_{{{{i}}}}}\log \left( {{{{{p}}}}_{{{{i}}}}} \right) + \left( {1 - \widehat {{{{{y}}}}_{{{{i}}}}}} \right)\log \left( {1 - {{{{p}}}}_{{{{i}}}}} \right)} \right)}$$, where N was the number of clips within each batch, $$\widehat {{{{{y}}}}_{{{{i}}}}}$$ was the ground truth label of the ith clip and pi was the output classification probability of the model.

The diagnostic model was developed based on the same EfficientNet-B4 backbone as the detection model. The only difference was that the output of the diagnostic model was activated by a five-category softmax function that indicated the probability of each class: $${{{{p}}}}_{{{{k}}}} = \frac{{{{{{\mathrm{e}}}}}^{{{{{x}}}}_{{{{k}}}}}}}{{\mathop {\sum }\nolimits_{{{{{j}}}} = 1}^5 {{{{\mathrm{e}}}}}^{{{{{x}}}}_{{{{j}}}}}}}$$, where xk was the output of the last classification layer for the kth class. The cost function of the network was given by the stochastic cross-entropy loss $${{{{L}}}} = - \frac{1}{{{{{N}}}}}\mathop {\sum}\nolimits_{{{{{i}}}} = 1}^{{{{N}}}} {\mathop {\sum}\nolimits_{{{{{k}}}} = 1}^5 {\left( {{{{\hat{ y}}}}_{{{{k}}}}^{{{{i}}}}\log \left( {{{{{p}}}}_{{{{k}}}}^{{{{i}}}}} \right)} \right)} }$$, where N was the batch size and $${{{\hat{y}}}}_{{{{k}}}}^{{{{i}}}} \in \{ 0{{{{,}}}}1\}$$ was the binary Boolean variable of the ith input clip within each batch, indicating whether the kth class matched the ground truth label of the ith clip.

Child-level classification was based on clip-level predictions. A sliding window integrated with the quality control module was applied along the temporal dimension of the whole video to extract high-quality clips. Such clips then served as the candidate inputs for detection/diagnostic models. For the detection model, if the average score of the clips exceeded 0.50 within each video, the child was eventually classified as a visually impaired individual. For the diagnostic model, the category with the highest average probability was treated as the final prediction outcome.

### Model training and internal validation

We first developed the data quality control module using both publicly available datasets and the ZOC dataset (Supplementary Table 1). Then, we trained and validated the detection/diagnostic models with the ground truth of the visual conditions using the development dataset at ZOC. In this stage, data collection preceded the development of the quality control module, so raw videos without quality checking were collected. In total, raw videos from 2,632 children undergoing ophthalmic examinations were collected by trained volunteers using the mHealth apps installed on iPhone-7 or iPhone-8 smartphones. After initial quality checking by the quality control module, qualified videos of 2,344 (89.1%) children were reserved as the development dataset, which was randomly split into training, tuning and validation (internal validation) sets using a stratified sampling strategy according to sex, age and the category of ophthalmic disorder to train and internally validate the detection/diagnostic models (Fig. 1b, Extended Data Fig. 3e and Supplementary Table 2). The age distribution and the proportions of children with unilateral and bilateral severe visual impairment for different datasets are shown in Supplementary Tables 15 and 16, respectively. Internal validation refers to the assessment of the performance of the selected optimized model, after training and hyperparameter selection and tuning, on the independent datasets from the same settings as training datasets. The top-performing checkpoint was selected on the basis of accuracy on the tuning set. In particular, the videos utilized for quality control module development did not overlap with those in the detection/diagnostic model validation.

### Finding a needle in a haystack test

To estimate the performance of the AIS system in the general population with a rare-case prevalence of visual impairment, we simulated a gradient of prevalences ranging from 0.1% to 9% to conduct a finding a needle in a haystack test. For each simulated prevalence, we resampled 10,000 children based on the internal validation dataset in a bootstrap manner to test whether the AIS system could pick up the ‘needle’ (visually impaired children at the simulated prevalence) in the ‘haystack’ (10,000 resampled children) and repeated this process 100 times to estimate the 95% CIs.

### Data augmentation

To ensure better model capacity and reliability in complex environments, data augmentation was performed during model training using brightness and contrast adjustments, together with blurring techniques. Specifically, the brightness of the input frames was randomly adjusted by a factor of 0.40, and the contrast was randomly adjusted by a factor of 0.20. Blurring techniques included Gaussian blur, median blur and motion blur. The factor of all blurring techniques was set to five. Each input frame had a probability of 0.50 to perform data augmentation (Supplementary Table 17). All data augmentation processes were based on a publicly available Python library known as Albumentations53.

### Multicenter external validation

External validation refers to the assessment of the performance of the AI system using independent datasets, captured from different clinical settings. This is to ensure the generalizability of the system to different settings. Trained volunteers used mHealth apps installed on iPhone-7 or iPhone-8 smartphones to perform external validation in the ophthalmology clinics of the Second Affiliated Hospital of Fujian Medical University, Shenzhen Eye Hospital, Liuzhou Maternity and Child Healthcare Hospital and Beijing Children’s Hospital of Capital Medical University. In this stage, the quality control module automatically reminded volunteers to repeat data collection when the videos were of low quality. In total, 305 children were recruited and qualified videos for 301 children (98.7%) were successfully collected. Qualified videos for 298 children undergoing ophthalmic examinations were reserved for final validation of the detection model (see Fig. 1b and Table 1 for details of the participants and the dataset used for external validation).

### Implementation by untrained parents or caregivers at home

We further challenged our system in an application administered by untrained parents or caregivers with their smartphones in daily routines (Fig. 1b). Children (independent from the development and external validation participants) were recruited online, and their parents or caregivers autonomously used AIS at home according to the system’s instructions to collect qualified videos and perform tests without pretraining or controlling any biases before testing, such as brands and models of smartphones and the home environment. This process generated data with huge variations of distributions that had an extremely high requirement of generalizability and extensibility for the DL-based system. Thus, before final implementation, we performed a pilot study to collect a dataset for fine-tuning our system to chaotic home environments. To efficiently evaluate the performance of AIS for identifying visual impairment in at-home settings, a sufficient proportion of visually impaired children with various ocular diseases were recruited. Of the 125 children recruited, 122 children (97.6%) successfully completed the detection tests and collected qualified videos, among whom 120 children undergoing ophthalmic examinations were enrolled to fine-tune and evaluate the detection model. We fine-tuned the detection model using qualified videos from 32 children collected first and then tested it by the subsequently collected validation set from another 88 children. See Fig. 1b and Table 1 for more information on the fine-tuning and implementation.

### Reliability analyses and adjusted analyses

To test the stability and generalizability of AIS under various conditions, investigators conducted a batch of reliability analyses and adjusted analyses (Fig. 1b and Table 1).

### Reliability across different smartphone platforms

We performed adjustments at different blur, brightness, color or Gaussian noise adjustment gradients to a dataset (n = 200 children and n = 200 qualified videos) randomly sampled from the ZOC validation set to simulate the characteristics of data collected by various cameras and evaluate the reliability of AIS. Furthermore, we collected another dataset in an independent population of children at ZOC to assess the reliability of the AIS system across different operating systems. In total, raw videos from 389 children undergoing ophthalmic examinations were collected by trained volunteers using two Android smartphones, Redmi Note 7 and Huawei Hornor-6 plus. After initial quality checking, qualified videos of 361 (92.8%) children were reserved for testing. The technical specifications of the smartphones used in this study are summarized in Supplementary Table 13.

### Retest reliability analysis

We performed detection tests for each child twice by two volunteers at least 1 day apart on another independent population recruited at ZOC to evaluate the retest reliability. Raw videos from 213 children undergoing ophthalmic examinations were collected using iPhone-7 or iPhone-8 smartphones. Qualified videos of 187 (87.8%) children were reserved for retest analysis after initial quality checking (Fig. 1b and Table 1). An intraclass correlation coefficient was calculated for repeated predicted probabilities of the detection model, and a Cohen’s κ was calculated for repeated predicted categories to evaluate retest reliability.

### Hard-to-spot test

To investigate the influence of the apparency of the phenotypic features on the AIS system, a panel of 14 community ophthalmologists with 3–5 years of clinical experience identified ‘likely impaired’ children based on the phenotypic videos in the ZOC validation dataset. The true impaired and nonimpaired children were mixed at a ratio of 1:1 during identification. Each case was independently reviewed by three ophthalmologists. When no more than one ophthalmologist provided ‘likely impaired’ labels for one true impaired child, this child was classified as a hard-to-spot case with insidious phenotypic features rather than a relatively evident case. The performance of the AIS system for relatively evident/hard-to-spot cases was assessed.

### Other reliability analyses

We tested AIS under different room illuminance conditions. Photometers (TESTES-1330A; TES Electrical Electronic Corp.) were used to measure the mean room illuminance intensity before and after data collection. The following criteria were applied to estimate the distances between the children and the smartphones to assess the reliability of AIS in different testing distance groups. When most of the vertical lengths of a child’s head regions were less than one-third of the height of the smartphone screen at the frame level, the video was determined to be taken from a long distance. When most of the lengths were between one-third and one-half of the height of the screen, the video was judged to be taken from a medium distance, and when most of the lengths were larger than one-half of the height of the screen, the video was judged to be taken at a close distance. For each full-length video, subvideos with various durations were generated to serve as inputs to evaluate the influence of the duration of the video recording on the performance of AIS. We also evaluated the performances of AIS grouped by patient-related factors including sex, age and laterality of the eye disorder.

To further verify that the predictions of this system were not solely mediated by sample characteristics as confounders, we performed adjusted analyses to examine the ORs of the predictions of the system adjusted for sample characteristics leveraging logistic regression models.

### Detection model visualization and explanation

Two strategies were used to interpret and visualize the detection model: t-distributed stochastic neighbor embedding (t-SNE) and gradient-weighted class activation mapping (Grad-CAM)54,55,56. The former was used to visualize the high-dimensional activation status of the deep CNN at the clip level by projecting its feature vector into a two-dimensional space, and the latter was adopted to create a heat map showing the area within each frame of the clip that contributed most to the output class of the network. In practice, the feature vectors output by the temporal average pooling layer and flatten operation and the feature maps output by the last convolutional layer before the temporal average pooling operation were chosen to visualize the results generated by t-SNE and Grad-CAM, respectively. Specifically, 1,200 visually impaired clips and 1,200 nonimpaired clips were randomly selected from the ZOC validation set to perform t-SNE analysis. To generate average heat maps, we randomly sampled ten videos for each ophthalmic disorder from the internal validation dataset. Since each video had multiple clips, we ranked these clips according to the model predicted probabilities and selected the two clips with the highest probabilities. For each selected clip, we took 30 frames at equal intervals to generate the corresponding average heat map. In summary, we had a total of 600 heat maps for each type of disorder, and we summed and averaged these heat maps to obtain the typical heat map for a certain disease. A public machine learning Python library named Scikit-learn was used to generate two-dimensional coordinates of t-SNE results, and Grad-CAM analysis was performed based on an open-source GitHub code set57.

Additionally, we compared the model-predicted probabilities of three groups of clips (clips randomly sampled from videos of nonimpaired children, clips randomly sampled from videos of visually impaired children, and clips annotated by experts as having abnormal behavioral patterns from videos of visually impaired children) to investigate whether the detection model focused on specific behavioral patterns in children (Fig. 3d and Supplementary Table 8).

### Triage-driven approach to select equivocal cases for manual review

We assessed a triage strategy to find a solution when the system was likely unreliable by choosing equivocal cases for manual review in the internal validation set. An equivocal case referred to a child predicted by the AIS system with a low confidence value, given by |p − 0.50|, where p was the predicted probability for the child. Three ophthalmologists from ZOC with over 10 years of clinical experience vetted the phenotypic videos of the equivocal cases and the AIS predictions in a voting manner. Additional information, including baseline information and medical histories, was provided when necessary. An increasing ratio from 0 to 19% of equivocal cases with the lowest confidence values was chosen for manual review to evaluate this triage strategy.

### Statistical analysis

The primary outcomes were the AUCs of the detection/diagnostic models. The secondary outcomes included the accuracy, sensitivity and specificity of the models and the reliability of the detection model under various settings. The 95% CIs of the AUC, accuracy, sensitivity and specificity of the models were estimated. Specifically, the DeLong CIs of AUCs were calculated at the child level. To eliminate bias due to the association of multiple clips for the same child, the bootstrap CIs of the AUCs of the detection model were calculated at the clip level. One clip for each child was randomly taken to form a bootstrap sample, and this process was repeated 1,000 times. Wilson CIs were reported for other proportional metrics. Descriptive statistics, including means, s.d., numbers and percentages, were used. Mann–Whitney U-tests were used to compare means on continuous variables, and Fisher exact tests were used to compare distributions on categorical variables. A two-sided P value of <0.05 indicates statistical significance. All statistical analyses were performed in R Statistics (v.4.1.2) or Python Programs (v.3.9.7), and plots were created with the ggplot2 package (v.3.3.5) in R Statistics.

### Computational hardware

Hardware information for this study is shown as follows: graphics processing unit (GPU), Nvidia Titan RTX 24 GB memory × 4, Driver v.440.82, Cuda v.10.2; central processing unit (CPU), Intel(R) Xeon(R) CPU E5-2678 v.3 @ 2.50 GHz × 2, 48 threads; random access memory (RAM), Samsung 64 GB RAM × 8, configured speed 2,133 MHz.

### Use of human data

The ethical review of this study was approved by the Institutional Review Board/Ethics Committee of ZOC. The test was prospectively registered at ClinicalTrials.gov (identifier: NCT04237350).

### Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.