Training and Profiling a Pediatric Facial Expression Classifier for Children on Mobile Devices: Machine Learning Study.

BACKGROUND: Implementing automated facial expression recognition on mobile devices could provide an accessible diagnostic and therapeutic tool for those who struggle to recognize facial expressions, including children with developmental behavioral conditions such as autism. Despite recent advances in facial expression classifiers for children, existing models are too computationally expensive for smartphone use. OBJECTIVE: We explored several state-of-the-art facial expression classifiers designed for mobile devices, used posttraining optimization techniques for both classification performance and efficiency on a Motorola Moto G6 phone, evaluated the importance of training our classifiers on children versus adults, and evaluated the models' performance against different ethnic groups. METHODS: We collected images from 12 public data sets and used video frames crowdsourced from the GuessWhat app to train our classifiers. All images were annotated for 7 expressions: neutral, fear, happiness, sadness, surprise, anger, and disgust. We tested 3 copies for each of 5 different convolutional neural network architectures: MobileNetV3-Small 1.0x, MobileNetV2 1.0x, EfficientNetB0, MobileNetV3-Large 1.0x, and NASNetMobile. We trained the first copy on images of children, second copy on images of adults, and third copy on all data sets. We evaluated each model against the entire Child Affective Facial Expression (CAFE) set and by ethnicity. We performed weight pruning, weight clustering, and quantize-aware training when possible and profiled each model's performance on the Moto G6. RESULTS: Our best model, a MobileNetV3-Large network pretrained on ImageNet, achieved 65.78% accuracy and 65.31% F1-score on the CAFE and a 90-millisecond inference latency on a Moto G6 phone when trained on all data. This accuracy is only 1.12% lower than the current state of the art for CAFE, a model with 13.91x more parameters that was unable to run on the Moto G6 due to its size, even when fully optimized. When trained solely on children, this model achieved 60.57% accuracy and 60.29% F1-score. When trained only on adults, the model received 53.36% accuracy and 53.10% F1-score. Although the MobileNetV3-Large trained on all data sets achieved nearly a 60% F1-score across all ethnicities, the data sets for South Asian and African American children achieved lower accuracy (as much as 11.56%) and F1-score (as much as 11.25%) than other groups. CONCLUSIONS: With specialized design and optimization techniques, facial expression classifiers can become lightweight enough to run on mobile devices and achieve state-of-the-art performance. There is potentially a "data shift" phenomenon between facial expressions of children compared with adults; our classifiers performed much better when trained on children. Certain underrepresented ethnic groups (e.g., South Asian and African American) also perform significantly worse than groups such as European Caucasian despite similar data quality. Our models can be integrated into mobile health therapies to help diagnose autism spectrum disorder and provide targeted therapeutic treatment to children.

Crowd annotations can approximate clinical autism impressions from short home videos with privacy protections.

Artificial Intelligence (A.I.) solutions are increasingly considered for telemedicine. For these methods to serve children and their families in home settings, it is crucial to ensure the privacy of the child and parent or caregiver. To address this challenge, we explore the potential for global image transformations to provide privacy while preserving the quality of behavioral annotations. Crowd workers have previously been shown to reliably annotate behavioral features in unstructured home videos, allowing machine learning classifiers to detect autism using the annotations as input. We evaluate this method with videos altered via pixelation, dense optical flow, and Gaussian blurring. On a balanced test set of 30 videos of children with autism and 30 neurotypical controls, we find that the visual privacy alterations do not drastically alter any individual behavioral annotation at the item level. The AUROC on the evaluation set was 90.0% ±7.5% for unaltered videos, 85.0% ±9.0% for pixelation, 85.0% ±9.0% for optical flow, and 83.3% ±9.3% for blurring, demonstrating that an aggregation of small changes across behavioral questions can collectively result in increased misdiagnosis rates. We also compare crowd answers against clinicians who provided the same annotations for the same videos as crowd workers, and we find that clinicians have higher sensitivity in their recognition of autism-related symptoms. We also find that there is a linear correlation (r = 0.75, p < 0.0001) between the mean Clinical Global Impression (CGI) score provided by professional clinicians and the corresponding score emitted by a previously validated autism classifier with crowd inputs, indicating that the classifier's output probability is a reliable estimate of the clinical impression of autism. A significant correlation is maintained with privacy alterations, indicating that crowd annotations can approximate clinician-provided autism impression from home videos in a privacy-preserved manner.

Causal Modeling to Mitigate Selection Bias and Unmeasured Confounding in Internet-Based Epidemiology of COVID-19: Model Development and Validation.

BACKGROUND: Selection bias and unmeasured confounding are fundamental problems in epidemiology that threaten study internal and external validity. These phenomena are particularly dangerous in internet-based public health surveillance, where traditional mitigation and adjustment methods are inapplicable, unavailable, or out of date. Recent theoretical advances in causal modeling can mitigate these threats, but these innovations have not been widely deployed in the epidemiological community. OBJECTIVE: The purpose of our paper is to demonstrate the practical utility of causal modeling to both detect unmeasured confounding and selection bias and guide model selection to minimize bias. We implemented this approach in an applied epidemiological study of the COVID-19 cumulative infection rate in the New York City (NYC) spring 2020 epidemic. METHODS: We collected primary data from Qualtrics surveys of Amazon Mechanical Turk (MTurk) crowd workers residing in New Jersey and New York State across 2 sampling periods: April 11-14 and May 8-11, 2020. The surveys queried the subjects on household health status and demographic characteristics. We constructed a set of possible causal models of household infection and survey selection mechanisms and ranked them by compatibility with the collected survey data. The most compatible causal model was then used to estimate the cumulative infection rate in each survey period. RESULTS: There were 527 and 513 responses collected for the 2 periods, respectively. Response demographics were highly skewed toward a younger age in both survey periods. Despite the extremely strong relationship between age and COVID-19 symptoms, we recovered minimally biased estimates of the cumulative infection rate using only primary data and the most compatible causal model, with a relative bias of +3.8% and -1.9% from the reported cumulative infection rate for the first and second survey periods, respectively. CONCLUSIONS: We successfully recovered accurate estimates of the cumulative infection rate from an internet-based crowdsourced sample despite considerable selection bias and unmeasured confounding in the primary data. This implementation demonstrates how simple applications of structural causal modeling can be effectively used to determine falsifiable model conditions, detect selection bias and confounding factors, and minimize estimate bias through model selection in a novel epidemiological context. As the disease and social dynamics of COVID-19 continue to evolve, public health surveillance protocols must continue to adapt; the emergence of Omicron variants and shift to at-home testing as recent challenges. Rigorous and transparent methods to develop, deploy, and diagnosis adapted surveillance protocols will be critical to their success.

Improved Digital Therapy for Developmental Pediatrics Using Domain-Specific Artificial Intelligence: Machine Learning Study.

BACKGROUND: Automated emotion classification could aid those who struggle to recognize emotions, including children with developmental behavioral conditions such as autism. However, most computer vision emotion recognition models are trained on adult emotion and therefore underperform when applied to child faces. OBJECTIVE: We designed a strategy to gamify the collection and labeling of child emotion-enriched images to boost the performance of automatic child emotion recognition models to a level closer to what will be needed for digital health care approaches. METHODS: We leveraged our prototype therapeutic smartphone game, GuessWhat, which was designed in large part for children with developmental and behavioral conditions, to gamify the secure collection of video data of children expressing a variety of emotions prompted by the game. Independently, we created a secure web interface to gamify the human labeling effort, called HollywoodSquares, tailored for use by any qualified labeler. We gathered and labeled 2155 videos, 39,968 emotion frames, and 106,001 labels on all images. With this drastically expanded pediatric emotion-centric database (>30 times larger than existing public pediatric emotion data sets), we trained a convolutional neural network (CNN) computer vision classifier of happy, sad, surprised, fearful, angry, disgust, and neutral expressions evoked by children. RESULTS: The classifier achieved a 66.9% balanced accuracy and 67.4% F1-score on the entirety of the Child Affective Facial Expression (CAFE) as well as a 79.1% balanced accuracy and 78% F1-score on CAFE Subset A, a subset containing at least 60% human agreement on emotions labels. This performance is at least 10% higher than all previously developed classifiers evaluated against CAFE, the best of which reached a 56% balanced accuracy even when combining "anger" and "disgust" into a single class. CONCLUSIONS: This work validates that mobile games designed for pediatric therapies can generate high volumes of domain-relevant data sets to train state-of-the-art classifiers to perform tasks helpful to precision health efforts.

A Method for Localizing Non-Reference Sequences to the Human Genome.

As the last decade of human genomics research begins to bear the fruit of advancements in precision medicine, it is important to ensure that genomics' improvements in human health are distributed globally and equitably. An important step to ensuring health equity is to improve the human reference genome to capture global diversity by including a wide variety of alternative haplotypes, sequences that are not currently captured on the reference genome.We present a method that localizes 100 basepair (bp) long sequences extracted from short-read sequencing that can ultimately be used to identify what regions of the human genome non-reference sequences belong to.We extract reads that don't align to the reference genome, and compute the population's distribution of 100-mers found within the unmapped reads. We use genetic data from families to identify shared genetic material between siblings and match the distribution of unmapped k-mers to these inheritance patterns to determine the the most likely genomic region of a k-mer. We perform this localization with two highly interpretable methods of artificial intelligence: a computationally tractable Hidden Markov Model coupled to a Maximum Likelihood Estimator. Using a set of alternative haplotypes with known locations on the genome, we show that our algorithm is able to localize 96% of k-mers with over 90% accuracy and less than 1Mb median resolution. As the collection of sequenced human genomes grows larger and more diverse, we hope that this method can be used to improve the human reference genome, a critical step in addressing precision medicine's diversity crisis.

The Performance of Emotion Classifiers for Children With Parent-Reported Autism: Quantitative Feasibility Study.

BACKGROUND: Autism spectrum disorder (ASD) is a developmental disorder characterized by deficits in social communication and interaction, and restricted and repetitive behaviors and interests. The incidence of ASD has increased in recent years; it is now estimated that approximately 1 in 40 children in the United States are affected. Due in part to increasing prevalence, access to treatment has become constrained. Hope lies in mobile solutions that provide therapy through artificial intelligence (AI) approaches, including facial and emotion detection AI models developed by mainstream cloud providers, available directly to consumers. However, these solutions may not be sufficiently trained for use in pediatric populations. OBJECTIVE: Emotion classifiers available off-the-shelf to the general public through Microsoft, Amazon, Google, and Sighthound are well-suited to the pediatric population, and could be used for developing mobile therapies targeting aspects of social communication and interaction, perhaps accelerating innovation in this space. This study aimed to test these classifiers directly with image data from children with parent-reported ASD recruited through crowdsourcing. METHODS: We used a mobile game called Guess What? that challenges a child to act out a series of prompts displayed on the screen of the smartphone held on the forehead of his or her care provider. The game is intended to be a fun and engaging way for the child and parent to interact socially, for example, the parent attempting to guess what emotion the child is acting out (eg, surprised, scared, or disgusted). During a 90-second game session, as many as 50 prompts are shown while the child acts, and the video records the actions and expressions of the child. Due in part to the fun nature of the game, it is a viable way to remotely engage pediatric populations, including the autism population through crowdsourcing. We recruited 21 children with ASD to play the game and gathered 2602 emotive frames following their game sessions. These data were used to evaluate the accuracy and performance of four state-of-the-art facial emotion classifiers to develop an understanding of the feasibility of these platforms for pediatric research. RESULTS: All classifiers performed poorly for every evaluated emotion except happy. None of the classifiers correctly labeled over 60.18% (1566/2602) of the evaluated frames. Moreover, none of the classifiers correctly identified more than 11% (6/51) of the angry frames and 14% (10/69) of the disgust frames. CONCLUSIONS: The findings suggest that commercial emotion classifiers may be insufficiently trained for use in digital approaches to autism treatment and treatment tracking. Secure, privacy-preserving methods to increase labeled training data are needed to boost the models' performance before they can be used in AI-enabled approaches to social therapy of the kind that is common in autism treatments.

Addendum to the Acknowledgements: Validity of Online Screening for Autism: Crowdsourcing Study Comparing Paid and Unpaid Diagnostic Tasks.

[This corrects the article DOI: 10.2196/13668.].

Validity of Online Screening for Autism: Crowdsourcing Study Comparing Paid and Unpaid Diagnostic Tasks.

BACKGROUND: Obtaining a diagnosis of neuropsychiatric disorders such as autism requires long waiting times that can exceed a year and can be prohibitively expensive. Crowdsourcing approaches may provide a scalable alternative that can accelerate general access to care and permit underserved populations to obtain an accurate diagnosis. OBJECTIVE: We aimed to perform a series of studies to explore whether paid crowd workers on Amazon Mechanical Turk (AMT) and citizen crowd workers on a public website shared on social media can provide accurate online detection of autism, conducted via crowdsourced ratings of short home video clips. METHODS: Three online studies were performed: (1) a paid crowdsourcing task on AMT (N=54) where crowd workers were asked to classify 10 short video clips of children as "Autism" or "Not autism," (2) a more complex paid crowdsourcing task (N=27) with only those raters who correctly rated ≥8 of the 10 videos during the first study, and (3) a public unpaid study (N=115) identical to the first study. RESULTS: For Study 1, the mean score of the participants who completed all questions was 7.50/10 (SD 1.46). When only analyzing the workers who scored ≥8/10 (n=27/54), there was a weak negative correlation between the time spent rating the videos and the sensitivity (ρ=-0.44, P=.02). For Study 2, the mean score of the participants rating new videos was 6.76/10 (SD 0.59). The average deviation between the crowdsourced answers and gold standard ratings provided by two expert clinical research coordinators was 0.56, with an SD of 0.51 (maximum possible SD is 3). All paid crowd workers who scored 8/10 in Study 1 either expressed enjoyment in performing the task in Study 2 or provided no negative comments. For Study 3, the mean score of the participants who completed all questions was 6.67/10 (SD 1.61). There were weak correlations between age and score (r=0.22, P=.014), age and sensitivity (r=-0.19, P=.04), number of family members with autism and sensitivity (r=-0.195, P=.04), and number of family members with autism and precision (r=-0.203, P=.03). A two-tailed t test between the scores of the paid workers in Study 1 and the unpaid workers in Study 3 showed a significant difference (P<.001). CONCLUSIONS: Many paid crowd workers on AMT enjoyed answering screening questions from videos, suggesting higher intrinsic motivation to make quality assessments. Paid crowdsourcing provides promising screening assessments of pediatric autism with an average deviation <20% from professional gold standard raters, which is potentially a clinically informative estimate for parents. Parents of children with autism likely overfit their intuition to their own affected child. This work provides preliminary demographic data on raters who may have higher ability to recognize and measure features of autism across its wide range of phenotypic manifestations.

Detecting Developmental Delay and Autism Through Machine Learning Models Using Home Videos of Bangladeshi Children: Development and Validation Study.

BACKGROUND: Autism spectrum disorder (ASD) is currently diagnosed using qualitative methods that measure between 20-100 behaviors, can span multiple appointments with trained clinicians, and take several hours to complete. In our previous work, we demonstrated the efficacy of machine learning classifiers to accelerate the process by collecting home videos of US-based children, identifying a reduced subset of behavioral features that are scored by untrained raters using a machine learning classifier to determine children's "risk scores" for autism. We achieved an accuracy of 92% (95% CI 88%-97%) on US videos using a classifier built on five features. OBJECTIVE: Using videos of Bangladeshi children collected from Dhaka Shishu Children's Hospital, we aim to scale our pipeline to another culture and other developmental delays, including speech and language conditions. METHODS: Although our previously published and validated pipeline and set of classifiers perform reasonably well on Bangladeshi videos (75% accuracy, 95% CI 71%-78%), this work improves on that accuracy through the development and application of a powerful new technique for adaptive aggregation of crowdsourced labels. We enhance both the utility and performance of our model by building two classification layers: The first layer distinguishes between typical and atypical behavior, and the second layer distinguishes between ASD and non-ASD. In each of the layers, we use a unique rater weighting scheme to aggregate classification scores from different raters based on their expertise. We also determine Shapley values for the most important features in the classifier to understand how the classifiers' process aligns with clinical intuition. RESULTS: Using these techniques, we achieved an accuracy (area under the curve [AUC]) of 76% (SD 3%) and sensitivity of 76% (SD 4%) for identifying atypical children from among developmentally delayed children, and an accuracy (AUC) of 85% (SD 5%) and sensitivity of 76% (SD 6%) for identifying children with ASD from those predicted to have other developmental delays. CONCLUSIONS: These results show promise for using a mobile video-based and machine learning-directed approach for early and remote detection of autism in Bangladeshi children. This strategy could provide important resources for developmental health in developing countries with few clinical resources for diagnosis, helping children get access to care at an early age. Future research aimed at extending the application of this approach to identify a range of other conditions and determine the population-level burden of developmental disabilities and impairments will be of high value.

Outgroup Machine Learning Approach Identifies Single Nucleotide Variants in Noncoding DNA Associated with Autism Spectrum Disorder.

Autism spectrum disorder (ASD) is a heritable neurodevelopmental disorder affecting 1 in 59 children. While noncoding genetic variation has been shown to play a major role in many complex disorders, the contribution of these regions to ASD susceptibility remains unclear. Genetic analyses of ASD typically use unaffected family members as controls; however, we hypothesize that this method does not effectively elevate variant signal in the noncoding region due to family members having subclinical phenotypes arising from common genetic mechanisms. In this study, we use a separate, unrelated outgroup of individuals with progressive supranuclear palsy (PSP), a neurodegenerative condition with no known etiological overlap with ASD, as a control population. We use whole genome sequencing data from a large cohort of 2182 children with ASD and 379 controls with PSP, sequenced at the same facility with the same machines and variant calling pipeline, in order to investigate the role of noncoding variation in the ASD phenotype. We analyze seven major types of noncoding variants: microRNAs, human accelerated regions, hypersensitive sites, transcription factor binding sites, DNA repeat sequences, simple repeat sequences, and CpG islands. After identifying and removing batch effects between the two groups, we trained an ℓ1-regularized logistic regression classifier to predict ASD status from each set of variants. The classifier trained on simple repeat sequences performed well on a held-out test set (AUC-ROC = 0.960); this classifier was also able to differentiate ASD cases from controls when applied to a completely independent dataset (AUC-ROC = 0.960). This suggests that variation in simple repeat regions is predictive of the ASD phenotype and may contribute to ASD risk. Our results show the importance of the noncoding region and the utility of independent control groups in effectively linking genetic variation to disease phenotype for complex disorders.

Mobile detection of autism through machine learning on home video: A development and prospective validation study.

BACKGROUND: The standard approaches to diagnosing autism spectrum disorder (ASD) evaluate between 20 and 100 behaviors and take several hours to complete. This has in part contributed to long wait times for a diagnosis and subsequent delays in access to therapy. We hypothesize that the use of machine learning analysis on home video can speed the diagnosis without compromising accuracy. We have analyzed item-level records from 2 standard diagnostic instruments to construct machine learning classifiers optimized for sparsity, interpretability, and accuracy. In the present study, we prospectively test whether the features from these optimized models can be extracted by blinded nonexpert raters from 3-minute home videos of children with and without ASD to arrive at a rapid and accurate machine learning autism classification. METHODS AND FINDINGS: We created a mobile web portal for video raters to assess 30 behavioral features (e.g., eye contact, social smile) that are used by 8 independent machine learning models for identifying ASD, each with >94% accuracy in cross-validation testing and subsequent independent validation from previous work. We then collected 116 short home videos of children with autism (mean age = 4 years 10 months, SD = 2 years 3 months) and 46 videos of typically developing children (mean age = 2 years 11 months, SD = 1 year 2 months). Three raters blind to the diagnosis independently measured each of the 30 features from the 8 models, with a median time to completion of 4 minutes. Although several models (consisting of alternating decision trees, support vector machine [SVM], logistic regression (LR), radial kernel, and linear SVM) performed well, a sparse 5-feature LR classifier (LR5) yielded the highest accuracy (area under the curve [AUC]: 92% [95% CI 88%-97%]) across all ages tested. We used a prospectively collected independent validation set of 66 videos (33 ASD and 33 non-ASD) and 3 independent rater measurements to validate the outcome, achieving lower but comparable accuracy (AUC: 89% [95% CI 81%-95%]). Finally, we applied LR to the 162-video-feature matrix to construct an 8-feature model, which achieved 0.93 AUC (95% CI 0.90-0.97) on the held-out test set and 0.86 on the validation set of 66 videos. Validation on children with an existing diagnosis limited the ability to generalize the performance to undiagnosed populations. CONCLUSIONS: These results support the hypothesis that feature tagging of home videos for machine learning classification of autism can yield accurate outcomes in short time frames, using mobile devices. Further work will be needed to confirm that this approach can accelerate autism diagnosis at scale.

Coalitional game theory as a promising approach to identify candidate autism genes.

Despite mounting evidence for the strong role of genetics in the phenotypic manifestation of Autism Spectrum Disorder (ASD), the specific genes responsible for the variable forms of ASD remain undefined. ASD may be best explained by a combinatorial genetic model with varying epistatic interactions across many small effect mutations. Coalitional or cooperative game theory is a technique that studies the combined effects of groups of players, known as coalitions, seeking to identify players who tend to improve the performance--the relationship to a specific disease phenotype--of any coalition they join. This method has been previously shown to boost biologically informative signal in gene expression data but to-date has not been applied to the search for cooperative mutations among putative ASD genes. We describe our approach to highlight genes relevant to ASD using coalitional game theory on alteration data of 1,965 fully sequenced genomes from 756 multiplex families. Alterations were encoded into binary matrices for ASD (case) and unaffected (control) samples, indicating likely gene-disrupting, inherited mutations in altered genes. To determine individual gene contributions given an ASD phenotype, a "player" metric, referred to as the Shapley value, was calculated for each gene in the case and control cohorts. Sixty seven genes were found to have significantly elevated player scores and likely represent significant contributors to the genetic coordination underlying ASD. Using network and cross-study analysis, we found that these genes are involved in biological pathways known to be affected in the autism cases and that a subset directly interact with several genes known to have strong associations to autism. These findings suggest that coalitional game theory can be applied to large-scale genomic data to identify hidden yet influential players in complex polygenic disorders such as autism.

Genetic Networks of Complex Disorders: from a Novel Search Engine for PubMed Article Database.

Finding genetic risk factors of complex disorders may involve reviewing hundreds of genes or thousands of research articles iteratively, but few tools have been available to facilitate this procedure. In this work, we built a novel publication search engine that can identify target-disorder specific, genetics-oriented research articles and extract the genes with significant results. Preliminary test results showed that the output of this engine has better coverage in terms of genes or publications, than other existing applications. We consider it as an essential tool for understanding genetic networks of complex disorders.