Marlena Duda

Searching for a minimal set of behaviors for autism detection through feature selection-based machine learning

Although the prevalence of autism spectrum disorder (ASD) has risen sharply in the last few years reaching 1 in 68, the average age of diagnosis in the United States remains close to 4—well past the developmental window when early intervention has the largest gains. This emphasizes the importance of developing accurate methods to detect risk faster than the current standards of care. In the present study, we used machine learning to evaluate one of the best and most widely used instruments for clinical assessment of ASD, the Autism Diagnostic Observation Schedule (ADOS) to test whether only a subset of behaviors can differentiate between children on and off the autism spectrum. ADOS relies on behavioral observation in a clinical setting and consists of four modules, with module 2 reserved for individuals with some vocabulary and module 3 for higher levels of cognitive functioning. We ran eight machine learning algorithms using stepwise backward feature selection on score sheets from modules 2 and 3 from 4540 individuals. We found that 9 of the 28 behaviors captured by items from module 2, and 12 of the 28 behaviors captured by module 3 are sufficient to detect ASD risk with 98.27% and 97.66% accuracy, respectively. A greater than 55% reduction in the number of behaviorals with negligible loss of accuracy across both modules suggests a role for computational and statistical methods to streamline ASD risk detection and screening. These results may help enable development of mobile and parent-directed methods for preliminary risk evaluation and/or clinical triage that reach a larger percentage of the population and help to lower the average age of detection and diagnosis.

Testing the accuracy of an observation-based classifier for rapid detection of autism risk.

Current approaches for diagnosing autism have high diagnostic validity but are time consuming and can contribute to delays in arriving at an official diagnosis. In a pilot study, we used machine learning to derive a classifier that represented a 72% reduction in length from the gold-standard Autism Diagnostic Observation Schedule-Generic (ADOS-G), while retaining >97% statistical accuracy. The pilot study focused on a relatively small sample of children with and without autism. The present study sought to further test the accuracy of the classifier (termed the observation-based classifier (OBC)) on an independent sample of 2616 children scored using ADOS from five data repositories and including both spectrum (n=2333) and non-spectrum (n=283) individuals. We tested OBC outcomes against the outcomes provided by the original and current ADOS algorithms, the best estimate clinical diagnosis, and the comparison score severity metric associated with ADOS-2. The OBC was significantly correlated with the ADOS-G (r=−0.814) and ADOS-2 (r=−0.779) and exhibited >97% sensitivity and >77% specificity in comparison to both ADOS algorithm scores. The correspondence to the best estimate clinical diagnosis was also high (accuracy=96.8%), with sensitivity of 97.1% and specificity of 83.3%. The correlation between the OBC score and the comparison score was significant (r=−0.628), suggesting that the OBC provides both a classification as well as a measure of severity of the phenotype. These results further demonstrate the accuracy of the OBC and suggest that reductions in the process of detecting and monitoring autism are possible.

The potential of accelerating early detection of autism through content analysis of YouTube videos.

Abstract Autism is on the rise, with 1 in 88 children receiving a diagnosis in the United States, yet the process for diagnosis remains cumbersome and time consuming. Research has shown that home videos of children can help increase the accuracy of diagnosis. However the use of videos in the diagnostic process is uncommon. In the present study, we assessed the feasibility of applying a gold-standard diagnostic instrument to brief and unstructured home videos and tested whether video analysis can enable more rapid detection of the core features of autism outside of clinical environments. We collected 100 public videos from YouTube of children ages 1–15 with either a self-reported diagnosis of an ASD (N = 45) or not (N = 55). Four non-clinical raters independently scored all videos using one of the most widely adopted tools for behavioral diagnosis of autism, the Autism Diagnostic Observation Schedule-Generic (ADOS). The classification accuracy was 96.8%, with 94.1% sensitivity and 100% specificity, the inter-rater correlation for the behavioral domains on the ADOS was 0.88, and the diagnoses matched a trained clinician in all but 3 of 22 randomly selected video cases. Despite the diversity of videos and non-clinical raters, our results indicate that it is possible to achieve high classification accuracy, sensitivity, and specificity as well as clinically acceptable inter-rater reliability with nonclinical personnel. Our results also demonstrate the potential for video-based detection of autism in short, unstructured home videos and further suggests that at least a percentage of the effort associated with detection and monitoring of autism may be mobilized and moved outside of traditional clinical environments.

Use of machine learning for behavioral distinction of autism and ADHD

Although autism spectrum disorder (ASD) and attention deficit hyperactivity disorder (ADHD) continue to rise in prevalence, together affecting >10% of today’s pediatric population, the methods of diagnosis remain subjective, cumbersome and time intensive. With gaps upward of a year between initial suspicion and diagnosis, valuable time where treatments and behavioral interventions could be applied is lost as these disorders remain undetected. Methods to quickly and accurately assess risk for these, and other, developmental disorders are necessary to streamline the process of diagnosis and provide families access to much-needed therapies sooner. Using forward feature selection, as well as undersampling and 10-fold cross-validation, we trained and tested six machine learning models on complete 65-item Social Responsiveness Scale score sheets from 2925 individuals with either ASD (n=2775) or ADHD (n=150). We found that five of the 65 behaviors measured by this screening tool were sufficient to distinguish ASD from ADHD with high accuracy (area under the curve=0.965). These results support the hypotheses that (1) machine learning can be used to discern between autism and ADHD with high accuracy and (2) this distinction can be made using a small number of commonly measured behaviors. Our findings show promise for use as an electronically administered, caregiver-directed resource for preliminary risk evaluation and/or pre-clinical screening and triage that could help to speed the diagnosis of these disorders.

Crowdsourced validation of a machine-learning classification system for autism and ADHD

Autism spectrum disorder (ASD) and attention deficit hyperactivity disorder (ADHD) together affect >10% of the children in the United States, but considerable behavioral overlaps between the two disorders can often complicate differential diagnosis. Currently, there is no screening test designed to differentiate between the two disorders, and with waiting times from initial suspicion to diagnosis upwards of a year, methods to quickly and accurately assess risk for these and other developmental disorders are desperately needed. In a previous study, we found that four machine-learning algorithms were able to accurately (area under the curve (AUC)>0.96) distinguish ASD from ADHD using only a small subset of items from the Social Responsiveness Scale (SRS). Here, we expand upon our prior work by including a novel crowdsourced data set of responses to our predefined top 15 SRS-derived questions from parents of children with ASD (n=248) or ADHD (n=174) to improve our model’s capability to generalize to new, ‘real-world’ data. By mixing these novel survey data with our initial archival sample (n=3417) and performing repeated cross-validation with subsampling, we created a classification algorithm that performs with AUC=0.89±0.01 using only 15 questions.

Brain-specific functional relationship networks inform autism spectrum disorder gene prediction

Autism spectrum disorder (ASD) is a neuropsychiatric disorder with strong evidence of genetic contribution, and increased research efforts have resulted in an ever-growing list of ASD candidate genes. However, only a fraction of the hundreds of nominated ASD-related genes have identified de novo or transmitted loss of function (LOF) mutations that can be directly attributed to the disorder. For this reason, a means of prioritizing candidate genes for ASD would help filter out false-positive results and allow researchers to focus on genes that are more likely to be causative. Here we constructed a machine learning model by leveraging a brain-specific functional relationship network (FRN) of genes to produce a genome-wide ranking of ASD risk genes. We rigorously validated our gene ranking using results from two independent sequencing experiments, together representing over 5000 simplex and multiplex ASD families. Finally, through functional enrichment analysis on our highly prioritized candidate gene network, we identified a small number of pathways that are key in early neural development, providing further support for their potential role in ASD.

A Review of Automated Methods for Detection of Myocardial Ischemia and Infarction Using Electrocardiogram and Electronic Health Records

There is a growing body of research focusing on automatic detection of ischemia and myocardial infarction (MI) using computer algorithms. In clinical settings, ischemia and MI are diagnosed using electrocardiogram (ECG) recordings as well as medical context including patient symptoms, medical history, and risk factors—information that is often stored in the electronic health records. The ECG signal is inspected to identify changes in the morphology such as ST-segment deviation and T-wave changes. Some of the proposed methods compute similar features automatically while others use nonconventional features such as wavelet coefficients. This review provides an overview of the methods that have been proposed in this area, focusing on their historical evolution, the publicly available datasets that they have used to evaluate their performance, and the details of their algorithms for ECG and EHR analysis. The validation strategies that have been used to evaluate the performance of the proposed methods are also presented. Finally, the paper provides recommendations for future research to address the shortcomings of the currently existing methods and practical considerations to make the proposed technical solutions applicable in clinical practice.

Cross-pollination of research findings, although uncommon, may accelerate discovery of human disease genes

BackgroundTechnological leaps in genome sequencing have resulted in a surge in discovery of human disease genes. These discoveries have led to increased clarity on the molecular pathology of disease and have also demonstrated considerable overlap in the genetic roots of human diseases. In light of this large genetic overlap, we tested whether cross-disease research approaches lead to faster, more impactful discoveries.MethodsWe leveraged several gene-disease association databases to calculate a Mutual Citation Score (MCS) for 10,853 pairs of genetically related diseases to measure the frequency of cross-citation between research fields. To assess the importance of cooperative research, we computed an Individual Disease Cooperation Score (ICS) and the average publication rate for each disease.ResultsFor all disease pairs with one gene in common, we found that the degree of genetic overlap was a poor predictor of cooperation (r2=0.3198) and that the vast majority of disease pairs (89.56%) never cited previous discoveries of the same gene in a different disease, irrespective of the level of genetic similarity between the diseases. A fraction (0.25%) of the pairs demonstrated cross-citation in greater than 5% of their published genetic discoveries and 0.037% cross-referenced discoveries more than 10% of the time. We found strong positive correlations between ICS and publication rate (r2=0.7931), and an even stronger correlation between the publication rate and the number of cross-referenced diseases (r2=0.8585). These results suggested that cross-disease research may have the potential to yield novel discoveries at a faster pace than singular disease research.ConclusionsOur findings suggest that the frequency of cross-disease study is low despite the high level of genetic similarity among many human diseases, and that collaborative methods may accelerate and increase the impact of new genetic discoveries. Until we have a better understanding of the taxonomy of human diseases, cross-disease research approaches should become the rule rather than the exception.

Sparsifying machine learning models identify stable subsets of predictive features for behavioral detection of autism

BackgroundAutism spectrum disorder (ASD) diagnosis can be delayed due in part to the time required for administration of standard exams, such as the Autism Diagnostic Observation Schedule (ADOS). Shorter and potentially mobilized approaches would help to alleviate bottlenecks in the healthcare system. Previous work using machine learning suggested that a subset of the behaviors measured by ADOS can achieve clinically acceptable levels of accuracy. Here we expand on this initial work to build sparse models that have higher potential to generalize to the clinical population.MethodsWe assembled a collection of score sheets for two ADOS modules, one for children with phrased speech (Module 2; 1319 ASD cases, 70 controls) and the other for children with verbal fluency (Module 3; 2870 ASD cases, 273 controls). We used sparsity/parsimony enforcing regularization techniques in a nested cross validation grid search to select features for 17 unique supervised learning models, encoding missing values as additional indicator features. We augmented our feature sets with gender and age to train minimal and interpretable classifiers capable of robust detection of ASD from non-ASD.ResultsBy applying 17 unique supervised learning methods across 5 classification families tuned for sparse use of features and to be within 1 standard error of the optimal model, we find reduced sets of 10 and 5 features used in a majority of models. We tested the performance of the most interpretable of these sparse models, including Logistic Regression with L2 regularization or Linear SVM with L1 regularization. We obtained an area under the ROC curve of 0.95 for ADOS Module 3 and 0.93 for ADOS Module 2 with less than or equal to 10 features.ConclusionsThe resulting models provide improved stability over previous machine learning efforts to minimize the time complexity of autism detection due to regularization and a small parameter space. These robustness techniques yield classifiers that are sparse, interpretable and that have potential to generalize to alternative modes of autism screening, diagnosis and monitoring, possibly including analysis of short home videos.

Rising interdisciplinary collaborations refine our understanding of autisms and give hope to more personalized solutions

Autism is heterogeneous, complex and arguably a condition of many conditions. Both the number of researchers and the number of research collaborations in the field of autism have been growing at unprecedented rates. Interdisciplinary collaborations have increased more than eightfold since the year 2000. In fact, most - if not all - areas of autism research are starting to converge, and these convergences are leading not only to a richer research network but also to a causal network for autism. This network can, and likely will, decode the many forms of autism into its various subcomponents, enabling increasingly more personalized approaches for both the detection and treatment of those different forms of autism.