The recent Deoni et al. (2013) manuscript proposed that breastfeeding was associated with increased cognitive ability and white-matter in older children (over 26 months), using ms-DESPOT MRI imaging to indirectly measure white matter in children who were either breastfed, formula fed, or combined breast+formula fed. In this response, we identify limitations in drawing causal inference among white matter, cognitive ability, and breastfeeding. We propose that the observed cognitive and neurodevelopmental differences between breastfed and formula-fed infants might actually be caused by the premature introduction of cow's milk in the second year of life, among other contributing factors. The implication of a causal relationship between intelligence and white matter metrics, especially in a developmentally young population, is premature given the recency of this field. The original analyses did not control for important covariates; when comparing both white matter and test scores, mothers were not controlled for age and socio-economic status (SES) and their children were not controlled for gender. Raw test scores, instead of age-adjusted test scores, were used even though the children were of different ages. Mothers were not controlled for reason(s) not to breastfeed, even though many prenatal factors are known to predict this such as stress, parity, obesity, and smoking habits. The observed cognitive ability and white matter benefits identified primarily within the long-term breastfed children are at least partially attributable to other factors such as age, gender, and SES. We suggest methodological approaches to removing such ambiguity, and ways to dissociate cause from effect. The formula and breastfeeding groups didn't show differences until the "formula fed" children likely had been fed cow's milk for longer than they had been fed formula, at 2.2 years. The greatest cognitive differences however were observed within the high SES breastfed infants depending on breastfeeding duration; infants who were breastfed over 15 months showed increased cognitive ability compared to those breastfed less than months. This implicates the source of dairy during the second year of life, and not other SES factors or infant formula, as the most likely nutritional factor responsible for the observed differences within the breastfed children. Given the known nutritional deficiencies of cow's milk, these findings imply infants who received cow's milk during the second year of life were at a disadvantage compared to those who were breastfed, independent of whether they were fed formula or breast milk during the first year of life. This evidence suggests that infants should receive formula in lieu of cow's milk when breast milk is unavailable as a dairy source, until roughly 2 years of age.

Functional network connectivity (FNC) is a method of analyzing the temporal relationship of anatomical brain components, comparing the synchronicity between patient groups or conditions. We use functional-connectivity measures between independent components to classify between Schizophrenia patients and healthy controls during resting-state. Connectivity is measured using a variety of graph-theoretic connectivity measures such as graph density, average path length, and small-worldness. The Schizophrenia patients showed significantly less clustering (transitivity) among components than healthy controls (p < 0.05, corrected) with networks less likely to be connected, and also showed lower small-world connectivity than healthy controls. Using only these connectivity measures, an SVM classifier (without parameter tuning) could discriminate between Schizophrenia patients and healthy controls with 65% accuracy, compared to 51% chance. This implies that the global functional connectivity between resting-state networks is altered in Schizophrenia, with networks more likely to be disconnected and behave dissimilarly for diseased patients. We present this research finding as a tutorial using the publicly available COBRE dataset of 146 Schizophrenia patients and healthy controls, provided as part of the 1000 Functional Connectomes Project. We demonstrate preprocessing, using independent component analysis (ICA) to nominate networks, computing graph-theoretic connectivity measures, and finally using these connectivity measures to either classify between patient groups or assess between-group differences using formal hypothesis testing. All necessary code is provided for both running command-line FSL preprocessing, and for computing all statistical measures and SVM classification within R. Collectively, this work presents not just findings of diminished FNC among resting-state networks in Schizophrenia, but also a practical connectivity tutorial.

This study used social network analysis to evaluate whether sex heterophily, the degree to which peers are different in sex, between 126 children with autism (ages 5-12 years) and their peers affected social network connectivity. Results indicate that: (1) the quantity and sex of friends were more important in predicting social network connectivity than the relational characteristics of the friends (friendship nominations and social network salience/popularity); and (2) sex heterophily is an important factor in predicting social network connectivity. For males with autism, having friends of the same sex was associated with better social network connectivity; this was not true for females with autism. These findings have important implications for the selection of peer models for elementary-aged children with autism.

The recent Chu et al. (2012) manuscript discusses two key findings regarding feature selection (FS): (1) data driven FS was no better than using whole brain voxel data and (2) a priori biological knowledge was effective to guide FS. Use of FS is highly relevant in neuroimaging-based machine learning, as the number of attributes can greatly exceed the number of exemplars. We strongly endorse their demonstration of both of these findings, and we provide additional important practical and theoretical arguments as to why, in their case, the data-driven FS methods they implemented did not result in improved accuracy. Further, we emphasize that the data-driven FS methods they tested performed approximately as well as the all-voxel case. We discuss why a sparse model may be favored over a complex one with similar performance. We caution readers that the findings in the Chu et al. report should not be generalized to all data-driven FS methods.

Although children with autism spectrum disorder are frequently included in mainstream classrooms, it is not known how their social networks change compared to typically developing children and whether the factors predictive of this change may be unique. This study identified and compared predictors of social connectivity of children with and without autism spectrum disorder using a social network analysis. Participants included 182 children with autism spectrum disorder and 152 children without autism spectrum disorder, aged 5-12 years in 152 general education K-5 classrooms. General linear models were used to compare how age, classroom size, gender, baseline connectivity, diagnosis, and intelligence quotient predicted changes in social connectivity (closeness). Gender and classroom size had a unique interaction in predicting final social connectivity and the change in connectivity for children with autism spectrum disorder; boys who were placed in larger classrooms showed increased social network fragmentation. This increased fragmentation for boys when placed in larger classrooms was not seen in typically developing boys. These results have implications regarding placement, intervention objectives, and ongoing school support that aimed to increase the social success of children with autism spectrum disorder in public schools.

Background

Brain networks in fMRI are typically identified using spatial independent component analysis (ICA), yet other mathematical constraints provide alternate biologically-plausible frameworks for generating brain networks. Non-negative matrix factorization (NMF) would suppress negative BOLD signal by enforcing positivity. Spatial sparse coding algorithms (L1 Regularized Learning and K-SVD) would impose local specialization and a discouragement of multitasking, where the total observed activity in a single voxel originates from a restricted number of possible brain networks.

New method

The assumptions of independence, positivity, and sparsity to encode task-related brain networks are compared; the resulting brain networks within scan for different constraints are used as basis functions to encode observed functional activity. These encodings are then decoded using machine learning, by using the time series weights to predict within scan whether a subject is viewing a video, listening to an audio cue, or at rest, in 304 fMRI scans from 51 subjects.

Results and comparison with existing method

The sparse coding algorithm of L1 Regularized Learning outperformed 4 variations of ICA (p<0.001) for predicting the task being performed within each scan using artifact-cleaned components. The NMF algorithms, which suppressed negative BOLD signal, had the poorest accuracy compared to the ICA and sparse coding algorithms. Holding constant the effect of the extraction algorithm, encodings using sparser spatial networks (containing more zero-valued voxels) had higher classification accuracy (p<0.001). Lower classification accuracy occurred when the extracted spatial maps contained more CSF regions (p<0.001).

Conclusion

The success of sparse coding algorithms suggests that algorithms which enforce sparsity, discourage multitasking, and promote local specialization may capture better the underlying source processes than those which allow inexhaustible local processes such as ICA. Negative BOLD signal may capture task-related activations.

Objective: Total scale scores derived by summing ratings from the 30-item PANSS are commonly used in clinical trial research to measure overall symptom severity, and percentage reductions in the total scores are sometimes used to document the efficacy of treatment. Acknowledging that some patients may have substantial changes in PANSS total scores but still be sufficiently symptomatic to warrant diagnosis, ratings on a subset of 8 items, referred to here as the "Remission set," are sometimes used to determine if patients' symptoms no longer satisfy diagnostic criteria. An unanswered question remains: is the goal of treatment better conceptualized as reduction in overall symptom severity, or reduction in symptoms below the threshold for diagnosis? We evaluated the psychometric properties of PANSS total scores, to assess whether having low symptom severity post-treatment is equivalent to attaining Remission. Design: We applied a bifactor item response theory (IRT) model to post-treatment PANSS ratings of 3,647 subjects diagnosed with schizophrenia assessed at the termination of 11 clinical trials. The bifactor model specified one general dimension to reflect overall symptom severity, and five domain-specific dimensions. We assessed how PANSS item discrimination and information parameters varied across the range of overall symptom severity (θ), with a special focus on low levels of symptoms (i.e., θ<-1), which we refer to as "Relief" from symptoms. A score of θ=-1 corresponds to an expected PANSS item score of 1.83, a rating between "Absent" and "Minimal" for a PANSS symptom. Results: The application of the bifactor IRT model revealed: (1) 88% of total score variation was attributable to variation in general symptom severity, and only 8% reflected secondary domain factors. This implies that a general factor may provide a good indicator of symptom severity, and that interpretation is not overly complicated by multidimensionality; (2) Post-treatment, 534 individuals (about 15% of the whole sample) scored in the "Relief" range of general symptom severity, but more than twice that number (n = 1351) satisfied Remission criteria (37%). 2 in 3 Remitted patients had scores that were not in a low symptom range (corresponding to Absent or Minimal item scores); (3) PANSS items vary greatly in their ability to measure the general symptom severity dimension; while many items are highly discriminating and relatively "pure" indicators of general symptom severity (delusions, conceptual disorganization), others are better indicators of specific dimensions (blunted affect, depression). The utility of a given PANSS item for assessing a patient depended on the illness level of the patient. Conclusion: Satisfying conventional Remission criteria was not strongly associated with low levels of symptoms. The items providing the most information for patients in the symptom Relief range were Delusions, Preoccupation, Suspiciousness Persecution, Unusual Thought Content, Conceptual Disorganization, Stereotyped Thinking, Active Social Avoidance, and Lack of Judgment and Insight. Lower scores on these items (item scores ≤2) were strongly associated with having a low latent trait θ or experiencing overall symptom relief. The inter-rater agreement between Remission and Relief subjects suggested that these criteria identified different subsets of patients. Alternative subsets of items may offer better indicators of general symptom severity and provide better discrimination (and lower standard errors) for scaling individuals and judging symptom relief, where the "best" subset of items ultimately depends on the illness range and treatment phase being evaluated.

Sensory features are common and impairing in autism spectrum disorder (ASD), but there are few observational sensory assessments that are valid across ages. We used the Sensory Processing 3-Dimensional (SP3-D) observed Assessment and parent-reported Inventory to examine sensory responsivity in 41 ASD and 33 typically-developing (TD) youth across 7-17 years. ASD youth had higher and more variable observed and reported sensory responsivity symptoms compared to TD, but the two measures were not correlated. Observed sensory over-responsivity (SOR) and sensory craving (SC) decreased with age in ASD, though SOR remained higher in ASD versus TD through adolescence. Results suggest that in ASD, the SP3-D Assessment can identify SOR through adolescence, and that there is value in integrating multiple sensory measures.

Objectives

An estimated 25% of type two diabetes mellitus (DM2) patients in the United States are undiagnosed due to inadequate screening, because it is prohibitive to administer laboratory tests to everyone. We assess whether electronic health record (EHR) phenotyping could improve DM2 screening compared to conventional models, even when records are incomplete and not recorded systematically across patients and practice locations, as is typically seen in practice.

Methods

In this cross-sectional, retrospective study, EHR data from 9948 US patients were used to develop a pre-screening tool to predict current DM2, using multivariate logistic regression and a random-forests probabilistic model for out-of-sample validation. We compared (1) a full EHR model containing commonly prescribed medications, diagnoses (as ICD9 categories), and conventional predictors, (2) a restricted EHR DX model which excluded medications, and (3) a conventional model containing basic predictors and their interactions (BMI, age, sex, smoking status, hypertension).

Results

Using a patient's full EHR or restricted EHR was superior to using basic covariates alone for detecting individuals with diabetes (hierarchical X(2) test, p<0.001). Migraines, depot medroxyprogesterone acetate, and cardiac dysrhythmias were associated negatively with DM2, while sexual and gender identity disorder diagnosis, viral and chlamydial infections, and herpes zoster were associated positively. Adding EHR phenotypes improved classification; the AUC for the full EHR Model, EHR DX model, and conventional model using logistic regression, were 84.9%, 83.2%, and 75.0% respectively. For random forest machine learning out-of-sample prediction, accuracy also was improved when using EHR phenotypes; the AUC values were 81.3%, 79.6%, and 74.8%, respectively. Improved AUCs reflect better performance for most thresholds that balance sensitivity and specificity.

Conclusions

EHR phenotyping resulted in markedly superior detection of DM2, even in the face of missing and unsystematically recorded data, based on the ROC curves. EHR phenotypes could more efficiently identify which patients do require, and don't require, further laboratory screening. When applied to the current number of undiagnosed individuals in the United States, we predict that incorporating EHR phenotype screening would identify an additional 400,000 patients with active, untreated diabetes compared to the conventional pre-screening models.