Integration of Brain and Behavior Measures for Identification of Data-Driven Groups Cutting Across Children with ASD, ADHD, or OCD

Jacobs, Grace; Voineskos, Aristotle N.; Hawco, Colin; Stefanik, Laura; Forde, Natalie J.; Dickie, Erin W.; Lai, Meng‐Chuan; Szatmári, Péter; Schachar, Russell; Crosbie, Jennifer; Arnold, Paul; Goldenberg, Anna; Erdman, Lauren; Lerch, Jason P.; Anagnostou, Evdokia; Ameis, Stephanie H.

doi:10.1101/2020.02.11.944744

Cited by 17 publications

(23 citation statements)

References 52 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…SNF belongs to a broader family of techniques of multiview learning algorithms (e.g., multiple kernel learning, multi-table matrix factorization). We opted to use SNF because (1) it is an unsupervised learning technique, (2) it is explicitly optimized to control for differing dimensionalities amongst input data modalities [87], and (3) it has been shown to be effective at disentangling heterogeneity in psychiatric populations [35,76].…”

Section: Similarity Network Fusionmentioning

confidence: 99%

“…Despite extensive validation in the original article [87], its successful application to similar neuroimaging datasets [35,76], and the exhaustive parameter search employed in the current article, we wanted to assess the stability of SNF as it applies to the current dataset. We first compare clustering assignments generated from SNF when varying the dimensionality of cortical thickness data (see Data dimensionality variation).…”

Section: Stability Of Snfmentioning

confidence: 99%

“…In the main text we use a squared Euclidean distance function to generate patient similarity networks from the original data feature matrices. While there is precedent for this choice [35,76,87], we also wanted to investigate the impact of the chosen distance metric on our results. Thus, we repeated the analyses described in the main text using two alternative distance metricscityblock (or Manhattan) distance and cosine distancefinding that both clustering results (NMI cityblock = 0.74, NMI cosine = 0.68) and embedding dimensions (average r cityblock = 0.93 ± 0.10 [0.65-0.99], r cosine = 0.87 ± 0.10 [0.70-0.96]) were highly consistent with squared Euclidean distance.…”

Section: Alternative Distance Metricsmentioning

confidence: 99%

“…A principal challenge is to parsimoniously integrate these multi-view data in order to take full advantage of each source of information [57]. How to account for multiple sources of data to characterize heterogeneous patient samples is a topic of significant interest in computational medicine [40,58], with important applications for oncology [62,87], psychiatry, [35,55,76], and neurology [21]. Indeed, recent advances in techniques like multiple kernel learning have yielded promis-ing results for integrating disparate data modalities in the context of supervised and unsupervised problems [19,99].…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Multimodal phenotypic axes of Parkinson’s disease

Markello¹,

Shafiei²,

Tremblay³

et al. 2020

Preprint

View full text Add to dashboard Cite

Individuals with Parkinson's disease present with a complex clinical phenotype, encompassing sleep, motor, cognitive, and affective disturbances. However, characterizations of PD are typically made for the "average" patient, ignoring patient heterogeneity and obscuring important individual differences. Modern large-scale data sharing efforts provide a unique opportunity to precisely investigate individual patient characteristics, but there exists no analytic framework for comprehensively integrating data modalities. Here we apply an unsupervised learning method-similarity network fusion-to objectively integrate MRI morphometry, dopamine active transporter binding, protein assays, and clinical measurements from n = 186 individuals with de novo Parkinson's disease from the Parkinson's Progression Markers Initiative. We show that multimodal fusion captures inter-dependencies among data modalities that would otherwise be overlooked by field standard techniques like data concatenation. We then examine how patient subgroups derived from fused data map onto clinical phenotypes, and how neuroimaging data is critical to this delineation. Finally, we identify a compact set of phenotypic axes that span the patient population, demonstrating that this continuous, low-dimensional projection of individual patients presents a more parsimonious representation of heterogeneity in the sample compared to discrete biotypes. Altogether, these findings showcase the potential of similarity network fusion for combining multimodal data in heterogeneous patient populations.

show abstract

Section: Similarity Network Fusionmentioning

confidence: 99%

Section: Stability Of Snfmentioning

confidence: 99%

Section: Alternative Distance Metricsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Multimodal phenotypic axes of Parkinson’s disease

Markello¹,

Shafiei²,

Tremblay³

et al. 2020

Preprint

View full text Add to dashboard Cite

show abstract

“…A principal challenge is to parsimoniously integrate these multi-view data in order to take full advantage of each source of information 15 . How to account for multiple sources of data to characterize heterogeneous patient samples is a topic of significant interest in computational medicine 16,17 , with important applications for oncology 18,19 , psychiatry [20][21][22] , and neurology 7 . Indeed, recent advances in techniques like multiple kernel learning have yielded promising results for integrating disparate data modalities in the context of supervised and unsupervised problems 23,24 .…”

Section: Introductionmentioning

confidence: 99%

Multimodal phenotypic axes of Parkinson’s disease

Markello

Shafiei

Tremblay

et al. 2021

npj Parkinsons Dis.

View full text Add to dashboard Cite

Individuals with Parkinson’s disease present with a complex clinical phenotype, encompassing sleep, motor, cognitive, and affective disturbances. However, characterizations of PD are typically made for the “average” patient, ignoring patient heterogeneity and obscuring important individual differences. Modern large-scale data sharing efforts provide a unique opportunity to precisely investigate individual patient characteristics, but there exists no analytic framework for comprehensively integrating data modalities. Here we apply an unsupervised learning method—similarity network fusion—to objectively integrate MRI morphometry, dopamine active transporter binding, protein assays, and clinical measurements from n = 186 individuals with de novo Parkinson’s disease from the Parkinson’s Progression Markers Initiative. We show that multimodal fusion captures inter-dependencies among data modalities that would otherwise be overlooked by field standard techniques like data concatenation. We then examine how patient subgroups derived from the fused data map onto clinical phenotypes, and how neuroimaging data is critical to this delineation. Finally, we identify a compact set of phenotypic axes that span the patient population, demonstrating that this continuous, low-dimensional projection of individual patients presents a more parsimonious representation of heterogeneity in the sample compared to discrete biotypes. Altogether, these findings showcase the potential of similarity network fusion for combining multimodal data in heterogeneous patient populations.

show abstract

Clinical, Brain, and Multilevel Clustering in Early Psychosis and Affective Stages

et al. 2022

View full text Add to dashboard Cite

IMPORTANCEApproaches are needed to stratify individuals in early psychosis stages beyond positive symptom severity to investigate specificity related to affective and normative variation and to validate solutions with premorbid, longitudinal, and genetic risk measures. OBJECTIVE To use machine learning techniques to cluster, compare, and combine subgroup solutions using clinical and brain structural imaging data from early psychosis and depression stages. DESIGN, SETTING, AND PARTICIPANTSA multisite, naturalistic, longitudinal cohort study (10 sites in 5 European countries; including major follow-up intervals at 9 and 18 months) with a referred patient sample of those with clinical high risk for psychosis (CHR-P), recent-onset psychosis (ROP), recent-onset depression (ROD), and healthy controls were

show abstract

Integration of Brain and Behavior Measures for Identification of Data-Driven Groups Cutting Across Children with ASD, ADHD, or OCD

Cited by 17 publications

References 52 publications

Multimodal phenotypic axes of Parkinson’s disease

Multimodal phenotypic axes of Parkinson’s disease

Multimodal phenotypic axes of Parkinson’s disease

Clinical, Brain, and Multilevel Clustering in Early Psychosis and Affective Stages

Contact Info

Product

Resources

About