Explainable Machine Learning Approach to Predict and Explain the Relationship between Task-based fMRI and Individual Differences in Cognition

Pornpattananangkul, Narun; Bartonicek, Adam; Wang, Yue; Stringaris, Argyris

doi:10.1101/2020.10.21.348367

Cited by 3 publications

(3 citation statements)

References 98 publications

(220 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For task-fMRI sets of features, we used unthresholded generalised-linear model (GLM) contrasts, averaged across two runs 9,56,57 . These contrasts were embedded in the brain parcels based on the FreeSurfer’s atlases 58 : 148 cortical-surface Destrieux parcels 59 and subcortical-volumetric 19 ASEG parcels 60 , leaving 167 features in each task-fMRI set of features.…”

Section: Methodsmentioning

confidence: 99%

The relationship between cognitive abilities and mental health as represented by cognitive abilities at the neural and genetic levels of analysis

Wang,

Anney,

Pat

2024

Preprint

Self Cite

View full text Add to dashboard Cite

SummaryBackgroundA leading transdiagnostic framework for psychiatry, NIMH Research Domain Criteria (RDoC), posits that cognitive abilities are a major factor (known as functional domain) underlying mental health across various diagnoses. Specifically, RDoC assumes the relationship between cognitive abilities and mental health to be 1) manifested across neural and geneticunits of analysis, 2) environmentally situated, and 3) reliable. These central assumptions have not been empirically validated. To address this, we used data from the Adolescent Brain Cognitive Development (ABCD) Study, which included 11,876 participants (5,680 females). The study spanned two years and focused on children aged 9-10 and 11-12.MethodsWe applied machine learning to make out-of-sample predictions of cognitive abilities based on measures of mental health (emotional/behavioural problems, at-risk personalities), neuroimaging (45 types of brain MRI), polygenic scores (three definitions), and socio-demographics, lifestyles and developments (44 variables e.g., parental income, screen use). Using linear-mixed-model commonality analyses, we then examined the extent to which the relationship between cognitive abilities and mental health was explained by the other measures.FindingsMental health predicted cognitive abilities of unseen children (i.e., those who were not part of the modelling process) atr=.39. At baseline, this cognitive-abilities-mental-health relationship was accounted for by neuroimaging (69%), by polygenic scores (18%) and by socio-demographics, lifestyles and developments (70%). Moreover, the variance in the relationship between cognitive abilities and mental health that was captured by socio-demographics, lifestyles and developments was explained by neuroimaging (66%) and polygenic scores (22%). These patterns were consistent across the two-time points.InterpretationConsistent with RDoC, the cognitive abilities and mental health relationship was 1) manifested in both neuroimaging and polygenic scores, 2) explained by socio-demographics, lifestyles and developments and 3) reliable across two years. This supports RDoC’s view of cognitive abilities as an integrative-functional domain for the aetiology of mental health.FundingHealth Research Council Funding of New Zealand (21/618) and University of OtagoResearch in ContextEvidence before this studyIt has been over a decade since the initiation of Research Domain Criteria (RDoC) by the US National Institute of Mental Health. The first article outlining its framework (Insel et al., 2010,AJP) has been cited over 6,650 times according to Google Scholar as of September 25th, 2023. In one estimate (Pacheco et al., 2022,JCPP), 644 RDoC-themed grants were funded between 2008-19 in the US. RDoC has also asserted its impact on psychiatric research outside of the US, as seen in frameworks such as the Psychiatric Ratings using Intermediate Stratified Markers (PRISM) funded by EU-Innovative Medicine Initiative (Kas et al., 2019,Neuroscience and Biobehavioral Reviews). Accordingly, RDoC has had a great influence on the landscape of modern psychiatric research. Yet, most RDoC-themed studies have focused on applying the framework (e.g., to improve the aetiology of certain mental illnesses), but few, if any, have tested its central assumptions regarding the properties of its functional domains. These assumptions were largely expert-drivenNY. Here, we focused on empirically validating the properties of one of the six major functional domains, cognitive systems (note we called them cognitive abilities here to map closer to a term used in cognitive literature).Added value of this studyDuring the time when RDoC was first developed, there was no large-scale neuroimaging and genetic data appropriate to validate the framework. The emergence of studies such as the Adolescent Brain Cognitive Development (ABCD) offers not only neuroimaging and genetic data but also high-quality data on phenotypes (e.g., mental health and cognitive abilities) along with socio-demographics and lifestyles which allows us to validate the RDoC framework. Using machine learning, we were able to quantify the extent to which the relationship between cognitive abilities and mental health was 1) manifested in both neuroimaging and polygenic scores, 2) explained by socio-demographics, lifestyles, and developments and 3) reliable across two years. Accordingly, our study validated the role of cognitive abilities as a functional domain according to RDoC.Implications of all of the available evidenceWe provide empirical evidence to support RDoC’s view of cognitive abilities as a transdiagnostic domain for mental health. By doing so, we laid a foundation to achieve the four goals set by RdoC (Pacheco et al., 2022,JCPP). First, we have addressed the aetiology of mental health by showing the reliable and predictable transdiagnostic relationship between cognitive abilities and mental health. Second, we also identified neuroimaging and genetic biomarkers that explain this relationship. Third, with our findings on the socio-demographics, lifestyles and developments, we touched on target identification. Finally, our use of machine learning to make out-of-sample predictions on individual children could lead to further development of personalised interventions. Overall, our findings lend credibility to RDoC research focusing on cognitive abilities.

show abstract

Section: Methodsmentioning

confidence: 99%

The relationship between cognitive abilities and mental health as represented by cognitive abilities at the neural and genetic levels of analysis

Wang,

Anney,

Pat

2024

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…It includes hyperparameters, which determine the degree to which the sum of either the squared (Ridge) or absolute (LASSO) slopes is penalized. Elastic Net is well suited to select the features that best predict the outcome 27,28 .…”

Section: Ethics Approval and Subject Consent The Clinical Research Et...mentioning

confidence: 99%

MRI based radiomics enhances prediction of neurodevelopmental outcome in very preterm neonates

Wagner

Guo

et al. 2022

Sci Rep

View full text Add to dashboard Cite

To predict adverse neurodevelopmental outcome of very preterm neonates. A total of 166 preterm neonates born between 24–32 weeks’ gestation underwent brain MRI early in life. Radiomics features were extracted from T1- and T2- weighted images. Motor, cognitive, and language outcomes were assessed at a corrected age of 18 and 33 months and 4.5 years. Elastic Net was implemented to select the clinical and radiomic features that best predicted outcome. The area under the receiver operating characteristic (AUROC) curve was used to determine the predictive ability of each feature set. Clinical variables predicted cognitive outcome at 18 months with AUROC 0.76 and motor outcome at 4.5 years with AUROC 0.78. T1-radiomics features showed better prediction than T2-radiomics on the total motor outcome at 18 months and gross motor outcome at 33 months (AUROC: 0.81 vs 0.66 and 0.77 vs 0.7). T2-radiomics features were superior in two 4.5-year motor outcomes (AUROC: 0.78 vs 0.64 and 0.8 vs 0.57). Combining clinical parameters and radiomics features improved model performance in motor outcome at 4.5 years (AUROC: 0.84 vs 0.8). Radiomic features outperformed clinical variables for the prediction of adverse motor outcomes. Adding clinical variables to the radiomics model enhanced predictive performance.

show abstract

“…To gain neurobiological insights of the G-Factor, the predictive models have to be interpretable 9 , allowing researchers to demonstrate which neural indices contribute to the prediction. For a typical unimodal analysis, we 25 recently proposed a framework, known as eNetXplorer 26 , that applies permutation along with Elastic Net to enable a statistical inference for each brain feature. Briefly, researchers first fit two sets of many Elastic Net models: one for predicting the true response variable (target models) and the other for predicting a randomly permuted response variable (null models).…”

Section: Introductionmentioning

confidence: 99%

Longitudinally stable, brain-based predictive models mediate the relationships between childhood cognition and socio-demographic, psychological and genetic factors

Pornpattananangkul

Wang

Stringaris³

2021

Preprint

Self Cite

View full text Add to dashboard Cite

General intelligence, or G-Factor, of children is associated with important life outcomes, including educational attainment, employment, health, and mortality. Thus, determining the neural predictor of the G-Factor of children has become one of the main tasks in neuroscience. Here we aim to build neuroimaging-based predictive models of children's general intelligence that are longitudinally stable across two years. To achieve this goal, we used large-scale, longitudinal data with multiple neuroimaging modalities from the Adolescent Brain Cognitive Development (ABCD) study (n ~11k). We first computed modality-specific models from six MRI modalities (three task-based, resting-state, structural, and diffusion tensor imaging) of the baseline data using Elastic Net. We, then, combined the predicted values of all modality-specific models using Random Forest Opportunistic Stacking. The stacked model allowed us to predict the G-Factor of unseen, same-age (9-10-year-old) children at Pearson r=.44, better than any other modality-specific models. Based on permutation tests, this prediction was predominantly driven by activity in the parietal and frontal areas during the N-Back working-memory task. Importantly, this model was generalizable to unseen children from the follow-up data who were two years older at r=.41. Moreover, this model allowed for missingness in the data, making it possible for us to maintain around 73% of the data that would otherwise be excluded due to different artifacts occurring to any modalities. Accordingly, we developed an MRI-multimodal predictive model for the G-Factor of children that is 1) stable across years, 2) interpretable and 3) able to handle missing values.

show abstract

Explainable Machine Learning Approach to Predict and Explain the Relationship between Task-based fMRI and Individual Differences in Cognition

Cited by 3 publications

References 98 publications

The relationship between cognitive abilities and mental health as represented by cognitive abilities at the neural and genetic levels of analysis

The relationship between cognitive abilities and mental health as represented by cognitive abilities at the neural and genetic levels of analysis

MRI based radiomics enhances prediction of neurodevelopmental outcome in very preterm neonates

Longitudinally stable, brain-based predictive models mediate the relationships between childhood cognition and socio-demographic, psychological and genetic factors

Contact Info

Product

Resources

About