SpExNeuro – A behavioural and brain imaging data collection of speech and executive functions

IntroductionThe relationship between speech and executive functions (EF) is well documented (Hagoort,2017; Novick et al., 2005), but the validity of speech biomarkers for cognitive performanceremains inconclusive (Robin et al., 2020). Moreover, the interactions between the subdomainsof EF and speech are under-researched, specifically regarding how they vary betweenindividuals. Foundational studies are needed to investigate these associations, focusing on thepotential predictive power of linguistic parameters. To this end, the SpEx study (Camilleri &Volkening et al., 2024) acquired speech and EF data from 148 participants, enablingcomprehensive analyses. Previous studies have used this dataset together with ML analysesto study the connection between EF and verbal fluency (Amunts et al., 2020, 2021). Adding tothese analyses, we here evaluated the predictiveness of EF by prosody. However, the SpExstudy only focused on behavioural data. To facilitate the investigation of neural mechanismsof EF-speech interactions, we here present the SpExNeuro study containing a larger sampleand neuroimaging data.MethodsAs part of the objectives of the SpEx study, we examined suprasegmental features aspredictors of EF performance (Kuhles et al., in review). EF performance was measured usingstandard tests, capturing 66 variables across cognitive flexibility, working memory, inhibition,and attention domains. 264 prosodic features, including frequency, energy, spectral, andtemporal dimensions, were extracted from the speech data using OpenSmile. ML analyseswith 10-fold cross-validation employed RF regressors to predict EF from prosodic features,evaluated by R² metrics. Confounding effects of sex, age, and education were regressed out.Building on SpEx, SpExNeuro expands the dataset with additional behavioral andneuroimaging data (Fig. 1). During a ~2 h MRI session, subjects perform EF core domain testsand speech tasks (verbal fluency, spontaneous speech). Brain activities are measured usingfMRI on a Siemens Prisma 3-T scanner with a 64-channel head coil. In addition to this task-based fMRI, RS, and DTI data are collected.ResultsAmong the predicted EF targets, reasonable model fits were found only for the EF targets TMTBTA and TMT BTB. However, deeper analyses uncovered significant confound leakage,revealing that the relationships between confounds (age, sex, and education) and the EFtargets inflated prediction accuracy (Fig. 2). These findings emphasise the critical need tocarefully control for confounding variables in ML pipelines and highlight significant risks in theinterpretation of ML predictions.To overcome the limitations of current datasets, a larger and more diverse validation set,including both behavioral and neuroimaging data, is urgently needed to ensure thegeneralisability of findings and elucidate underlying mechanisms. The ongoing SpExNeurostudy addresses these challenges by capturing a comprehensive range of variables acrossbehavioral, linguistic, neuroimaging, and neuroendocrine domains. The data will be openlyaccessible to maximise its utility for future research on individual differences in speech andEF.ConclusionsTogether, our studies highlight the need for robust methodologies, both experimental andcomputational, to deepen our understanding of EF and speech relationships. The SpEx andSpExNeuro datasets provide a critical resource for investigating complex interactions betweenspeech and EF, integrating behavioral, linguistic, and neuroimaging data. The data enablesthe study of individual differences in EF and speech and their interaction with demographic,hormonal, and neuroanatomical factors. Using our data, advanced ML analyses can uncovershared brain activation patterns, bridging behavioral and neural data, and enhancingindividualized biomarkers. We encourage scientists to leverage this growing dataset forcollaborative research.