Confound-free hand grip strength prediction: Synergy of advanced machine learning and neuroimaging

1427 

Abstract Submission 

Vera Komeyer^1,2,3, Simon Eickhoff^1,2, Jan Kasper^1,2, Christian Grefkes^4,5, Kaustubh Patil^1,2, Federico Raimondo^1,2

¹Institute of Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf, Duesseldorf, Germany, ²Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Center Juelich, Juelich, Germany, ³Department of Biology, Faculty of Mathematics and Natural Sciences, Heinrich-Heine-University Duesseldorf, Duesseldorf, Germany, ⁴Department of Neurology, University Hospital Frankfurt, Goethe University Frankfurt, Frankfurt am Main, Germany, ⁵Department of Neurology, University Hospital Cologne and Medical Faculty, University Cologne, Cologne, Germany

Vera Komeyer  
Institute of Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf|Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Center Juelich|Department of Biology, Faculty of Mathematics and Natural Sciences, Heinrich-Heine-University Duesseldorf
Duesseldorf, Germany|Juelich, Germany|Duesseldorf, Germany

Simon Eickhoff  
Institute of Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf|Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Center Juelich
Duesseldorf, Germany|Juelich, Germany

Jan Kasper  
Institute of Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf|Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Center Juelich
Duesseldorf, Germany|Juelich, Germany

Christian Grefkes  
Department of Neurology, University Hospital Frankfurt, Goethe University Frankfurt|Department of Neurology, University Hospital Cologne and Medical Faculty, University Cologne
Frankfurt am Main, Germany|Cologne, Germany

Kaustubh Patil  
Institute of Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf|Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Center Juelich
Duesseldorf, Germany|Juelich, Germany

Federico Raimondo  
Institute of Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf|Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Center Juelich
Duesseldorf, Germany|Juelich, Germany

Hand grip strength (HGS) not only reflects overall strength, but is also closely related to physical disability, cognitive decline and mortality [2,8,5]. Recognized by the WHO as a key marker for vitality in aging populations [1], HGS is a cost-efficient and reliable measure in clinical practice. Despite its ubiquity, the neural mechanisms governing HGS remain unclear. Our study systematically developed and evaluated the combination of neuroimaging-derived features with machine learning models to predict HGS. The aim was to identify models that are solely driven by brain information free from confounding factors such as sex, age and body composition, of which particularly sex but als age strongly correlates with HGS. Such confound-free models are essential to shed light on the neural basis of HGS.

We predicted HGS from 9 feature categories in the UK Biobank [3] (N = 22554-33136) including gray matter volume (GMV) [10], functional fALFF, LCOR, GCOR [7] (each 1088 ROIs), cortical thickness, white surface area, white matter hyperintensity (WMH) with PSMD [4], gray white contrast and a collection of 6 white matter microstructural characteristics (Fractional Anisotropy (FA), Mean Diffusivity (MD), free water volume fraction (ISOVF), orientation dispersion index (OD), intra-cellular volume fraction (ICVF), diffusion tensor mode (MO)). For each of the 9 categories 80% data were used to train seven algorithms in a 5-fold (nested) cross validation (CV) (Fig. 1A). Features were univariately, linearly adjusted for six confounder setups (Fig. 1B) in a CV-consistent manner. A final model was trained on the entire training data for out of sample (OOS) predictions on 20% held-out test data. Because of the particularly high influence of sex on linear predictions but less on non-linear predictions (Fig. 1B), we additionally classified sex with a XGBoost classifier with and without linear sex-adjustment to investigate the residual non-linear sex information after confound removal. The same analyses were performed on sex-split data. Six most successful sex-split and age-regressed models underwent SHAP analysis [6] to investigate relevant features for successful models.

The sex-mixed sample analysis identified GMV, white surface, fALFF and white matter as most predictive features (Fig. 1B). Predictability decreased noticeably when adjusting for sex and age, but didn't drop further when removing more confounders (Fig. 1B). Non-linear algorithms performed better than linear ones in the sex-age-adjusted scenario (Fig. 1B purple vs. blue). Classifying sex with and without linear removal of sex suggests that non-linear algorithm superiority is driven by residual non-linear sex information in the features (Fig. 1C). Nonetheless, non-linear approaches also showed superior performance in "sex-split" models, even after controlling for age (Fig. 1D). GMV, fALFF and white matter were most resilient for this very stringent confounder control (Fig. 1D). For these three feature categories XGBoost excelled other non-linear algorithms, leading to the six (3 per sex) best models: r(m)GMV = 0.18, r(f)GMV = 0.20; r(m)fALFF = 0.18, r(f)fALFF = 0.23; r(m)WM = 0.21, r(f)WM = 0.23 (Fig. 1F). Interpretative SHAP analyses suggested GMV's importance in anterior globus pallidus (Fig. 2A, B) and microstructural characteristics of sensory input bundles to the thalamus and thalamo-cortical tracts (Fig. 2E, F) as neural correlates for successful, confound-free HGS predictions.

Our exhaustive evaluation of ML models and features from diverse MRI modalities identified six effective models for predicting HGS under stringent confounding constraints. Such strict constraints allow us to conclude that the successful models rely on neural information independent of any sex and linear age signal in the features. Further investigations will delve into feature collinearities and the intricate interplay of feature modalities to gain a better understanding of the underlying biology.

Classification and Predictive Modeling ¹

Multivariate Approaches ²

Motor Behavior Other

Cerebellum

Cortex

FUNCTIONAL MRI

Machine Learning

Motor

MRI

Multivariate

STRUCTURAL MRI

Sub-Cortical

White Matter