Machine learning to discover factors predicting volume of white matter hyperintensities: Insights from the UK Biobank.

INTRODUCTION

Brain white matter hyperintensities (WMHs) reflect the risks of stroke, dementia, and overall mortality.

METHODS

We used a hypothesis-free gradient boosting decision tree (GBDT) approach and conventional statistical methods to discover risk factors associated with volume of WMHs. The GBDT models considered data on 2891 input features, collected ∼10 years prior to volume of WMH measurements from 44,053 participants. Top 3% of features, ranked by Shapley values, were taken forward to epidemiological analyses using linear regression.

RESULTS

Adiposity, lung function, and indicators of metabolic health (eg, glycated hemoglobin, hypertension, alkaline phosphatase, microalbumin, and urate) contribute to WMH prediction. Of lifestyle factors, smoking had the strongest association. Time spent outdoors, creatinine, and several red blood cell indices were among the identified less-known predictors of WMHs.

CONCLUSIONS

Obesity, high blood pressure, lung function, metabolic abnormalities, and lifestyle are key contributors to WMHs, providing opportunities to prevent or reduce their development.

HIGHLIGHTS

Obesity and related metabolic abnormalities were linked with WMHs.Associations with time spent outdoors, creatinine, some red blood cell indices and height were among the less-known risk factors identified.Action on blood pressure, metabolic abnormalities, and adequate oxygenation may help to prevent WMHs.Biomarker links may suggest simple blood tests could aid in early dementia prediction.