Fig. 4.

Machine learning model based on cell cluster frequencies can partly reconstruct lifestyle score. A) Performance of an elastic net machine learning model based on cell cluster frequencies (n = 80), age and sex trained to predict lifestyle score. Observed compared to predicted lifestyle score based on training (80%) and test data (20%; n = 5 samples per location) are shown. Using cell frequency data, we can explain ∼30% of the variance in lifestyle scores (leave-out test data). B) Feature importance of all features that remained in the model after feature shrinkage/regularization. Clusters previously associated with either location or lifestyle score (n = 17) are indicated (∗). Three clusters have not been associated with location nor lifestyle score in previous analyses. C) Feature stability across bootstraps. All features from the models fitted with the optimized tuning parameters (penalty/mixture) were extracted. The number of times a feature was selected across bootstrap samples serves as a score for stability of that feature (maximum score = 2000).