Figure 2.

The structured deep embedding model has no adverse effects on the model performance

(A) The prediction performance scores for all five models were similar although slightly improved with the structured deep embedding model that has more hierarchical groupings (e.g., base model versus composite model_g1). The boxplots show the “minimum,” 1^st quartile (Q1, 25^th percentile), median (Q2, 50^th percentile), and 3^rd quartile (Q3, 75^th percentile), and the “maximum,” where the minimum and maximum values are defined as Q1 – 1.5 ∗ interquartile range (IQR) and Q3 + 1.5 ∗ IQR, respectively.

(B) Early detection performance by measuring predictivity with the limited dataset available at the months prior to diagnosis. The error bars indicate 95% confidence intervals.

(C) We measured total entropy of clusters (i.e., purity of the clusters classified into PC and non-PC) resulted from each model. The error bars indicate 95% confidence intervals.