Figure 1.

The WSS-DL method pipeline. Step 1: creation of silver standard labels $Y^{\ast }$ from $X_{surr}$ using the MAP algorithm; this step may be omitted when $X_{surr}$ is a viable surrogate that can be directly used as $Y^{\ast }$. Step 2: creation of the enhanced-silver-standard label $Y^{**}$ using neural network, with $Y^{\ast }$ as the outcome, $X_{full}$ as input feature, and a subset of $Y$ for fine-tuning. Step 3: final prediction of $Y$ using logistic regression, with the enhanced-silver-standard labels and a minimal set of informative features $X_{sel}$ as input features. The different sizes of $\text{Y}$ represented in steps 2 and 3 indicate that a mere subset of $\text{Y}$ is included for algorithm fine-tuning in Step 2, and the full vector $\text{Y}$ in Step 3 for the final patient-level phenotype classification.