Implement AUROC metric for generic cardiac records stratification task

didactic/tasks/cardiac_records_stratification.py

@@ -1,6 +1,6 @@
 import itertools
 import logging
-from typing import Sequence, Tuple
+from typing import Dict, Sequence, Tuple
 
 import hydra
 import numpy as np
@@ -9,6 +9,7 @@
 from omegaconf import DictConfig
 from pytorch_lightning.trainer.states import TrainerFn
 from sklearn.base import ClassifierMixin
+from sklearn.metrics import accuracy_score, roc_auc_score
 from vital.data.cardinal.config import TabularAttribute
 from vital.data.cardinal.data_module import CardinalDataModule
 from vital.data.cardinal.datapipes import MISSING_CAT_ATTR
@@ -31,6 +32,10 @@ def __init__(
             tabular_attrs: List of tabular attributes to use as input features for the classifier.
             target_attr: Tabular attribute to use as the target label for the classifier.
         """
+        for method in ["fit", "predict_proba", "save_model"]:
+            if not callable(getattr(model, method, None)):
+                raise ValueError(f"Model must implement method: {method}")
+
         self.model = model
 
         # Ensure string tags are converted to their appropriate enum types
@@ -102,20 +107,26 @@ def fit(self, data: CardinalDataModule) -> "CardiacRecordsStratificationTask":
 
         return self
 
-    def score(self, data: CardinalDataModule) -> float:
+    def score(self, data: CardinalDataModule) -> Dict[str, float]:
         """Measure the model's performance on the test set.
 
         Args:
             data: CARDINAL data module.
 
         Returns:
-            Model's accuracy on the test set.
+            Dictionary of model's metrics (e.g. accuracy, AUROC, etc.) on the test set.
         """
         # Extract the tabular data (i.e. inputs and target labels) from the test set
         X, y = self._prepare_data_subset(data, "test")
 
-        # Compute the model's performance on the test set
-        return self.model.score(X, y)
+        # Perform inference on the test set, keeping intermediate predictions (i.e. class probabilities)
+        predict_proba = self.model.predict_proba(X)
+        y_hat = np.argmax(predict_proba, axis=1)
+
+        # Compute the model's performance metrics
+        scores = {"acc": accuracy_score(y, y_hat), "auroc": roc_auc_score(y, predict_proba, multi_class="ovr")}
+
+        return scores
 
     def save(self, path: str) -> None:
         """Save the model to disk.
@@ -145,18 +156,18 @@ def main(cfg: DictConfig):
     task.fit(data)
 
     # Evaluate the model's performance on the test set
-    score = task.score(data)
+    scores = task.score(data)
 
     hydra_output_dir = Path(HydraConfig.get().runtime.output_dir)
 
     # Save the model
     task.save(hydra_output_dir / cfg.model_ckpt)
 
     # Save the model's performance
-    score_df = pd.Series({"acc": score})
+    score_df = pd.Series(scores)
     score_df.to_csv(hydra_output_dir / cfg.scores_filename, header=["value"])
 
-    logger.info(f"Logging model and its score ({score:.2%}) to {hydra_output_dir}")
+    logger.info(f"Logging model and its scores: {scores} to {hydra_output_dir}")
 
 
 if __name__ == "__main__":