Fig. 3: The tradeoff between model discrimination and complexity.
From: A flexible symbolic regression method for constructing interpretable clinical prediction models
Each point shows the cross-validation testing AUPRC (y-axis) and size (x-axis) for models trained in 50 repeat trials for each method. Each subplot represents a different expert-curated heuristic (top row) or chart review phenotype (bottom row). The ideal model is discriminative and simple, meaning it is near the top left corner. DT models occupy this zone for modeling simple heuristics, whereas FEAT models tend to be simple and accurate across all experiments.
