What does a learning curve tell you about whether more data would help?

Why

Collecting more samples in biology is expensive, so it is worth asking before you ask for budget. The scikit-learn guide describes the learning curve as showing training and validation scores for increasing numbers of training samples, to find out how much more data would help and whether error comes mainly from variance or bias. Its example shows a naive Bayes model whose two curves converge at a low score, so more data would not help, and an SVM whose training score stays well above its validation score, so more data probably would.

Read the shape, not one point. Suppose a classifier trained on 50, 100, 200 and 400 patients reaches validation AUC of 0.66, 0.71, 0.74 and 0.76 while training AUC sits near 0.95. The gap is large and validation is still rising, so 800 patients might plausibly reach about 0.78, though the gains are shrinking.

If instead both curves had flattened at 0.70, a richer feature set or a more flexible model would be the next step, not more patients. Each point should be averaged over several splits with its spread shown, because small subsets are noisy. A learning curve also assumes new data come from the same population.

Adding patients from a new site is a different question about transfer.