With severely unbalanced classes, overall accuracy is not a good metric. So we use recall of the symptom class, and how it varies with the precision. Recall = 0

With severely unbalanced classes, overall accuracy is not a good metric. So we use recall of the symptom class, and how it varies with the precision. Recall = 0.6 to 0.9. All results are based on cross-validation. There is a lot more sophistication we can add (e.g. markov properties), these are almost first-order results. Nonetheless important to validate with other populations. That would be a big part of testing any product.