Xjyuk

Usually in logistic regression, we fit a model and get some predictions on the training set. We then cross-validate on those training predictions (something like here) and decide the optimal threshold value based on something like the ROC curve.

Why don't we incorporate cross-validation of the threshold INTO the actual model, and train the whole thing end-to-end?

It isn't because logistic regression isn't a classifier (cf., Why isn't Logistic Regression called Logistic Classification?). It is a model to estimate the parameter, $p$, that governs the behavior of the Bernoulli distribution. That is, you are assuming that the response distribution, conditional on the covariates, is Bernoulli, and so you want to estimate how the parameter that controls that variable changes as a function of the covariates. It is a direct probability model only. Of course, it can be used as a classifier subsequently, and sometimes is in certain contexts, but it is still a probability model.

Regardless of the underlying model, we can work out the sampling distributions of TPR and FPR at a threshold. This implies that we can characterize the variability in TPR and FPR at some threshold, and we can back into a desired error rate trade-off.

A ROC curve is a little bit deceptive because the only thing that you control is the threshold, however the plot displays TPR and FPR, which are functions of the threshold. Moreover, the TPR and FPR are both statistics, so they are subject to the vagaries of random sampling. This implies that if you were to repeat the procedure (say by cross-validation), you could come up with a different FPR and TPR at some specific threshold value.

However, if we can estimate the variability in the TPR and FPR, then repeating the ROC procedure is not necessary. We just pick a threshold such that the endpoints of a confidence interval (with some width) are acceptable. That is, pick the model so that the FPR is plausibly below some researcher-specified maximum, and/or the TPR is plausibly above some researcher-specified minimum. If your model can't attain your targets, you'll have to build a better model.

Of course, what TPR and FPR values are tolerable in your usage will be context-dependent.

For more information, see ROC Curves for Continuous Data
by Wojtek J. Krzanowski and David J. Hand.

It's because the optimal threshold is not only a function of the true positive rate (TPR), the false positive rate (FPR), accuracy or whatever else. The other crucial ingredient is the cost and the payoff of correct and wrong decisions.

If your target is a common cold, your response to a positive test is to prescribe two aspirin, and the cost of a true untreated positive is an unnecessary two days' worth of headaches, then your optimal decision (not classification!) threshold is quite different than if your target is some life-threatening disease, and your decision is (a) some comparatively simple procedure like an appendectomy, or (b) a major intervention like months of chemotherapy! And note that although your target variable may be binary (sick/healthy), your decisions may have more values (send home with two aspirin/run more tests/admit to hospital and watch/operate immediately).

Bottom line: if you know your cost structure and all the different decisions, you can certainly train a decision support system (DSS) directly, which includes a probabilistic classification or prediction. I would, however, strongly argue that discretizing predictions or classifications via thresholds is not the right way to go about this.

See also my answer to the earlier "Classification probability threshold" thread. Or this answer of mine. Or that one.