Machine-learning improves the prediction of stroke recovery

For one thing, the figure that the linked article presents is very misleading. It is not the performance of the model! It's just the raw data. Patients by how well we think they might recover vs. how well they discovered. The goal of the model is to figure out how will recover well. Literally, nothing in that figure is an output of the model.

> In case of stroke lesions affecting the performance of parcellation, the correspondence part of the lesion on the unaffected hemisphere was used and interpolated over the lesioned brain as a brain transplant and fed into the above-described structural imaging processing.

The classifier will trivially discover that you are doing this, and it's almost certainly correlated with terrible outcomes. (PS: should be "corresponding" part) This alone calls into question all of the results and never should have gotten past anyone. Just because you can't tell as a human, doesn't mean the data manipulation isn't immediately obvious to even a simple model like an SVM.

> By means of k-means clustering, patients were identified as fitters, i.e. revealing a proportional recovery, and non-fitters, i.e. patients lacking proportional recovery.

Good god, why? Just set up a simple criteria and use it. This kind of strange arbitrary boundary in your data that comes from k-means can only hurt your model.

Finally: The results are terrible. When cross-validating the model does well, 83% accuracy (87% precision). But on the external data the model is worthless. 60% accuracy, 53% precision.

They provide no trivial baselines. They don't even list of the base rates for their dataset (I'm guessing that out of their main dataset the classifier is just operating at chance at 60% accuracy / 53% precision is just the base rate).

Their data seems to be horrific. Half the patients in the 2nd dataset don't have data for both time points. How can they be assigned to groups that fit or don't fit without this basic data? They don't even describe results on their 3rd dataset.

They don't bother to compute a confidence interval, which will be huge because they have so little data.

Seriously. Authors. If you are reading this, reach out to a machine learning person. If you don't know anyone, reach out to me, I'm a scientist at a major US university. Papers like this just pollute our science and are a total waste of time on your part.