A machine-learning algorithm spotted signs of “covert consciousness” in coma patients—in some cases, days before doctors could do so.