An alternative way to throw money at the problem: Instead of trying to further improve accuracy, build out space for more ER beds, and implement continuous monitoring of marginal patients.

Or, build devices to send home with patients which allow for cheap, continuous self-monitoring. That might be a legitimate application of AI actually, if you could use e.g. phone camera tricks to measure more health parameters. Even if imperfect, it could still pick up a few patients who should not have been sent home.