The training data was fed into the neural network, so that it could spot patterns and make recommendations for new patients based on their records.
The problem with this approach is that human illnesses, accidents etc may not have consistent patterns, they can be completely random. Two patients with the same gross symptoms and history may have completely different outcomes. That's why strict protocol based medicine (the patient has this, so we will do that) has proven to be ineffectual compared to more traditional methods.
Whilst not quite the same, I have first-hand experience of this misguided attempt to predict future outcomes based on historical data. In the mid-nineties I was working as a paramedic with a UK ambulance service.
A local university research team came up with the idea that they could improve ambulance response times by predicting where emergency calls would occur, based on supposed patterns in historical data of previous years. The idea was that ambulances would be sent to loiter in the predicted area, so that they would be closer when a call came in.
Instead of trialling this with any simulation, the ambulance service decided to do it live, so they started using the software to position the ambulances around the county.
It soon became apparent that except for a few isolated occasions where it guessed correctly, the overall impact was that the ambulances were nearly always in the wrong place, and that response times were worsening, not improving.
The trial was abandoned early, despite the university's insistence that their idea would work if it was given more time.
If you take a step back you would realise that the idea that previous years' data of emergency calls would have any bearing on future occurrences is unlikely at best, but the university research group were convinced that it was a reasonable assumption.
In the same way, trying to predict how a patient will react to treatment, based on patterns of historical data of previous patients, is a flawed idea.