When training neural networks (NNs) on time-series inpatient data, as the number of outcomes predicted diversifies, the NN both generalizes better on external validation and reaches higher performance in similar numbers of training epochs. We demonstrated this in the context of predicting decompensation in Acute Respiratory Distress Syndrome (ARDS). The NN outperformed gradient boosted trees, achieving an area under the receiver operating characteristic of 0.86 on an external hold out test set of hospitals not included in the training set. We estimated real world benefit by comparing mortality rates between similarly at risk patients when diagnosed before or after the time of algorithm evaluation. We showed that, among similarly at risk patients, earlier diagnosis of ARDS nearly doubles the rate of in-hospital survival. Using cluster analysis of the algorithm’s internal representations, we identified distinct ARDS sub-groups, some of which had similar mortality rates but different clinical presentations.