Predicted-by-Observed Charts

Figure 1. Predicted-by-observed chart for Length of stay

For scale-dependent variables, the predicted-by-observed chart displays a scatterplot of predicted values on the y axis by observed values on the x axis for the combined training and testing samples. Ideally, values should lie roughly along a 45-degree line starting at the origin. The points in this plot form vertical lines at each observed number of days of Length of stay.

Looking at the plot, it appears that the network does a reasonably good job of predicting Length of stay. The general trend of the plot is off the ideal 45-degree line in the sense that predictions for observed lengths of stay under five days tend to overestimate the length of stay, while predictions for observed lengths of stay beyond six days tend to underestimate the length of stay.

The cluster of patients in the lower left part of the plot are likely to be patients who did not undergo surgery. There is also a cluster of patients in the upper left part of the plot, where the observed length of stay is one to three days and the predicted values are much greater. It is likely that these cases are patients who died in the hospital post-surgery.

Figure 2. Predicted-by-observed chart for Treatment costs

The network also appears to be reasonably good at predicting Treatment costs. There appear to be three primary clusters of patients:

In the lower left are primarily patients who did not have surgery. Their costs are relatively low and are differentiated by the type of Clot-dissolving drugs [clotsolv] administered in the emergency room.
The next cluster of patients have treatment costs of about $30,000. These are patients who had percutaneous transluminal coronary angioplasty (PTCA).
The final cluster have treatment costs in excess of $40,000. These are patients who had coronary artery bypass surgery (CABG). This surgery is somewhat more expensive than PTCA, and patients have a longer hospital recovery time, which increases costs further.

There are also a number of cases with costs in excess of $50,000 that the network does not predict very well. These are patients who experienced complications during surgery, which can increase the cost of surgery and length of stay.