Fig. 4.

Fig 4. compares graphically three methods of outcome prediction analysis in infants having MRI scans. (data from Table 1). The whole cohort (n = 168) has the darkest colour grade fading towards the smallest cohort n = 146 where infants with postnatal collapse and/or additional diagnosis to HIE were excluded.

The 146 cohort would fulfill the original cooling entry-criteria in the CoolCap and TOBY trials. The upper panel shows the positive predictive value (PPV) for adverse outcome. The first 4 shaded bars show results from binary logistic regression from the best model allowing all six MRI and all clinical and biochemical variables. WMxBGT is the strongest MRI variable. For the n = 146-group (palest colour), the best PPV from logistic regression is 95%. In the second vertical set of bars, WMxBGT is removed from the allowed variables and total injury score (TIS) is now the most significant. Again the 146 group has the best prediction, now 90%. The next two vertical sets of bars use cross-tabulation analysis with the best cut-off for a single MRI variable, either WMxBGT or TIS. The sequence of results show that logistic regression is better than cross-tabulation and that WMxBGT is better than TIS for outcome prediction. The middle horizontal panel shows that the negative predictive value (NPV) for poor outcome is good, 90–93% between all groups and methods. The lowest horizontal panel shows the predictive accuracy (PA, the sum of all correct predictions, both adverse and favourable) compared to the whole group. Again, there is little difference between methods. In a dataset with 75% favourable outcome, it is the PPV for adverse outcome that is the most important predictor.