a) Physician Certified VA Confusion Matrix | b) Random Allocation Confusion Matrix |
|---|
| | Predicted | | Predicted |
|---|
Stroke | Diabetes | Other | Stroke | Diabetes | Other |
|---|
True | Stroke | 123 | 18 | 125 | True | Stroke | 87 | 84 | 95 |
| | Diabetes | 5 | 86 | 55 | | Diabetes | 32 | 61 | 53 |
| | Other | 83 | 95 | 2112 | | Other | 746 | 780 | 764 |
- Panel A shows the confusion matrix for physician certified verbal autopsy (with a length-three cause list for clarity). The entry in each cell counts the number of deaths truly due to the row cause that were predicted to be due to the column cause. For example, the value 83 in the “other” row, “stroke” column indicates that 83 deaths truly due to causes other than stroke or diabetes were (incorrectly) attributed to stroke by physicians. This table demonstrates that (for this dataset) physicians are right more often than they are wrong when they predict stroke as the cause of death, but wrong more than they are right when they predict diabetes. Panel B shows the confusion matrix for Random Allocation with the same dataset, where random chance predicts stroke and diabetes incorrectly for a vast majority of the cases. True and PCVA data from Lozano et al. [18, 22], where physicians were presented with VAI data where the underlying cause was known to meet stringent clinical diagnostic criteria, and their results compared to the truth
