Table 2.
Head-to-head performance of 6 analytical models across 500 splits (number)
|
CSMF accuracy |
Chance-corrected concordance |
|||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Tariff | SSP | RF | PCVA | King-Lu | InterVA | Tariff | SSP | RF | PCVA | InterVA | ||
|
Adult |
No HCE |
162 |
156 |
150 |
8 |
13 |
11 |
5 |
493 |
2 |
0 |
0 |
|
HCE |
156 |
142 |
189 |
10 |
0 |
3 |
4 |
495 |
1 |
0 |
0 |
|
|
Child |
No HCE |
232 |
166 |
68 |
15 |
18 |
1 |
169 |
195 |
136 |
0 |
0 |
|
HCE |
264 |
141 |
69 |
21 |
5 |
0 |
236 |
191 |
55 |
18 |
0 |
|
| Neonate |
No HCE |
203 |
44 |
44 |
24 |
163 |
22 |
46 |
300 |
154 |
0 |
0 |
| HCE | 201 | 50 | 62 | 30 | 138 | 19 | 47 | 254 | 199 | 0 | 0 | |
Table 2 gives the number of 500 test-train datasets for which each method performs best for chance-corrected concordance (CCC) and cause-specific mortality fraction (CSMF) accuracy, by age and health care experience (HCE). King-Lu (KL) does not estimate individual causes so chance-corrected concordance cannot be calculated. PCVA, physician-certified VA; RF, Random Forest; SSP, Simplified Symptom Pattern; VA, verbal autopsy.