Table 3.
Performance of ML on ER40 and BLERT datasets compared to performance inferred from t-tests in SCOPE study. (Classification target: SZ vs. HC). For ER401, 120 features (cognition: 40 RT, 40 Corr and meta-cognition: 40 CR) were considered while for BLERT 63 features (cognition: 21 RT, 21 Corr and meta-cognition: 21 CR) were considered as input.
| Dataset | Features | Ranked feature/totalb, c | Count of selected negative emotion featuresb | Count of (CR) featuresb, c | Count of negative CR featuresb | Best performing model | AUC | F1 | Equivalent Cohen's d |
|---|---|---|---|---|---|---|---|---|---|
| ER40 | GINI filtered (RT + CR + Accuracy) |
25/120 | 20 | 12 | 11 | Stacka | 0.81 | 0.78 | 1.24e |
|
t-Test Accuracy |
– | – | – | – | – | 0.63! | – | 0.48d | |
| t-Test Reaction Time |
– | – | – | – | – | 0.59! | – | 0.32d | |
|
t-Test Confidence |
– | – | – | – | – | 0.52! | – | 0.08d | |
| BLERT | GINI filtered (RT + CR + Accuracy) |
33/63 | 26 | 16 | 13 | Neural Network, ReLu | 0.73 | 0.71 | 0.87e |
| t-test Accuracy |
– | – | – | – | – | 0.66! | – | 0.58d | |
| t-test Reaction Time |
– | – | – | – | – | 0.54! | – | 0.16d | |
| t-Test Confidence |
– | – | – | – | – | 0.59! | – | 0.32d |
a
Stack method implies stacking of methods (Naïve Bayes, Neural Network (ReLU), Random Forest, Tree).
b
Negative features considered for ER40 are S, A and F while negative features considered for BLERT are S, A, F, D and SU.
c
Positive features considered for ER40 and BLERT are N, H.
d
Table 8 from (Pinkham et al., 2018a).
e
Table 2 from (Salgado, 2018).