Table 2.
Results of K = 100 cross-validation experiments with the HDRS total score (left columns) and the YMRS total score (right columns) models based on all, mandatory and mood self-assessment items, respectively.
| HDRS total score | YMRS total score | |||
|---|---|---|---|---|
| Model | R2 (SD) ↑a | RMSE (SD) ↓b | R2 (SD) ↑a | RMSE (SD) ↓b |
| All self-assessment items | ||||
| Pooled naïve mean | −0.02 (0.03) | 5.99 (0.37) | −0.04 (0.05) | 4.18 (0.70) |
| Pooled Ridge | 0.37 (0.10) | 4.68 (0.48) | 0.02 (0.15) | 4.03 (0.60) |
| Pooled XGBoost | 0.44 (0.10) | 4.40 (0.41) | −0.04 (0.21) | 4.11 (0.53) |
| Pooled Bayesian | 0.36 (0.12) | 4.72 (0.51) | 0.00 (0.21) | 4.04 (0.56) |
| Separate naïve mean | 0.47 (0.11) | 4.29 (0.47) | −0.00 (0.33) | 4.00 (0.53) |
| Separate Ridge | 0.47 (0.12) | 4.30 (0.49) | 0.04 (0.30) | 3.92 (0.54) |
| Separate XGBoost | 0.27 (0.15) | 5.03 (0.49) | −0.38 (0.50) | 4.64 (0.45) |
| Hierarchical Bayesian | 0.57 (0.10) | 3.85 (0.47) | 0.12 (0.31) | 3.74 (0.46) |
| Mandatory self-assessment items | ||||
| Pooled naïve mean | −0.02 (0.03) | 5.94 (0.37) | −0.04 (0.06) | 4.25 (0.71) |
| Pooled Ridge | 0.21 (0.07) | 5.24 (0.34) | 0.01 (0.09) | 4.12 (0.65) |
| Pooled XGBoost | 0.37 (0.12) | 4.63 (0.39) | −0.06 (0.18) | 4.23 (0.57) |
| Pooled Bayesian | 0.21 (0.10) | 5.22 (0.37) | 0.03 (0.13) | 4.08 (0.61) |
| Separate naïve mean | 0.46 (0.16) | 4.28 (0.57) | −0.01 (0.30) | 4.08 (0.54) |
| Separate Ridge | 0.46 (0.16) | 4.29 (0.57) | 0.00 (0.29) | 4.06 (0.54) |
| Separate XGBoost | 0.25 (0.18) | 5.06 (0.54) | −0.34 (0.39) | 4.68 (0.42) |
| Hierarchical Bayesian | 0.54 (0.13) | 3.94 (0.53) | 0.10 (0.27) | 3.85 (0.49) |
| Mood self-assessment item | ||||
| Pooled naïve mean | −0.02 (0.02) | 5.91 (0.41) | −0.05 (0.05) | 4.20 (0.77) |
| Pooled Ridge | 0.21 (0.06) | 5.19 (0.35) | 0.02 (0.07) | 4.05 (0.70) |
| Pooled XGBoost | 0.34 (0.11) | 4.75 (0.35) | 0.01 (0.18) | 4.03 (0.54) |
| Pooled Bayesian | 0.20 (0.12) | 5.23 (0.45) | 0.04 (0.12) | 4.00 (0.63) |
| Separate naïve mean | 0.44 (0.15) | 4.31 (0.47) | 0.02 (0.27) | 3.98 (0.59) |
| Separate Ridge | 0.45 (0.15) | 4.29 (0.48) | 0.03 (0.27) | 3.96 (0.59) |
| Separate XGBoost | 0.42 (0.15) | 4.42 (0.42) | −0.04 (0.34) | 4.05 (0.51) |
| Hierarchical Bayesian | 0.51 (0.14) | 4.05 (0.45) | 0.16 (0.25) | 3.68 (0.54) |
The hierarchical Bayesian model achieved the best overall performance in every case and could predict the clinical severity ratings within 4 points of RMSE on the original rating scales. The best HDRS total result was achieved using all self-assessment items while the best YMRS total result was achieved using only the mood self-assessment item.
Bold values indicates the best results within each set of self-assessment items.
aCoefficient of determination. Higher is better.
bRoot Mean Square Error. Lower is better.