Table 2. Modeling Performance Metrics.
| Modela | AUC, mean (95% CI) | Cut point | Sensitivity | Specificity | PPV, % | NPV, % | BAC | PSI |
|---|---|---|---|---|---|---|---|---|
| Conv-LSTM model | ||||||||
| Held-out test set (activity only) | 0.66 (0.64-0.68) | 0.48 | 0.79 | 0.50 | 5.77 | 98.43 | 0.65 | 10.02 |
| Cross-validation sets (activity only) | 0.67 (0.64-0.71) | 0.49 | 0.68 | 0.60 | 6.90 | 98.10 | 0.64 | 10.03 |
| Conv-LSTM model with depression scores | ||||||||
| Held-out test set (PHQ-9 plus activity) | 0.66 (0.65-0.67) | 0.50 | 0.74 | 0.56 | 6.13 | 98.26 | 0.65 | 10.10 |
| Cross-validation sets (PHQ-9 plus activity) | 0.70 (0.65-0.75) | 0.50 | 0.73 | 0.60 | 7.10 | 98.35 | 0.66 | 10.20 |
| Wavelet LogReg model | ||||||||
| Held-out test set | 0.64 (0.64-0.64) | 0.08 | 0.72 | 0.56 | 5.95 | 98.15 | 0.64 | 3.40 |
| Cross-validation sets | 0.65 (0.60-0.71) | 0.06 | 0.72 | 0.54 | 6.05 | 98.22 | 0.63 | 3.41 |
Abbreviations: AUC, area under the receiver operating characteristic curve; BAC, balanced accuracy; Conv-LSTM, convolutional–long short-term memory; LogReg, logistic regression; NPV, negative predictive value; PHQ-9, 9-item Patient Health Questionnaire; PPV, positive predictive value; PSI, population stability index.
Given the small size of our selective serotonin reuptake inhibitor group, we present these metrics for each model (Conv-LSTM run with movement data alone, Conv-LSTM run with movement data and depression scores, and logistic regression operating on wavelet-derived features) to ensure comprehensive report of model performance. Sensitivity, specificity, PPV, NPV, BAC, and PSI are mean values across 10 cross-validation sets or across distinct model runs of held-out test set; AUC remains the primary outcome metric, which is discussed in the Results and Discussion sections.