Table 3.
Average runtime (in seconds) of the different predictive modeling pipeline task types for the three different EHR data sets.
| Task Type | Runtime (seconds) | ||
|---|---|---|---|
| Small | Medium | Large | |
| Cohort Construction | 6.14 | 6.53 | 9.32 |
| Feature Construction (Diagnosis) | 15.77 | 275.05 | 266.64 |
| Feature Construction (Lab) | 27.94 | 551.49 | 104.29 |
| Feature Construction (Medication) | 26.43 | 203.75 | 131.47 |
| Feature Construction (Procedure) | 13.93 | -- | 199.25 |
| Feature Construction (Symptoms) | -- | 88.33 | -- |
| Feature Construction (Merge) | 1.32 | 12.16 | 16.28 |
| Cross-validation | 7.00 | 42.37 | 248.41 |
| Feature Selection (Fisher Score) | 5.47 | 65.87 | 72.64 |
| Feature Selection (Information Gain) | 8.50 | 103.95 | 370.83 |
| Classification (K-Nearest Neighbor) | 0.16 | 64.02 | 1651.12 |
| Classification (Naïve Bayes) | 0.20 | 7.74 | 34.25 |
| Classification (Logistic Regression) | 0.18 | 24.56 | 342.65 |
| Classification (Random Forest) | 1.21 | 190.66 | 1649.13 |