Table 1 Measures for the performance of the Twitter supervised models (national and subnational) and the Google model (national) in estimating the overall intensity, onset and peaks of the 2015/16 influenza season when compared to RCGP ILI rates (national and subnational).
From: The added value of online user-generated content in traditional methods for influenza surveillance
| Â | Subnational level | National level | ||||
|---|---|---|---|---|---|---|
London Twitter | Midlands and East Twitter | North Twitter | South Twitter | England Twitter | England Google | |
Overall intensity | ||||||
r | 0.37 | 0.37 | 0.40 | 0.31 | 0.67 | 0.96 |
MSE | 69.46 | 70.06 | 71.72 | 131.5 | 36.57 | 3.86 |
RMSE | 8.33 | 8.37 | 8.47 | 11.47 | 6.05 | 1.96 |
MAE | 6.28 | 5.95 | 6.13 | 7.83 | 4.27 | 1.47 |
MAPE | 39.10% | 45.55% | 40.85% | 45.31% | 29.95% | 14.10% |
ME | −2.58 | −5.43 | −5.13 | −6.99 | −2.17 | 0.54 |
Max Error | 18.05 | 19.95 | 23.42 | 32.81 | 13.88 | 4.32 |
Week Max Error | 47 | 13 | 11 | 12 | 12 | 52 |
Max Percentage Error | 214.88% | 84.53% | 83.05% | 207.50% | 75.72% | 66.12% |
Week of Max Percentage Error | 47 | 13 | 11 | 46 | 47 | 52 |
Onset | ||||||
Alert week | 1 | 47 | 51 | 3 | 2 | 1 |
Time difference | 2 | −2 | 0 | −1 | −1 | 0 |
Peaks | ||||||
Magnitude of 1st peak-to-peak difference | −0.55 (2.04%) | −3.57 (17.76%) | 3.93 (18.36%) | −2.92 (10.25%) | 1.32 (6.03%) | 0.15 (0.68%) |
Temporal offset of 1st peaks | 12 | 2 | 4 | 2 | 4 | 2 |
Magnitude of 1st peak-to-model difference (same week as RCGP estimate) | −4.79 (17.74%) | −15.63 (77.76%) | −13.35 (62.38%) | −18.82 (66.04%) | −2.12 (9.68%) | −0.31 (1.42%) |
Magnitude of 2nd peak-to-peak difference | 6.05 (21.61%) | −11.86 (48.61%) | −7.81 (27.70%) | −25.24 (66.07%) | −4.28 (14.91%) | −3.16 (11.01%) |
Temporal offset of 2nd peaks | 6 | 5 | 4 | 2 | 4 | 1 |
Magnitude of 2nd peak-to-model difference (same week as RCGP estimate) | −15.03 (53.68%) | −19.81 (81.19%) | −23.42 (83.05%) | −32.81 (85.89%) | −11.87 (41.36%) | −3.86 (13.45%) |
