Table 3 Methods of addressing temporality in time-series data
Level of addressing temporality | Method of addressing temporality | Description of the method | Sample size | Frequency |
|---|---|---|---|---|
Addressed by the model | Recurrent Neural Network (RNN) | Data transformed to 3D array and fed in the LSTM model [26] Events encoded by Adaptive Net, pooled by LSTM model [27] | 823 9,500 | Weekly Irregulara |
RNN with GRU, considering each treatment as timestep [28] | 105,129 | Irregulara | ||
Addressed in the features | Measured change | Change of measurement from baseline [84] Change in mean measurements from baseline [88] Mean daily change from the 24-h baseline period [87] Change in symptom severity from previous report [78] | 245 31,700 116 34 | Every 90 days Daily Twice a day Weekly |
Binary outcome | Variable indicated if a report is followed by exacerbation event [80] Occurrence of symptom in any day of a time window [85] | 2,374 182,991 | Daily Daily (3 days) | |
Dichotomised score one week following the prediction date [104] | 210 | Weekly | ||
Feature for each timeline | Score added as an input feature at every measurement [82] Created a timeline of best overall responses (BORs) [103] | 83 31 | Weekly Weekly | |
Not considered | Model for each timeline | Treated the 2- and 8-week measures as if assessed at baseline [81] Three models that used 7, 14, and 21 days as inputs [116] | 1,003 20 | 3 time-points 3 time-points |
Selected 1 value for analysis | If patient had multiple follow-up events, the first was chosen [77] The assessment with the highest overall score was used [79] | 494 11,761 | Irregulara Bi-weekly | |
Score was updated at each assessment [86] | 212,615 | Irregulara |