Table 2.
Challenge participating system method comparison
| Team (Continent Organization) | Main algorithm | Temporal and cross-sectional modeling consideration | Pre-filling |
|---|---|---|---|
| Ping An [14] (A, I) | LightGBM | Multi-directional temporal and cross-sectional features | |
| AstraZeneca [16] (E, P) | MICE | GP estimated trend features and summary statistics to augment cross-sectional features for MICE | |
| Vanderbilt [17] (N, U) | XGBoost | Measurements from concurrent and pre- and post-three time-stamps | Global mean, local mean, SVD and Soft-Impute |
| HKBU [20] (A, U) | Fusion layer to combine RNN and MLP outputs | Temporal features for RNN, cross-sectional features for MLP | Temporal decay |
| Padova [21] (E, U) | Weighted average of KNN and linear interpolation outputs | Temporal features for linear interpolation, cross-sectional features for KNN | |
| TSU [22] (N, U) | Piecewise linear interpolation and non-linear extension of MICE | Temporal features for linear interpolation, cross-sectional features for non-linear extension of MICE | Linear interpolation for tests with over 0.5 correlation with time |
| DLUT [23] (A, U) | KNN, Soft-Impute and nuclear norm matrix factorization | No explicit modeling for temporal trends | |
| NCSU [24] (N, U) | Matrix factorization methods | Regularization term for modeling temporal locality | |
| Drexel [25] (N, U) | Similarity weighting | Time window based similarity to capture temporal locality | |
| Buffalo/Virginia [26] (N, U) | Fusion gate to combine RNN outputs | Separate RNNs for temporal features of each test and for cross-sectional features | |
| IBM [27] (N, T) | XGBoost as the ensemble method | Base models: linear model for temporal imputation; KNN, RF, MLP for cross-sectional imputation; and bi-directional GRU and LSTM for combined imputation | |
| Iowa [28] (N, U) | Ridge regression, LASSO or gradient boosting as the ensemble method | Base models: spline basis functions for temporal imputation; RBF neural network for cross-sectional imputation; and bi-directional LSTM for combined imputation | RBF interpolation |
Continent: A – Asia, N – North America, E – Europe. Organization: I – Insurance, P – Pharmaceutical, U – University, T – Technology.