Embedding and Data Representation for Time-Coursed Tabular Healthcare Data

Abstract:
Structured healthcare datasets contain heterogeneous variables whose underlying statistical and clinical characteristics are not always adequately preserved by conventional data representations. This thesis investigates representation-learning methods that account for ordinal structure, temporal dependencies and relationships between variables across psychosocial and clinical datasets.

The first part introduces a Monotonic Ordinal Embedding Framework (MOEF), which learns ordered, non-uniform representations of ordinal variables for depression prediction using the English Longitudinal Study of Ageing. The second part investigates temporal and relational dependencies through an LSTM–graph attention model for early sepsis prediction, alongside network analyses of psychosocial and health factors. The final part develops a two-phase self-supervised Transformer for learning representations from heterogeneous intensive-care data, supporting mortality prediction and patient subgroup analysis.

Together, these studies demonstrate the importance of preserving the underlying structure of tabular healthcare data to improve predictive modelling and the interpretability of learned representations.