PE-Tab2Conv: Clinically structured convolutional representation learning for preeclampsia risk modelling from routine electronic health records

Abstract Background: Machine-learning decision support from routine electronic health records requires representations that handle class imbalance while retaining clinically interpretable structure. We developed PE-Tab2Conv, a clinically structured tabular-to-matrix representation for preeclampsia risk modelling.

Methods

Encoded maternal variables were arranged in a fixed $4 \times 6$ matrix spanning baseline and socioeconomic characteristics, obstetric history, extended obstetric and family history, and comorbidities. A lightweight two-dimensional convolutional neural network produced 32-dimensional embeddings for downstream classifiers. Models were evaluated on a locked 2023 temporal cohort from RS Adhyatma ($n=531$), comparing a portable core feature space (17 predictors) with an extended space (21 predictors). Interpretability used SHapley Additive exPlanations and matrix-cell occlusion. A second hospital contributed 72 preeclampsia-positive records for positive-case transportability assessment.