Tangle

Tangle applies a time-span-guided neural attention mechanism to sequences of Medicare services to generate interpretable patient representations for predicting progression to second-line diabetes therapies.

Key Features:

• Time-Span-Guided Attention: Incorporates the duration between consecutive Medicare service claims into the attention mechanism to capture temporal patterns in patient trajectories.
• Pre-trained Word Embeddings: Uses pre-trained embeddings such as GloVe to encode sequences of Medicare Benefits Schedule (MBS) codes.
• Interpretable Patient Representations: Integrates sequences of MBS items and their time spans to produce representations intended to be interpretable for clinical prediction tasks.
• Predictive Performance: Evaluated on a 10% sample of de-identified Australian Medicare data (MBS and Pharmaceutical Benefits Scheme [PBS]) for predicting need for second-line diabetes therapies (e.g., when metformin monotherapy is insufficient) with a reported Area Under the ROC Curve (AUC) of 90%.
• Comparison with State-of-the-Art Models: In comparative studies against sequence classification methods including recurrent neural networks (RNNs) and attention-based models, Tangle achieved higher performance (reported AUC 90%).

Scientific Applications:

• Therapy review recommendations: Predicts timing for transition to second-line diabetes therapies to inform automatic therapy review recommendations.
• Proactive disease management: Predicts future therapeutic needs from Medicare service sequences to support earlier clinical intervention decisions.

Methodology:

Implemented in Python using the Keras framework, Tangle combines pre-trained word embeddings (e.g., GloVe) with a time-span-guided neural attention mechanism that incorporates durations between consecutive Medicare claims; evaluation used MBS and PBS claim sequences from a 10% de-identified Australian Medicare sample.