Sequential pattern transformer (SPT): a generative and interpretable framework for predicting disease trajectories

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Machine Learning for Healthcare

Background:

• Integrating AI into clinical workflows necessitates models offering explainable and actionable disease trajectories.
• Opaque deep learning models and noisy electronic health records (EHRs) present significant limitations.
• Existing predictive models often lack the comprehensive, interpretable outputs required for clinical decision-making.

Purpose of the Study:

• To introduce the Sequential Pattern Transformer (SPT), a novel framework for generating explainable and actionable disease trajectories.
• To address the limitations of current AI models in clinical settings by enhancing transparency and interpretability.
• To develop a robust and scalable AI solution for mapping complex disease dynamics.

Main Methods:

• Synergized sequential pattern mining (PrefixSpan) with generative transformer modeling.
• Distilled EHR data from 258,460 type 2 diabetes patients into 95,630 validated disease progression patterns.
• Trained a decoder-only transformer on these patterns to learn temporal disease dynamics, shifting from classification to trajectory generation.

Main Results:

• Achieved 85.78% Top-5 accuracy, significantly outperforming a standard LSTM baseline (71.47%).
• Developed a dynamic Disease Atlas visualizing future patient pathways using explainable AI (XAI) techniques.
• Demonstrated the framework's domain-agnostic nature and efficient fine-tuning capabilities.

Conclusions:

• The Sequential Pattern Transformer (SPT) provides a transparent, robust, and scalable framework for disease trajectory modeling.
• SPT bridges the gap between high-performance AI and interpretable clinical applications.
• The methodology is transferable to diverse clinical conditions and healthcare settings, enhancing AI integration in medicine.