Sequential Pattern Transformer (SPT)

A generative and interpretable framework for predicting disease trajectories using transformer architectures.

Overview

The Sequential Pattern Transformer (SPT) is a novel deep learning framework designed to predict disease trajectories by modeling temporal patterns in patient health records. Unlike traditional approaches that treat health events as static features, SPT captures the sequential, time-dependent nature of disease progression.

The Problem

Predicting how a patient’s health will evolve over time is one of the most challenging problems in clinical informatics. Traditional machine learning models often flatten temporal data into static feature vectors, losing critical information about the order and timing of clinical events. This limits their ability to forecast future diagnoses, complications, and care needs.

Our Approach

SPT leverages a transformer architecture — the same family of models behind modern large language models — but adapts it specifically for clinical event sequences:

  • Sequential Pattern Mining — Extracts frequent temporal patterns from electronic health records to create a structured vocabulary of disease progressions
  • Generative Prediction — Produces probabilistic forecasts of future diagnoses, enabling clinicians to anticipate emerging conditions
  • Interpretable Attention — Attention weights reveal which prior events most influence each prediction, providing built-in explainability

Key Features

Feature Description
Temporal Encoding Time-aware positional embeddings that capture irregular intervals between clinical events
Multi-horizon Forecasting Predictions at multiple future time points (30-day, 90-day, 1-year)
Interpretability Attention-based explanations showing which historical events drive each prediction
Scalability Validated on datasets with millions of patient records

Technical Architecture

The framework consists of three stages:

  1. Pattern Discovery — Sequential pattern mining identifies common disease progression pathways from large-scale EHR data
  2. Sequence Encoding — Patient histories are encoded as ordered sequences of clinical events with temporal embeddings
  3. Transformer Prediction — A modified transformer decoder generates probability distributions over future diagnoses

Impact

SPT has been validated on real-world datasets and demonstrates superior performance compared to traditional approaches (LSTM, GRU) for disease trajectory prediction. The framework’s interpretable design makes it suitable for clinical deployment where understanding the “why” behind predictions is essential for trust and adoption.

Published in: Neural Computing and Applications (2025) Read the paper →