DxFormer

DxFormer separates symptom inquiry from disease diagnosis using an inverted Transformer architecture to improve symptom-token representation and sequence-classification-based disease prediction.

Key Features:

• Decoupled Framework: The diagnostic process is divided into separate symptom inquiry and disease diagnosis phases to allow independent optimization of each task.
• Inverted Transformer Architecture: A decoder-encoder Transformer structure is used to learn dense representations of symptoms.
• Symptom Tokenization: Each symptom is represented as a token to enable structured modeling of symptom information.
• Task Formalization: Symptom inquiry is formulated as a language generation task and disease diagnosis as a sequence classification task.
• Parametric Stopping Criteria: The transition from symptom inquiry to disease diagnosis is governed by parametric stopping criteria.
• Joint Optimization: A dual optimization strategy combines reinforcement learning rewards with cross-entropy loss to enhance symptom recall and diagnostic precision.

Scientific Applications:

• Interactive Medical Diagnosis: Supports automated diagnostic workflows that require iterative patient questioning to improve disease prediction accuracy.
• Symptom-Recall Improvement: Enhances recovery of relevant symptoms to increase diagnostic accuracy in clinical decision-support contexts.

Methodology:

Each symptom is tokenized and processed by an inverted Transformer (decoder-encoder) that learns dense symptom representations, formalizes symptom inquiry as language generation and diagnosis as sequence classification, uses parametric stopping criteria to switch phases, employs joint optimization combining reinforcement learning rewards with cross-entropy loss to improve symptom recall, and is experimentally evaluated on three real-world medical dialogue datasets demonstrating improved diagnostic accuracy over prior RL-based methods.