A few-shot disease diagnosis decision making model based on meta-learning for general practice

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Machine Learning

Background:

• Diagnostic errors pose a significant risk to patient safety in primary healthcare.
• General practitioners face challenges diagnosing rare diseases due to limited knowledge and clinical experience.
• Clinical decision-making tools are crucial for improving diagnostic accuracy in primary care, especially for conditions with limited data.

Purpose of the Study:

• To propose a novel few-shot disease diagnosis decision-making model (FSDD-MAML) utilizing model-agnostic meta-learning.
• To enhance the diagnostic capabilities of general practitioners, particularly for rare or underrepresented diseases.
• To address the challenges posed by long-tailed class distributions in medical datasets for deep learning models.

Main Methods:

• A knowledge graph-based disease diagnosis model is enhanced with the model-agnostic meta-learning (MAML) algorithm.
• FSDD-MAML optimizes model parameters and learns optimal learning rates for all modules.
• The model employs inner and outer loops for gradient updates and meta-objective optimization in n-way, k-shot learning tasks.

Main Results:

• The FSDD-MAML model achieved a precision@1 of 90.02%, outperforming existing models.
• Significant improvements were observed in the prediction performance for few-shot diseases, with relative precision@1 increases of 29.13% and 21.63% for the rarest disease groups.
• The study demonstrates state-of-the-art performance in disease diagnosis, particularly for conditions with limited data.

Conclusions:

• The proposed meta-learning based decision-making model supports rapid disease diagnosis in general practice.
• FSDD-MAML is particularly effective in assisting general practitioners with diagnosing few-shot diseases.
• This research highlights the significance of applying meta-learning to few-shot disease assessment in primary care settings.