A Few-Shot Learning-Based Framework for Predicting Likelihood of In-patient's Mortality.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Data Science

Background:

• Electronic Health Records (EHR) digitization offers vast data for healthcare enhancement.
• Challenges in EHR data analysis include high dimensionality, sparsity, imbalance, and unstructured features.
• Predicting in-patient mortality and length of stay is crucial for improving healthcare delivery.

Purpose of the Study:

• To develop a deep learning framework for predicting in-patient mortality and length of stay using EHR data.
• To address the challenges of high dimensionality, sparsity, imbalance, and unstructured features in EHR data.
• To enhance model adaptability through few-shot learning capabilities.

Main Methods:

• A deep multitask and few-shot learning framework was proposed.
• A mixed-expert neural network was adopted for multitask learning.
• The network architecture served as an encoder for few-shot learning.

Main Results:

• The proposed framework achieved an accuracy of 72.41%.
• The framework demonstrated a recall of 51.75%, significantly outperforming SVM.
• An F1-score of 60.40% was achieved, surpassing other models.

Conclusions:

• The deep learning framework shows promise in enhancing healthcare delivery by effectively utilizing EHR data.
• The model's superior recall and F1-score indicate its potential for improving prediction tasks despite lower accuracy compared to SVM.
• The framework's adaptability through few-shot learning is a key advantage for handling diverse healthcare scenarios.