Privacy-Preserving Deep Learning NLP Models for Cancer Registries

Area of Science:

• Computational pathology
• Medical informatics
• Artificial intelligence in healthcare

Background:

• Population cancer registries process vast amounts of unstructured pathology reports annually.
• Deep learning (DL) requires large labeled datasets for effective model training.
• Privacy and confidentiality concerns hinder data sharing between cancer registries, impeding DL adoption.

Purpose of the Study:

• To propose and evaluate privacy-preserving transfer learning approaches for distributing DL natural language processing (NLP) models among cancer registries.
• To enable the extraction of key cancer characteristics without sharing sensitive patient data.

Main Methods:

• A multitask convolutional neural network (MT-CNN) NLP model was developed to extract six cancer characteristics.
• Privacy-preserving transfer learning techniques were employed for model distribution.
• The proposed methods were compared against conventional transfer learning, single-registry models, and a centralized model using 410,064 pathology documents from two registries.

Main Results:

• Transfer learning approaches, including data sharing and model distribution, significantly outperformed single-registry models.
• The best privacy-preserving model distribution approach achieved performance statistically indistinguishable from the centralized model (micro/macro-F1 scores of 0.823/0.580 vs. 0.827/0.585).

Conclusions:

• Privacy-preserving transfer learning is a viable strategy for distributing DL NLP models across cancer registries.
• This approach effectively addresses data privacy concerns while enabling collaborative model development and improving cancer data extraction accuracy.