AI Department Makes Strides in Predicting COVID-19 Prognosis with New Research

Researchers from the AI department in collaboration with several universities (such as Geneva University, University of British Columbia, Isfahan University of Medical Sciences, etc.) developed a deep learning model to predict COVID-19 outcomes using chest CT scans. This model uses federated learning to protect patient privacy and achieves comparable accuracy.

Dec 06-2023

15 min to read

The new year began with a beacon of hope in the fight against COVID-19, as the AI department unveiled a groundbreaking research paper utilizing deep learning technology to predict patient outcomes. This study, published in [Journal Name], tackled a crucial challenge hindering wider clinical adoption of deep learning models for COVID-19 prognostication: data privacy concerns.

Researchers developed a deep learning model that analyzes chest CT scans to predict patient outcomes. However, traditional approaches often require centralizing large amounts of sensitive patient data, raising privacy concerns. This research addressed this issue by employing a novel technique called deep privacy-preserving federated learning (DPFL).
With DPFL, data remains securely stored at individual participating institutions, like hospitals. Instead of sharing the data itself, only model updates are exchanged, allowing for collaborative training without compromising patient privacy. This approach also protects against membership inference attacks, where attackers attempt to identify which institutions contributed to the model.

Achieving Accuracy While Protecting Privacy:
The study involved over 3,000 patients from 19 centers, their chest CT images analyzed using a specific DensNet deep learning model. This model was trained in two ways:
• Centralized: All data is gathered in one location for training, the traditional approach.
• DPFL: Data stays at each institution, with only model updates shared, ensuring patient privacy.

Both models demonstrated remarkable accuracy in predicting patient outcomes. The centralized model reached an accuracy of 76%, while the DPFL model closely followed with 75%. Notably, both models exhibited high specificity (correctly identifying healthy patients) and sensitivity (identifying critically ill patients).
Crucially, statistically significant differences were not found between the two models. This means the DPFL approach achieved performance comparable to the centralized approach while protecting patient privacy. Additionally, the model proved resistant to membership inference attacks, further solidifying its privacy-preserving capabilities.
Revolutionizing Personalized Medicine:
This research marks a significant breakthrough in utilizing AI for personalized medicine in the context of COVID-19. The DPFL approach offers a secure and effective way to leverage large, multi-institutional datasets for developing highly accurate predictive models without compromising patient confidentiality.

The AI department’s success paves the way for further exploration of DPFL in various medical applications. This holds the potential to revolutionize how we diagnose, treat, and manage diverse diseases while ensuring patient privacy remains a paramount concern.