Study Finds AI Safeguards Patient Data & Detects Sleep ApneaStudy Finds AI Safeguards Patient Data & Detects Sleep Apnea

Sleep apnea patients may eventually achieve higher quality sleep, thanks to a team of researchers at the University at Buffalo, who have successfully created a prototype encrypted network that uses artificial intelligence (AI) to process encrypted patient data safely and is highly effective in detecting sleep apnea.

A study of the prototype was presented at the 2024 International Conference on Pattern Recognition (ICPR) in December in Kolkata, India. 

“My core research focus is on how to enhance privacy,” said lead research investigator Nalini Ratha, PhD, an Empire Innovation Professor in the Department of Computer Science and Engineering at the University at Buffalo, which is part of the State University of New York (SUNY). “Medical data is one of the classic examples of where any privacy enhancement method will have a great positive impact.”

Ratha and his research students chose sleep apnea because it is a very commonly observed disease worldwide, plus the availability of data. They have been researching encrypted sleep apnea data for over two years, for which they have developed several algorithms.

Advanced, AI-based diagnostic tools like the new prototype strive to safely process encrypted medical data on third-party cloud service providers, such as Google or Amazon, and send the encrypted results to physicians and their patients.

Related:Navigating FDA’s Draft Guidance for AI-Enabled Medical Devices

“Our research uses encrypted data; we don’t invent any new encryption,” Ratha told MD+DI. “The new twist we add is how we process encrypted data. For instance, if you take the number three and then encrypt it using an existing encryption method, it will no longer look like three, but look like some other number based on the private key that you use.”

By taking the original encrypted three and using the same encryption algorithm for both encrypted two and three, the investigators used fully homomorphic encryption techniques to add the two numbers in an encrypted Word document and then return the results in an encrypted form.

By taking the original encrypted three and using the same encryption algorithm for both encrypted two and three, the investigators used fully homomorphic encryption techniques to add the two encrypted numbers and then return the results in encrypted form.

Every complex operation must be performed on encrypted data for the ECG signal processing.

“There are deep learning or machine learning algorithms in general that process sleep apnea signals from an electrocardiogram (ECG),” Ratha said. “We take one of those encrypted signals and make the algorithm computable on an encrypted signal. It is similar to a credit card number being transmitted over the internet. Only the other end sees it.”

Related:FDA Issues Draft Guidance for AI-Enabled Devices, Seeks Public Feedback

Deep learning algorithms can identify patterns in the ECG signals that indicate disruptions in breathing or decreased oxygen levels during sleep. By analyzing copious amounts of ECG data, these models can also learn to detect subtle abnormalities that may be challenging for doctors to identify. 

The major obstacle of the prototype encrypted network is mapping the machine learning, in particular a deep learning algorithm, into the encrypted ECG signal. “A deep learning algorithm has a convolutional layer, a pooling layer, a nonlinear layer, and a fully connected layer,” Ratha said. “All four of these layers need to be transported or mapped into our fully homomorphic encryption (FHE) processing.”

The study found the new method was 99.56% effective in detecting sleep apnea from the deidentified ECG dataset available for research, while also supporting patient privacy.

Ratha attributed the high efficacy rate to selecting the right approximation algorithm. And by using FHE for encryption, the study accommodated 128-bit security for the entire pipeline of cloud-based medical diagnosis, including during inference.

Related:AI-Driven App Helps Combat Stress & Anxiety

However, processing sleep apnea data must be accomplished in a reasonable amount of time to be clinically beneficial. The study was able to process encrypted data within minutes, courtesy of a large number of multicore processors.

“ECG signals are cost-effective and convenient,” Ratha said. “By analyzing the variations and patterns in ECG data, advanced algorithms can detect anomalies in encrypted ECG data indicative of sleep apnea with high accuracy, contributing to more effective and timely diagnosis and treatment.”

The researchers plan on partnering with an existing healthcare company to commercialize their prototype, optimistically within the next year.

“I believe we’ve done the heavy lifting on our end, so it should be easy for physicians to embrace our technology that allows for medical data that is private, safe, and accurate,” Ratha said. “We are excited that we were able to solve this problem to address medical data privacy.”

The technology is likely applicable to other devices that rely on signal data, foremost X-ray images, CT scans, and MRIs. Ratha’s group is currently evaluating MRIs and X-ray images.