Noninvasive AI Tech Unlocks Hidden Insights in Heart Cells

Scientists have created a groundbreaking AI model that decodes heart cell signals without invasive procedures, paving the way for faster, safer drug development and personalized treatments.

Research: Intelligent in-cell electrophysiology: Reconstructing intracellular action potentials using a physics-informed deep learning model trained on nanoelectrode array recordings. Image Credit: FlashMovie / Shutterstock

A team of researchers led by the University of California San Diego and Stanford University have developed a noninvasive method to monitor the electrical activity inside heart muscle cells from the outside, avoiding the need to physically penetrate the cells. The method, published in the journal Nature Communications, relies on recording electrical signals from outside the cells and using a physics-informed deep learning model called PIA-UNET to reconstruct the signals within the cells with impressive accuracy.

The electrical signals inside heart muscle cells provide insights into how the heart functions, how its cells communicate, and how they respond to drugs. However, capturing these signals typically involves puncturing the cells with tiny electrodes, which can damage them and make large-scale testing complicated.

Now, researchers have found a way to peer inside the cells without actually going in.

a Scanning Electron Microscope (SEM) image of an NEA channel with nine nanocrown electrodes, alongside a schematic of the NEA setup. b A methodological comparison for capturing cardiac action potentials, presenting NEA alongside the reference patch clamp (PC) technique, ordered by their degree of invasiveness. The NEA’s functionality to convert eAPs into iAPs through precise biphasic electric pulses (electroporation) is demonstrated. c Simultaneous iAP recording from neighboring channels with a multi-step addition of Dofetilide, administered in concentrations of 0.3, 1, 3, and 10 nM. The process was conducted in three stages at approximately 400, 800, and 1200 s during the recording session to collect a diverse range of iAP shapes. d Comparison of iAP recording from neighboring channels by normalizing and segmenting them into arrays of length 8000 indices or 1.6 s. e Box plot distributions of difference in cycle time (dCT), correlation coefficient (r), mean absolute error (MAE), and APD50% and APD90% errors between neighboring and normalized iAP pairs with S/N > S/N* (= 90) for n = 2661 samples, comparing 22 pairs of neighboring iAP channels from two independent cell cultures. The box plot shows the median (center line), interquartile range (IQR; box bounds), whiskers (1.5×IQR), and outliers (points beyond whiskers). f Examples of iAP pairs from neighboring channels with the highest MAE, APD50, and APD90% errors, as indicated in the box plots. g collecting diverse iAPs and corresponding eAPs waveforms from neighboring channels on NEA from hiPSC-CMs, through the addition of drugs. h Collection of iAP/eAP pairs from neighboring channels by applying electroporation to one channel. Both iAP and eAP waveforms are normalized according to the Methods section. The signals are then segmented into windows of 800 indices or 1.6 s. The figure presents a wide range of collected eAPs and iAPs, including an overlay of the non-normalized eAP and iAP pairs. The study goal is to reconstruct iAP from eAP as illustrated.

The key lies in extracting the relationship between the signals inside the cells (intracellular signals) and those recorded on their surface (extracellular signals). The researchers' analysis revealed strong correlations between specific extracellular and intracellular features, such as amplitude and spiking velocity, leading to the discovery that extracellular signals hold the information needed to unlock intracellular features. "We discovered that extracellular signals hold the information we need to unlock the intracellular features that we're interested in," said Zeinab Jahed, a professor in the Aiiso Yufeng Li Family Department of Chemical and Nano Engineering at UC San Diego, who is one of the study's senior authors. Keivan Rahmani, a nano engineering Ph.D. student in Jahed's lab, is the first author on the study.

While extracellular signals can be captured with less invasive methods, they do not provide much detail about the cell's electrical activity. "It is like listening to a conversation through a wall–you can detect that communication is happening, but you miss the specific details," explained Jahed. "In contrast, intracellular signals offer the details, making you feel like you are sitting inside the room hearing every word clearly, but they can only be captured by invasive and more technically challenging methods." Using the PIA-UNET deep learning model, Jahed, Rahmani, and colleagues developed a method to correlate extracellular signals with specific intracellular signals.

To develop the new method, the team first engineered an array of nanoscale, crown-shaped, needle-shaped electrodes. These electrodes, each up to 200 times smaller than a single heart muscle cell, are made of silica-coated with platinum. Heart muscle cells, derived from stem cells, were grown and then placed on the electrode array.

The researchers collected a massive dataset—thousands of pairs of electrical signals—each pair linking an extracellular recording with its corresponding intracellular signal. The data included how the cells responded when exposed to various drugs, such as dofetilide, quinidine, nifedipine, and propranolol. This offered a rich library of data on how heart muscle cells behave under different conditions.

When analyzing these pairs, researchers identified patterns between the extracellular and intracellular signals. They then trained the physics-informed deep learning model PIA-UNET, which incorporates domain-specific knowledge to improve accuracy, to predict what the intracellular signals looked like based solely on the extracellular recordings. In tests, their model created accurate and complete reconstructions of the intracellular signals.

This work has important applications in drug screening, said Jahed. Every new pharmaceutical must undergo rigorous testing to ensure it does not adversely affect the heart—a process known as cardiotoxicity testing. Part of this process involves collecting detailed intracellular data from heart cells. Subtle changes in these electrical signals can provide clues about a drug's effects on the heart, which can help drug developers assess the safety of new medications. "Currently, this is a lengthy and costly process. It typically starts with tests on animal models, which don't always predict human outcomes," said Jahed.

By using this AI-driven approach combined with high-throughput nanoelectrode arrays, researchers can screen drugs directly on human heart cells. This can offer a more accurate picture of how a drug will behave in the human body and potentially bypass the need for early-stage animal testing.

"This could dramatically reduce the time and cost of drug development," said Jahed. "And because the cells used in these tests are derived from human stem cells, it also opens the door to personalized medicine. Drugs could be screened on patient-specific cells, enabling customized treatments to predict how an individual might respond to these treatments."

While the current study focused on heart muscle cells, the researchers are already working to expand their method to other types of cells, including neurons, with the goal of applying this technology to neurological disorders and other tissues. Their goal is to apply this technology to better understand a wide array of cellular activities in different tissues.