An autonomous platform uses machine learning and patterned light to detect and terminate cardiac arrhythmias in real time without electrical shocks.
(Nanowerk Spotlight) A sudden cardiac arrhythmia, a disruption of the heart’s normal electrical rhythm, can rapidly become life-threatening. The electrical waves that normally sweep across the heart in orderly sequence collapse into chaotic spirals, each rotating around an invisible organizing center that drifts unpredictably through the tissue.
Cardiologists have two main tools to fight back, and neither is adequate. Electrical cardioversion shocks the entire heart back into rhythm, but the jolt is painful and indiscriminate, damaging healthy tissue alongside diseased. Catheter ablation can destroy the precise spot generating a stable spiral, but if that spot moves, the procedure cannot follow.
What both approaches lack is adaptivity: the ability to watch the arrhythmia evolve and respond to it in real time. Optogenetics offers a fundamentally different mode of intervention. The technique involves genetically modifying cells to express light-sensitive proteins called channelrhodopsins, enabling light to trigger or suppress electrical activity with spatial precision that electricity cannot match.
Separately, optical voltage mapping can visualize membrane potential changes across tissue at high resolution, and machine learning algorithms have grown capable of recognizing complex spatiotemporal patterns in biological signals. Until now, no system had combined all three into a single autonomous control loop fast enough for cardiac tissue.
A study published in Advanced Science (“Smart Optogenetics for Real‐Time Automated Control of Cardiac Electrical Activity”) presents such a system. The research team built an integrated platform combining optical voltage mapping, a convolutional neural network, and patterned optogenetic stimulation to detect and terminate cardiac arrhythmias in cell cultures. Once running, the system operates autonomously; human intervention is required only to initiate the arrhythmia.
Schematic overview of the integrated software and hardware platform (system) for real-time optical voltage mapping (OVM) and manipulation of cardiomyocyte (hiAM) monolayers. The system is divided in three main components: software (orange background) for real-time image processing, AI-driven spiral wave detection, and adaptive action pattern generation; hardware (green background) featuring an mLED matrix with custom driver circuitry, PCB design, and optimized spatial pitch for precise pattern illumination; and a closed-loop interface (blue background) that synchronizes real-time optical data acquisition with targeted illumination, enabling automated initiation, detection, and termination of spiral waves. This bidirectional system provides rapid feedback for investigating arrhythmia dynamics and intervention strategies. (Image: Reproduced from DOI:10.1002/advs.202522759, CC BY) (click on image to enlarge)
The platform has three interconnected components. The first is biological: monolayers of human conditionally immortalized atrial myocytes engineered to express CheRiff, a blue light-sensitive channelrhodopsin that depolarizes cells upon illumination. These cells sustain the high-frequency reentrant circuits characteristic of atrial arrhythmias. Because CheRiff’s activation spectrum avoids overlap with the voltage-sensitive imaging dye, the system can stimulate and observe simultaneously.
To enable reentry induction in cells whose action potential duration was too long for spirals to form, the researchers added the potassium channel opener P1075 to the culture medium, a pharmacological adjustment that highlights the biological fine-tuning the system still requires.
The second component is hardware. The team fabricated two custom miniature LED matrices on flexible polyimide substrates, with individual LEDs measuring 125 × 250 µm². A smaller matrix (48 × 32 pixels, 0.6 mm pitch) suited single wells of a 12-well plate. A larger version (160 × 160 pixels, 0.9 mm pitch) could cover monolayers up to 14 cm across.
Both matrices achieved roughly 80% illumination uniformity and produced negligible heating. Optimization experiments showed that pixel pitches between 0.8 and 0.9 mm minimized the total power needed to terminate reentrant waves. Getting this pitch right was critical, since the entire closed-loop system depends on tight spatial alignment between the LED matrix, the camera, and the detection algorithm.
The third component is software. Experimental recordings of arrhythmias typically lack the spatial resolution to pinpoint spiral cores reliably. To overcome this, the researchers generated synthetic training data using a validated computational model of cardiac conduction that incorporated realistic cell-to-cell variability.
A shallow convolutional neural network was trained on 7,776 simulated spiral patterns to locate phase singularities from just 30 milliseconds of activity, less than a single spiral rotation. The model occupies only 62 kilobytes.
Although trained on simulated rat ventricular myocyte data, the network accurately identified spiral cores in the human atrial myocyte monolayers used for experiments, without any transfer learning. This cross-species generalization suggests it captured fundamental spiral wave dynamics rather than species-specific signal features. The network performed well on classical spiral patterns but was challenged by atypical wave shapes, a limitation that became apparent during closed-loop testing.
In operation, a camera captures membrane voltage images at 8 ms intervals. The machine learning module processes these frames and produces a probability map of spiral core locations. When confidence exceeds 70% across 10 consecutive frames, the system generates a line of light connecting the detected core to the nearest tissue boundary.
The LED matrix delivers this pattern for 0.5 seconds. The overall loop time is approximately 80 to 100 milliseconds, fast enough to respond within one rotational period of a spiral wave. The spatial resolution of 0.9 mm is smaller than typical spiral core sizes in atria (1 to 5 mm) and ventricles (3 to 10 mm).
To induce arrhythmias for testing, the team used a programmed stimulation protocol in which an initial electrical pulse was followed by a timed optical pulse covering half the monolayer, triggering reentry. A 3-second delay then allowed spirals to stabilize, during which roughly 10% terminated spontaneously.
Of the remaining spirals, tested across 73 reentry events in five samples, roughly half were terminated with a single attempt within 2 seconds. More than 80% were eliminated within 10 seconds using fewer than three attempts. All were terminated within 30 seconds. In control experiments without light, fewer than 10% stopped on their own.
The system also handled multiple simultaneous rotors in larger monolayers. In one experiment, it located spiral cores across three wells within 376 milliseconds and terminated all of them with a single simultaneous light pulse, while reentry persisted in an unilluminated control well.
The spatial tolerance between camera, algorithm, and LED matrix was experimentally determined to be 1.8 mm, meaning the system can tolerate moderate misalignment. Spirals with nonclassical morphology required longer detection times and more attempts, confirming the earlier observation that atypical wave shapes remain the algorithm’s primary weakness.
The system remains a proof of concept. Optical voltage mapping cannot be used inside living organisms, so clinical translation would require electrode-based sensing, likely through implantable micro-electrode arrays paired with integrated LED matrices. The rigid prototype would need to become stretchable and foldable to conform to the heart’s surface.
Newer, more light-sensitive channelrhodopsins such as bReaChES, ChRmine, or ChReef could reduce optical power requirements, an important factor for implanted devices. The convolutional neural network struggled under high-noise conditions, and future versions might benefit from transformer-based approaches.
The researchers envision reinforcement learning enabling the system to optimize its stimulation strategies autonomously over time. Concurrent work by another group, still at the preprint stage, has demonstrated a similar real-time LED-based optogenetic approach, indicating that the field is converging on closed-loop optical control of cardiac rhythm.
The platform establishes that continuous monitoring, pattern recognition, and targeted light can operate together fast enough to track and terminate drifting spiral waves. It provides a concrete technical foundation for shock-free, adaptive arrhythmia therapy.
For authors and communications departmentsclick to open
Lay summary
Prefilled posts
ORCID information
Daniël A. Pijnappels (Leiden University Medical Center)
, 0000-0001-6731-4125 corresponding author
Nanowerk Newsletter
Get our Nanotechnology Spotlight updates to your inbox!
Thank you!
You have successfully joined our subscriber list.
Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com.
