AI Detects Baby Heart Issues Without Skin Contact

A New AI Model Can Reconstruct a Premature Baby’s Heart Signal with 96% Accuracy

Adhesive ECG electrodes can damage the fragile skin of newborns in the NICU. To reduce this risk, a research team led by Choi, S., Gu, G., Kim, Y., and others has developed a method to reconstruct missing sections of a neonatal electrocardiogram (ECG) using a less invasive sensor. Their model, detailed in a new preprint, filled in lost ECG data with a 0.96 Pearson correlation to the original signal. The work focuses on a major technical challenge: matching the heart’s electrical activity from an ECG with blood volume pulses from a photoplethysmogram (PPG) in premature infants, whose physiology is highly variable.

The Challenge of Protecting Neonatal Skin While Monitoring the Heart

Continuous ECG monitoring is a standard of care in the NICU, but it comes with a cost. The repeated application and removal of adhesive electrodes can strip the underdeveloped outer layer of a preterm infant’s skin, leading to injury, pain, and increased infection risk. Researchers have explored using a single, often less-irritating PPG sensor—the technology used in pulse oximeters—to estimate the ECG. Most prior studies, however, used data from adults.

This presents a problem. Adults have a relatively stable pulse arrival time (PAT), the delay between the heart’s electrical impulse and the resulting pulse of blood reaching a peripheral site like the foot. In premature infants, PAT is highly inconsistent. Their immature autonomic nervous systems and rapidly changing hemodynamics make direct, forced alignment of PPG and ECG signals unreliable. The team set out to build a model that could learn this complex relationship without needing to artificially sync the two signals first.

How the Dual-Stream Transformer Learns from Messy Neonatal Data

Choi and colleagues trained their model on 52,566 ten-second windows of synchronized ECG and PPG data, taken from 159 NICU patients. They used a patient-level split to ensure the model was tested on entirely new babies, not just new segments from familiar patients.

The core of their approach is an “alignment-free RoPE-based dual-stream Transformer.” In practice, this means the model processes two streams of information simultaneously. One stream analyzes the PPG signal to understand the baby’s unique, moment-to-moment hemodynamic timing. The other stream examines the available ECG context from before and after a missing segment. A positional encoding technique called Rotary Position Embedding (RoPE) helps the model understand the order and timing of these signals without a rigid alignment step.

The model’s task was inpainting: given a PPG signal and an ECG with a large missing section, generate a plausible reconstruction for that gap. The researchers tested it under harsh conditions, including randomly deleting 40% of the ECG data, simulating a 4.0-second continuous blackout, and a 60% random patch loss.

High Accuracy in Simulated Monitoring Failures and Future Steps

Under the 40% random missing scenario, the model’s reconstructed ECG matched the original with a Pearson correlation coefficient (PCC) of 0.96. The mean absolute error was 0.04, and the root mean square error was 0.07. These results indicate a very close match to the true heart signal.

Performance remained strong in more extreme simulations. With a 4-second block of continuous data missing, the PCC stayed above 0.90. It also maintained a PCC of at least 0.90 under the 60% patch loss condition. This suggests the framework could be resilient to various types of signal interruption common in busy clinical settings.

The authors propose their model could eventually serve as a signal imputation module within NICU monitors. If an ECG electrode becomes loose or is temporarily removed for skin care, the system could use the ongoing PPG signal to fill in the gaps, preserving a continuous record for clinicians.

This study has important limitations. It is a preprint and has not undergone peer review. The data is retrospective, and the model’s performance was measured against the original ECG signal in a controlled analysis. Prospective validation in a live NICU environment is the essential next step to see if the model performs reliably in real time with new patients and different monitoring equipment. Furthermore, the reconstructed signal is intended for monitoring continuity, not for primary diagnostic use, which would require even more rigorous testing and regulatory approval.

Source:
Alignment-Free RoPE-Based Dual-Stream Transformer for PPG-Guided Neonatal ECG Segment Inpainting in the NICU (medRxiv preprint, 2026-07-10)