Software-Only Method Syncs Heterogeneous Biomedical Sensors Within Milliseconds

Biomedical researchers who want to combine signals from several wearable devices often run into a deceptively simple problem: the clocks inside those devices do not agree. An electrocardiogram monitor, an inertial measurement unit and a reference recorder may each stamp their data with their own internal time, and even small differences in clock offset and clock rate can smear measurements apart by tens or hundreds of milliseconds. A study published in BMC Biomedical Engineering now presents a software-only answer to this problem, reporting that a high-level acquisition protocol paired with offline linear regression kept ECG R-peak timing aligned with an independent gold-standard device to within an average of 20.1 milliseconds across a full hour of recording.

The research, led by Damiano Fruet, Stefano Cimignolo and Giandomenico Nollo, addresses a gap that has grown as physiological monitoring spreads from the clinic to wearables. Modern sensors can capture electrocardiograms, photoplethysmography, body temperature, arterial pressure and respiratory rate, and many are now embedded in smartphones and wearable devices for minimally invasive vital-sign monitoring. Combining these signals with environmental factors and clinical observations promises better diagnoses, but the authors note that multi-device acquisition is complicated by heterogeneous sampling frequencies, resolutions, signal-to-noise ratios and, critically, temporal references. Without a unified time domain, data fusion and artificial intelligence analyses lose their footing.

Existing synchronization strategies each carry trade-offs, and the new paper positions its method against them directly. Hardware triggering distributes a common signal from a shared reference generator to every device, achieving high timing accuracy but demanding extra equipment and physical input channels that low-cost compact wearables often lack. Software cross-correlation methods align streams offline by detecting a shared physical pattern recorded on two independent sensors, but this requires signal redundancy, which rules out integrating entirely distinct signal types. Open-source frameworks such as the Lab Streaming Layer offer high-level synchronization, yet they require devices to natively support the protocol or to be bridged through a compatible streaming layer, leaving many closed, SDK-restricted commercial devices out of reach.

The proposed framework takes a different route: rather than a new mathematical optimization algorithm, its novelty lies in a system-level, integration-focused software design aimed at non-programmable, SDK-constrained biomedical devices. A master-slave architecture places a central PC at the hub of the network. Each device streams its raw data packets, tagged with its own internal timestamps, to the PC, which immediately labels every arriving packet with its own reception time. The method models each device’s local clock as a linear function of the master reference time, with a scaling factor capturing clock skew, the drift that changes the offset over time, and an offset term capturing the absolute difference between the device clock and the PC clock at a given instant.

The workflow unfolds in two steps. During data acquisition, a continuous loop checks each device for available packets and stores three things for every one: the sensor data, the device’s internal timestamp and the PC’s nanosecond-resolution reception time. Because the two timestamps are read on consecutive lines of code, a small delay is unavoidable, arising from software execution overhead, communication latency, operating system scheduling and the intrinsic latency of clock access. In the offline synchronization step, a least-squares linear regression fits each device’s internal timestamps to its paired PC timestamps, yielding a slope and intercept that map the device clock into the common PC time domain. Because every device is aligned to the PC, all devices are aligned to each other.

The authors emphasize that this regression-based design is inherently resilient to network imperfections. If frames are lost, the offline process simply works with fewer paired observations without compromising the global regression line. When Bluetooth buffering or protocol retransmission causes multiple packets to arrive in a single burst, the PC records nearly identical reception times, but because the hardware-generated internal timestamps remain unique and sequential, the global line of best fit averages out these temporary arrival fluctuations. The framework does, however, depend on access to raw data and some form of internal timestamp or sequential counter through a device SDK; if a manufacturer’s ecosystem is entirely closed, the method cannot separate hardware clock drift from variable transmission latency.

To validate the approach, the team built a multimodal network of two acquisition devices and one independent reference. A CGX AIM system from Cognionics acquired single-lead ECG at 500 Hz over Bluetooth Classic, with electrodes placed on the clavicle and lower rib cage to optimize R-peak detection. A Movella DOT inertial measurement unit sampled acceleration at 60 Hz over Bluetooth Low Energy through a dedicated SDK. The gold standard was an Equivital Eq. 02 Lifemonitor sampling at 256 Hz, which recorded ECG and acceleration autonomously to its internal flash memory, untouched by the synchronization process, and was downloaded afterward over USB. The Movella DOT was physically affixed to the Equivital unit and aligned along the same coordinate axes.

The validation protocol consisted of six 10-minute phases, each containing a resting period followed by ten jumps that produced high-dynamic acceleration peaks for cross-correlating the Movella DOT with the Equivital reference. After acquisition, synchronization was finalized in MATLAB R2024b. The regression models showed strong correlations, with root-mean-squared errors of 7.5 milliseconds for the CGX AIM and 83.2 milliseconds for the Movella DOT, the latter constrained by its lower 60 Hz streaming rate and extensive SDK buffering. The one-hour session with a single healthy subject was then segmented into six stationary resting windows of 600 seconds each, and R-peak occurrence times from the CGX AIM and Equivital ECG signals were compared using a modified Pan-Tompkins detection algorithm and Bland-Altman analysis.

The results showed an overall mean R-peak synchronization delay of 20.1 milliseconds over the hour, a 0.03 percent difference across the acquisition period. Individual windows ranged from -124.1 milliseconds to 70.8 milliseconds in mean difference, with standard deviations between 7.8 and 37.4 milliseconds; windows 3 and 4 were the most consistent. The authors attribute the inter-window variability to minor hardware delays, automatic internal clock adjustments during acquisition and detection sensitivity within the Pan-Tompkins algorithm, and they note that a frame-by-frame analysis yielded a local mean absolute error of 58 milliseconds. Because the Equivital operated independently of the synchronized network, the observed agreement serves as a check against an unbiased reference. The authors also report that existing software-level cross-correlation methods have achieved mean delays around 130 milliseconds under optimal conditions, a figure their approach improved upon in this test.

The team is candid about the method’s boundaries. It is designed for offline post-processing, since linear regression needs the full session dataset to compute a stable fit; extending it to real-time or closed-loop systems would require sliding-window or adaptive-filter approaches to update clock coefficients as packets arrive. Multi-day monitoring could expose non-linear thermal drift that a linear model may not capture, potentially calling for higher-order polynomial regression or adaptive filtering. The validation itself rested on a single healthy subject in a controlled environment, a deliberate proof-of-concept that isolates algorithmic precision from biological variability, and the authors say future work must test the framework across larger populations, more dynamic movement and the higher packet-dropout rates such conditions bring. Within those limits, they argue, the demonstration that high-level software control can achieve millisecond-scale multimodal synchronization, with no external hardware and scalability limited only by PC connectivity, lays a practical foundation for body area networks and the validation of emerging wearable devices.

Subject of Research: A software-based method for synchronizing timestamps from heterogeneous biomedical devices using offline linear regression

Article Title: Multimodal data synchronization: a high-level software methodology for heterogeneous devices

Article References: Fruet, D., Cimignolo, S., & Nollo, G. (2026). Multimodal data synchronization: a high-level software methodology for heterogeneous devices. BMC Biomedical Engineering, 8(1), Article 12. https://doi.org/10.1186/s42490-026-00114-x

Image Credits: AI Generated

DOI: 10.1186/s42490-026-00114-x

Keywords: data synchronization, biomedical engineering, wearable sensors, ECG, inertial measurement unit, clock skew, linear regression, Bland-Altman analysis, Bluetooth, multimodal sensing, Multimodal, data