Multimodal data synchronization: a high-level software methodology for heterogeneous devices Authors: Damiano Fruet, Stefano Cimignolo, Giandomenico Nollo Published: 2026 ABSTRACT BackgroundMultimodal data acquisition from heterogeneous devices is frequently compromised by synchronization errors arising from clock skew and offset. Current solutions often necessitate complex external hardware or shared reference signals, which complicates deployment and limits system scalability.MethodsThis research introduces a high-level software methodology that utilizes a dedicated data acquisition protocol to record device-specific timestamps alongside a master PC timestamp. An offline linear regression model is then employed to convert internal timestamps to a common reference, compensating for clock discrepancies. The method was validated using electrocardiographic and Inertial Measurement Unit devices compared against a gold-standard reference. The ECG R-peaks fiducial points from the synchronized devices and the gold standard were compared.ResultsValidation showed an average R-peak synchronization delay of 20.1 ms over a one-hour acquisition (a 0.03% difference). The method successfully eliminates the need for external synchronization hardware, allowing scaling to be limited only by PC connectivity.ConclusionWhile dependent on device Software Development Kits, this methodology provides a robust, scalable foundation for precise multimodal synchronization. This proof-of-concept demonstrates that high-level software control can achieve high accuracy for dynamic clock adjustment. Introduction Biomedical signals are observations of physiological data, such as electrocardiograms (ECG), photoplethysmography (PPG) data, body temperature, arterial pressure, and respiratory rate [1]. To acquire these signals, novel sensors have been developed in recent years thanks to technological advances, including tactile [2], silicon [3], and multi-domain systems. Specifically, sensors can be integrated into common devices such as smartphones [4] or wearable devices [5], allowing for minimally invasive monitoring of vital parameters. Combining biomedical signals, environmental factors, and clinical observations offers significant promise for improving diagnoses [6, 7]. However, substantial challenges result from the inherent complexity of human physiological behavior and the difficulties associated with multi-device data acquisition. These challenges originate primarily from the heterogeneity of acquisition devices, varying acquisition protocols, and different characteristics such as sampling frequencies, resolutions, ranges, signal-to-noise ratios (SNRs), and, critically, temporal references. The lack of a unified temporal reference significantly complicates sensor integration, particularly for biomedical applications demanding precise synchronization [8, 9]. Consequently, achieving effective sensor integration necessitates robust synchronization strategies, implemented through either hardware [10] or software [11] solutions. Such synchronization is essential for accurately correlating acquired data within the same time domain, creating a unified domain for post-processing analyses using data fusion or artificial intelligence algorithms [12, 13]. The concept of synchronization was investigated by researchers across diverse scenarios, encompassing low-level communication within Wireless Sensor Networks (WSNs) [14] and Internet of Things (IoT) applications [15], as well as high-level communication, where device synchronization relies on methodologies beyond programmable hardware [11]. The first category, comprising intelligent and programmable devices, enables complex and flexible synchronization algorithms. These algorithms, fully adaptable to diverse needs, utilize specific protocols like IEEE 802.15.4 [16] and manage internal device resources in a distributed manner for optimization. Conversely, high-level networks with non-programmable commercial devices must rely on simpler, centralized mechanisms. Because the communication protocols implemented in these commercial devices are closed, their analog-to-digital conversion, internal timing stages, and hardware configurations are practically unknown to the end user, necessitating a high-level software synchronization procedure. While some studies have assumed that strict synchronization is not fundamental under the assumption that internal device clocks are already sufficiently synchronous [17, 18], these approaches fail to consider the critical limitations imposed by cumulative clock offset and clock delay over time [19]. To overcome these barriers without low-level hardware modifications, several alternative frameworks exist in the literature, though each carries distinct operational trade-offs. Hardware triggering methodologies often rely on a common signal distributed to each device via an auxiliary physical channel by a shared reference generator [10, 22]. While this approach provides high timing accuracy, it introduces significant hardware complexity and prevents the use of low-cost or compact wearable sensors that lack physical input channels. Alternatively, software-based cross-correlation methods synchronize multimodal streams offline by detecting a shared physical pattern recorded simultaneously across separate devices [8]. However, the primary limitation lies in the strict requirement for signal redundancy, meaning two independent sensors must record overlapping modalities, which severely restricts the integration of entirely distinct signal types. Finally, open-source protocols like the Lab Streaming Layer (LSL) [20] facilitate high-level synchronization across software applications [20, 21]. Yet, LSL requires devices to either natively support its communication protocol or rely on an architectural layer capable of streaming data through specific LSL functions, making it incompatible with many closed, SDK-restricted commercial devices. To explicitly position these distinct methodologies against our framework, Table 1 provides a structured comparative overview. Table 1Structured comparison of multi-device synchronization approachesSynchronization ApproachHardware Modifications Required?Signal Redundancy Required?Requires Device-Side Open Architecture?Handles Clock Skew and Drift?Main LimitationHardware TriggeringYes (Dedicated cables/generators)NoYes (Requires physical input channels)YesHigh complexity, poor scalability, costlyCross-CorrelationNoYes (Sensors must record identical signal types)NoPoorly (Typically handles offset, not dynamic skew)Restricts network to overlapping modalitiesLab Streaming Layer (LSL)NoNoYes (Requires native LSL support or custom apps)YesIncompatible with closed, SDK-restricted commercial devicesProposed FrameworkNoNoNo (Requires only standard manufacturer SDK)YesDependent on raw timestamp availability Table 1Structured comparison of multi-device synchronization approachesSynchronization ApproachHardware Modifications Required?Signal Redundancy Required?Requires Device-Side Open Architecture?Handles Clock Skew and Drift?Main LimitationHardware TriggeringYes (Dedicated cables/generators)NoYes (Requires physical input channels)YesHigh complexity, poor scalability, costlyCross-CorrelationNoYes (Sensors must record identical signal types)NoPoorly (Typically handles offset, not dynamic skew)Restricts network to overlapping modalitiesLab Streaming Layer (LSL)NoNoYes (Requires native LSL support or custom apps)YesIncompatible with closed, SDK-restricted commercial devicesProposed FrameworkNoNoNo (Requires only standard manufacturer SDK)YesDependent on raw timestamp availability Structured comparison of multi-device synchronization approaches Rather than introducing a novel mathematical optimization algorithm, the primary novelty of this research lies in a system-level and integration-focused software advancement designed specifically for non-programmable, SDK-constrained biomedical devices. This approach utilizes a master-slave communication architecture, where a central PC coordinates with each device. The core concept involves collecting device-specific timestamped data alongside master PC reception time and subsequently utilizing a centralized linear regression mapping to align all streams into a common time domain. The primary contributions of this integration-focused framework are: Simplified Device Integration: Synchronization is achieved through high-level software, removing the requirement for low-level device programming and leveraging readily available Software Development Kits (SDKs).Hardware Independence: The methodology eliminates the need for additional hardware components beyond the devices and the central PC.No External Reference Signals: The synchronization method does not require the devices to record or utilize any external reference signals, allowing devices without additional recording channels to synchronize.Accurate Time Alignment: The proposed approach specifically addresses both clock skew and offset, ensuring precise temporal alignment of data from different devices.Scalability: Enabling synchronization across numerous devices, limited only by PC connectivity. Simplified Device Integration: Synchronization is achieved through high-level software, removing the requirement for low-level device programming and leveraging readily available Software Development Kits (SDKs). Simplified Device Integration: Synchronization is achieved through high-level software, removing the requirement for low-level device programming and leveraging readily available Software Development Kits (SDKs). Hardware Independence: The methodology eliminates the need for additional hardware components beyond the devices and the central PC. Hardware Independence: The methodology eliminates the need for additional hardware components beyond the devices and the central PC. No External Reference Signals: The synchronization method does not require the devices to record or utilize any external reference signals, allowing devices without additional recording channels to synchronize. No External Reference Signals: The synchronization method does not require the devices to record or utilize any external reference signals, allowing devices without additional recording channels to synchronize. Accurate Time Alignment: The proposed approach specifically addresses both clock skew and offset, ensuring precise temporal alignment of data from different devices. Accurate Time Alignment: The proposed approach specifically addresses both clock skew and offset, ensuring precise temporal alignment of data from different devices. Scalability: Enabling synchronization across numerous devices, limited only by PC connectivity. Scalability: Enabling synchronization across numerous devices, limited only by PC connectivity. To evaluate the effectiveness of this system-level synchronization framework, a validation experiment was conducted using a multimodal network consisting of a physiological monitoring system and an Inertial Measurement Unit (IMU) compared directly against an independent, unbiased gold-standard reference device. Materials and methods Synchronization methodology The framework utilized a versatile data acquisition procedure, enabling the synchronization of diverse devices across various fields through a high-level software approach. This methodology resolved synchronization challenges by compensating for clock offset and skew, where the clock offset refers to the absolute time difference between a device’s clock and the master clock at a given instant, while the clock skew represents the difference in their clock rates, that is, the drift that causes the offset to change over time. A central PC served as master, distributing its time across the network, with other devices adhering to the relationship:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{dn}\left(t\right)=\:{\alpha\:}_{n}*t+{c}_{o}$$\end{document} Equation (1) defines the local clock of device n, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{dn}\left(t\right)$$\end{document}, as a linear function of the master reference time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document}. The term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{n}$$\end{document} is the scaling factor (clock skew) that accounts for the difference in packet arrival time, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{o}$$\end{document} represents the absolute time difference (clock offset) between the master and device n’s clock. The goal of the synchronization process is to minimize the error defined as the difference between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{dn}\left(t\right)$$\end{document} and t. Specifically, the proposed procedure consists of two sequential steps: Data Acquisition: In this first step, raw data and timestamps are acquired from each device, along with the corresponding PC reception time. This process ensures continuous data acquisition, minimizing synchronization errors.Offline Synchronization: This step transforms the device data from their respective internal time domains to a unified PC time domain, thereby achieving synchronization. Signal alignment is accomplished through linear regression, which calculates both a scaling factor and an offset. The offset directly corrects for the difference in starting points between the device’s clock and the PC’s clock, while the scaling factor compensates for discrepancies in clock rates between the devices and the PC during acquisition. Data Acquisition: In this first step, raw data and timestamps are acquired from each device, along with the corresponding PC reception time. This process ensures continuous data acquisition, minimizing synchronization errors. Data Acquisition: In this first step, raw data and timestamps are acquired from each device, along with the corresponding PC reception time. This process ensures continuous data acquisition, minimizing synchronization errors. Offline Synchronization: This step transforms the device data from their respective internal time domains to a unified PC time domain, thereby achieving synchronization. Signal alignment is accomplished through linear regression, which calculates both a scaling factor and an offset. The offset directly corrects for the difference in starting points between the device’s clock and the PC’s clock, while the scaling factor compensates for discrepancies in clock rates between the devices and the PC during acquisition. Offline Synchronization: This step transforms the device data from their respective internal time domains to a unified PC time domain, thereby achieving synchronization. Signal alignment is accomplished through linear regression, which calculates both a scaling factor and an offset. The offset directly corrects for the difference in starting points between the device’s clock and the PC’s clock, while the scaling factor compensates for discrepancies in clock rates between the devices and the PC during acquisition. High-level software framework for data acquisition To ensure synchronization between multiple devices, the proposed methodology required full access to the raw data from each device, including both the data itself and its associated timestamp. This allowed for the integration of various data types from different domains (e.g., physiological signals, acceleration data, force data), each sample having a corresponding internal timestamp. These timestamps provided common points of reference for data sharing between devices. The main idea behind the acquisition procedure is graphically depicted in Fig. 1. The infrastructure consisted of a master (central PC) responsible for acquiring all the data from the devices connected to the network, and a set of devices connected to the master through their own specific protocol (e.g., Bluetooth or Ethernet). When a data packet becomes available (Data_Device n), the devices send the packet to the central PC, along with the timestamp associated with the packet (InternalTime_Device n). When a data packet reaches the central PC, it is immediately labeled with the PC’s current timestamp (PCTime_Device n), which uniquely associates the packet with the precise reception time and the device of origin. This information was stored for use in the offline synchronization procedure. This procedure is repeated for each packet and for all connected devices during the entire acquisition process. Fig. 1Synchronization scheme: Simultaneous acquisition over wireless Bluetooth connections of data packets (containing raw data, e.g., acceleration/physiological signals, and internal timestamps) and PC timestamps from the two devices Synchronization scheme: Simultaneous acquisition over wireless Bluetooth connections of data packets (containing raw data, e.g., acceleration/physiological signals, and internal timestamps) and PC timestamps from the two devices Details of the data acquisition procedure are summarized in Pseudocode 1, using two devices as an example. The process begins by initializing and establishing communication with all connected devices. This initialization includes configuring parameters such as sampling frequency and packet type, defining potential interrupts, setting the communication protocol, and configuring the devices to send raw data to the PC (Pseudocode 1 - lines 1–4). A central loop continuously checks each device for available data packets (Pseudocode 1 - line 5). When a packet is available (Pseudocode 1 - lines 6 and 13 for Device 1 and 2, respectively), its data (Data_device n), along with the PC’s current timestamp (PCTime_Device n) and the device’s internal timestamp (InternalTime_Device n), are retrieved and stored (Pseudocode 1 - lines 7–9 and 14–16 for Device 1 and 2, respectively). This process aims to acquire two distinct timestamps for each sample from each device: one from the device’s internal clock and one from the master PC’s clock, establishing a relationship between the two. This dual timestamping is crucial for post-processing synchronization, enabling alignment of the data streams despite potential variations in packet arrival times. Specifically, each sample’s device timestamp is paired with the corresponding PC timestamp. While these two timestamps are acquired in close proximity, separated only by the execution time of two consecutive lines of code, acquiring them on separate lines, even if adjacent, introduces a small delay, preventing perfect simultaneity. Therefore, a post-processing timestamp alignment procedure is necessary to accurately correlate the timestamps for each device. The loop continues until a defined exit condition is met (Pseudocode 1 - line 20). Data acquisition from both devices is then stopped, and the collected data, including PC timestamps, internal device timestamps, and sensor readings, is saved to a file, likely a CSV, for subsequent offline synchronization (Pseudocode 1 - lines 22–24). Pseudocode 1. Synchronization acquisition protocol// Initialization1:Device1_Initialize(): Initialize the first sensor to communicate with the main PC2:Device2_Initialize(): Initialize the second sensor to communicate with the main PC // Setup and Connection 3:Device1_SetupAndConnect(sampling_frequency, num_channels, packet_type): Configure and connect the first sensor4:Device2_SetupAndConnect(sampling_frequency, num_channels, packet_type): Configure and connect the second sensor // Data Acquisition Loop 5:whileTRUE: Continuously acquire data until exit condition6:ifDevice1_PacketAvailable(): Check if Device 1 data is available7:packet1 ← Device1_GetPacketData(): Retrieve the next available packet of Device 18:PCTime_Device1 ← time.time_ns(): Get current PC time (nanoseconds)9:InternalTime_Device1 ← packet1.time: Get Device 1 internal timestamp10:Data_Device1 ← packet1.data: Get Device 1 data (e.g. acceleration, ECG)11:Device1_data.append(PCTime_Device1, InternalTime_Device1, Data_Device1): Store Device 1 data12: e nd if 13:ifDevice2_PacketAvailable(): Check if Device 2 data is available14:packet2 ← Device2_GetPacketData(): Retrieve the next available packet of Device 215:PCTime_Device2 ← time.time_ns(): Get current PC time (nanoseconds)16:InternalTime_Device2 ← packet2.time: Get Device 2 internal timestamp17:Data_Device2 ← packet2.data: Get Device 2 data (e.g. acceleration, ECG)18:Device2_data.append(PCTime_Device2, InternalTime_Device2, Data_Device2): Store Device 2 data19: end if 20:Exit Condition Check (e.g., if key_pressed(ESC))21: end while // Cleanup and Save 22:Device1_Stop()23:Device2_Stop()24:SaveData(Device1_data, Device2_data, “data.csv”): Save data to a CSV file or appropriate format Pseudocode 1. Synchronization acquisition protocol// Initialization1:Device1_Initialize(): Initialize the first sensor to communicate with the main PC2:Device2_Initialize(): Initialize the second sensor to communicate with the main PC // Setup and Connection 3:Device1_SetupAndConnect(sampling_frequency, num_channels, packet_type): Configure and connect the first sensor4:Device2_SetupAndConnect(sampling_frequency, num_channels, packet_type): Configure and connect the second sensor // Data Acquisition Loop 5:whileTRUE: Continuously acquire data until exit condition6:ifDevice1_PacketAvailable(): Check if Device 1 data is available7:packet1 ← Device1_GetPacketData(): Retrieve the next available packet of Device 18:PCTime_Device1 ← time.time_ns(): Get current PC time (nanoseconds)9:InternalTime_Device1 ← packet1.time: Get Device 1 internal timestamp10:Data_Device1 ← packet1.data: Get Device 1 data (e.g. acceleration, ECG)11:Device1_data.append(PCTime_Device1, InternalTime_Device1, Data_Device1): Store Device 1 data12: e nd if 13:ifDevice2_PacketAvailable(): Check if Device 2 data is available14:packet2 ← Device2_GetPacketData(): Retrieve the next available packet of Device 215:PCTime_Device2 ← time.time_ns(): Get current PC time (nanoseconds)16:InternalTime_Device2 ← packet2.time: Get Device 2 internal timestamp17:Data_Device2 ← packet2.data: Get Device 2 data (e.g. acceleration, ECG)18:Device2_data.append(PCTime_Device2, InternalTime_Device2, Data_Device2): Store Device 2 data19: end if 20:Exit Condition Check (e.g., if key_pressed(ESC))21: end while // Cleanup and Save 22:Device1_Stop()23:Device2_Stop()24:SaveData(Device1_data, Device2_data, “data.csv”): Save data to a CSV file or appropriate format Offline synchronization The data acquisition procedure results in n separate files, each from a different device and each composed of the following elements: Data_Device n: data sent to the PC by the device.InternalTime_Device n: timestamp associated with the packet, generated by the device.PCTime_Device n: timestamp acquired by the PC immediately after the packet arrives from the Device n. Data_Device n: data sent to the PC by the device. Data_Device n: data sent to the PC by the device. InternalTime_Device n: timestamp associated with the packet, generated by the device. InternalTime_Device n: timestamp associated with the packet, generated by the device. PCTime_Device n: timestamp acquired by the PC immediately after the packet arrives from the Device n. PCTime_Device n: timestamp acquired by the PC immediately after the packet arrives from the Device n. The aim of the offline synchronization procedure is to align the InternalTime_Device n with the associated PCTime_Device n, thus establishing a common time reference for all devices on the network. Even if the InternalTime_Device n and the PCTime_Device n timestamps for a given packet are acquired nearly simultaneously - separated only by the execution time of two consecutive lines of code (i.e., Pseudocode 1 - lines 8–9 for Device 1) - a potential delay is still introduced. This delay stems from multiple factors, including software execution overhead, communication latency between the device and the PC, operating system scheduling, and the intrinsic latency of system clock access. To compensate for this delay and accurately align the clocks, a linear regression model is applied, which fits the internal timestamp data from each device to its corresponding PCTime_Device n, effectively creating a close association between the device’s clock and the master PC’s clock. In relation to Eq. (1), for each device n, a linear equation is used to obtain the predicted PC Time (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:PCTim{e}_{fit}\:n$$\end{document}) from the acquired Device Internal Time data (InternalTime_Device n):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eqalign{ & PCTim{e_{fit}}n \cr & = {\beta _1}*InternalTime\_Devicen + {\beta _0} \cr} $$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{1}$$\end{document} represents the scaling factor (clock skew) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{0}$$\end{document} is the offset between the device’s internal clock and the PC’s clock. This mathematical formulation provides the framework with inherent resilience against network transmission errors, such as packet loss and jitter. In the case of missing frames, the offline synchronization process simply operates with fewer paired observations for that device without compromising the global calculation of the regression line. Conversely, when network jitter or protocol buffering (e.g., via Bluetooth or TCP retransmission) causes multiple packets to arrive simultaneously in a single burst, the master PC records nearly identical reception timestamps (PCTime_Device n). However, because the corresponding hardware-generated InternalTime_Device n values remain unique and sequential, the linear regression model naturally averages out these temporary arrival fluctuations by calculating the global line of best fit across the entire session. The coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{0}$$\end{document} are calculated as the slope and intercept, respectively, of the linear regression line that best describes the relationship between the device’s internal clock and the PC clock, based on m paired observations acquired during the procedure. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$InternalTim{e_{Device}}{n_i}$$\end{document}be the internal timestamp of device n for sample i, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PCTime\_Device\:{n_i}$$\end{document} be the PC timestamp associated with sample i of device n. Also, let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline {PCTime\_Devicen}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\overline{PCTime\_Device\:n\:}$$\end{document} represent the mean of these respective timestamp values. The scaling factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{1}$$\end{document} is first calculated using the standard formula for the slope of the least-squares regression line: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&{\beta _1} = \\&\frac{\begin{aligned}&\mathop \sum \nolimits_{i = 1}^m \left( {InternalTime\_Device{n_i} - \overline {InternalTime\_Devicen} } \right)\cr&\quad\left( {PCTime\_Device{n_i} - \overline {PCTime\_Devicen} } \right)\end{aligned}}{\mathop \sum \nolimits_{i = 1}^m {{\left( {InternalTime\_Device{n_i} - \overline {InternalTime\_Devicen} } \right)}^2}}\end{aligned}$$\end{document} Once \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{1}\:$$\end{document}is determined, the offset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{0}$$\end{document} is calculated using the means and the determined slope:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\beta _0} = & \overline {PCTime\_Device\,n} - {\beta _1} \\ & *\overline {InternalTime\_Device\,n} \\ \end{aligned}$$\end{document} This process is repeated for each of the n devices in the network. This results in n distinct regression curves, where each curve represents the mapping between a particular device’s internal clock and a common reference time maintained by the PC. By transforming each device’s timestamp to this shared PC time, the system establishes synchronization between each individual device and the PC (master). Consequently, because all devices are synchronized with the PC’s time, they are also synchronized with each other. Synchronization procedure validation Experimental setup and instrumentation The proposed methodology was validated using a multimodal sensor network synchronized to a master PC’s clock (Fig. 2). The setup comprised two primary acquisition devices and one independent reference gold standard: Physiological Sensing: A CGX AIM system (Cognionics, Inc., San Diego, CA, USA) [23] acquired single-lead ECG data at 500 Hz. Electrodes were placed on the clavicle and lower rib cage to optimize R-peak detection [24]. Data was transmitted wirelessly to the host PC via a standard Bluetooth Classic connection.Inertial Sensing: A DOT IMU (Movella, Enschede, The Netherlands) [25] with a sampling rate of 60 Hz was used to collect acceleration data, communicating via a dedicated SDK over Bluetooth Low Energy (BLE) connection.Gold Standard Reference: An Eq. 02 Lifemonitor (Equivital Eq. 02; Hidalgo Ltd., Cambridge, UK) [26] with a sampling rate of 256 Hz served as the validation benchmark This device, independent of the primary wireless acquisition network, simultaneously recorded ECG and acceleration directly to its internal flash memory to ensure an unbiased reference. Following the recording session, this data was downloaded to the host PC via a wired USB interface for offline synchronization analysis. Physiological Sensing: A CGX AIM system (Cognionics, Inc., San Diego, CA, USA) [23] acquired single-lead ECG data at 500 Hz. Electrodes were placed on the clavicle and lower rib cage to optimize R-peak detection [24]. Data was transmitted wirelessly to the host PC via a standard Bluetooth Classic connection. Physiological Sensing: A CGX AIM system (Cognionics, Inc., San Diego, CA, USA) [23] acquired single-lead ECG data at 500 Hz. Electrodes were placed on the clavicle and lower rib cage to optimize R-peak detection [24]. Data was transmitted wirelessly to the host PC via a standard Bluetooth Classic connection. Inertial Sensing: A DOT IMU (Movella, Enschede, The Netherlands) [25] with a sampling rate of 60 Hz was used to collect acceleration data, communicating via a dedicated SDK over Bluetooth Low Energy (BLE) connection. Inertial Sensing: A DOT IMU (Movella, Enschede, The Netherlands) [25] with a sampling rate of 60 Hz was used to collect acceleration data, communicating via a dedicated SDK over Bluetooth Low Energy (BLE) connection. Gold Standard Reference: An Eq. 02 Lifemonitor (Equivital Eq. 02; Hidalgo Ltd., Cambridge, UK) [26] with a sampling rate of 256 Hz served as the validation benchmark This device, independent of the primary wireless acquisition network, simultaneously recorded ECG and acceleration directly to its internal flash memory to ensure an unbiased reference. Following the recording session, this data was downloaded to the host PC via a wired USB interface for offline synchronization analysis. Gold Standard Reference: An Eq. 02 Lifemonitor (Equivital Eq. 02; Hidalgo Ltd., Cambridge, UK) [26] with a sampling rate of 256 Hz served as the validation benchmark This device, independent of the primary wireless acquisition network, simultaneously recorded ECG and acceleration directly to its internal flash memory to ensure an unbiased reference. Following the recording session, this data was downloaded to the host PC via a wired USB interface for offline synchronization analysis. To ensure comparability, the Movella DOT was physically affixed to the Equivital unit and aligned along the same coordinate axes. Data acquisition framework and protocol A custom Python-based software [27] managed the simultaneous data streams according to the procedure in Pseudocode 1. For the CGX AIM and Movella DOT, the software captured incoming data packets and immediately paired the sensor’s internal timestamp with the PC system time. Conversely, the Equivital device operated autonomously; its data was downloaded post-session, ensuring its internal timekeeping remained independent of the active synchronization process. The validation protocol consisted of six 10-minute phases. Each phase included a resting period followed by ten jumps. These jumps provided high-dynamic acceleration peaks necessary for precise cross-correlation between the Movella DOT and the Equivital reference. Fig. 2Synchronization Validation Scheme: Physiological data (CGX AIM) and acceleration data (Movella DOT) are streamed via wireless Bluetooth interfaces to be synchronized. Equivital data (physiological and acceleration), logged locally and downloaded via a wired USB interface, serves as the reference for evaluating synchronization accuracy Synchronization Validation Scheme: Physiological data (CGX AIM) and acceleration data (Movella DOT) are streamed via wireless Bluetooth interfaces to be synchronized. Equivital data (physiological and acceleration), logged locally and downloaded via a wired USB interface, serves as the reference for evaluating synchronization accuracy Offline synchronization and mapping Synchronization was finalized post-acquisition using MATLAB R2024b [28] via a linear regression model. First, the clock mapping procedure aligned the CGX AIM and Movella DOT timestamps with the recorded PC timestamps using linear correlation. This produced a scaling factor (clock skew) and an offset for each device. Second, the temporal alignment step mapped all data streams into the common PC time domain. To integrate the gold standard, the Equivital’s acceleration signal was cross-correlated with the pre-synchronized Movella DOT acceleration. This allowed the Equivital’s internal ECG data to be mapped into the same PC time domain, enabling a direct comparison between the gold standard and the synchronized signals. Performance evaluation System stability was tested over a one-hour session with a single healthy subject. This choice provides a technical proof-of-concept focused on the synchronization framework’s accuracy rather than clinical variability. This controlled environment allowed for the isolation and characterization of the methodology’s technical performance under stable conditions. Synchronization efficacy was quantified by comparing R-peak occurrence times in the ECG signals of the CGX AIM and the Equivital. R-peaks were identified using a modified Pan-Tompkins algorithm [29], and the temporal agreement between the two devices was analyzed using Bland–Altman plots to calculate the mean difference and standard deviation of R-peak timing across the hour-long session. Results High accuracy was achieved in the offline synchronization process, which employed linear regression to map device-specific timestamps to PC timestamps where strong correlations between device and PC timestamps were evident in the linear regression models. Specifically, the Root Mean Squared Error (RMSE) values, which quantify the prediction accuracy of these models, were 7.5 ms for the CGX AIM and 83.2 ms for the Movella DOT. These low RMSE values indicate a precise mapping of device timestamps to the PC time domain. Following the alignment of acceleration data from the Movella DOT and Equivital devices, the continuous one-hour acquisition was segmented into six discrete, stationary resting windows of 600 s each. This segmentation allowed for the evaluation of synchronization stability at different chronological stages of the long-duration recording. Within each of these resting windows, a comparison was made of the R-peak occurrence times between the CGX AIM and Equivital ECG signals. Figure 3 illustrates this synchronization procedure, displaying an example of 8-second window comparing raw ECG data from CGX AIM and ECG data from Equivital devices. Visual evaluation indicates good alignment of R-peaks after both 10 and 60 min, with a minor, quantifiable time delay. Fig. 3Visual comparison of ECG synchronization between CGX AIM and Equivital devices after 10 min (left) and 60 min (right) of acquisition Visual comparison of ECG synchronization between CGX AIM and Equivital devices after 10 min (left) and 60 min (right) of acquisition To statistically quantify the precise delay across these intervals, Bland–Altman analysis was performed for each window. Figure 4 displays a representative Bland–Altman plot for a single 600-second resting window, demonstrating that the R-peak timing differences remain stable and centered, proving that the linear regression model successfully and continuously compensated for clock drift throughout the session. Fig. 4Bland–Altman plot showing the differences in R-peak occurrence times between the physiological monitoring system (CGX AIM) and the gold standard reference (Equivital) during a single, representative 600-second resting window Bland–Altman plot showing the differences in R-peak occurrence times between the physiological monitoring system (CGX AIM) and the gold standard reference (Equivital) during a single, representative 600-second resting window The detailed statistical metrics for all six analyzed intervals are summarized in Table 2. The mean differences in R-peak occurrence times ranged from − 124.1 ms to 70.8 ms, reflecting minor, expected fluctuations due to localized hardware delays and variable detection sensitivity within the Pan-Tompkins algorithm. Despite this inter-window variability, the overall mean delay for the entire one-hour acquisition was only 20.1 ms. This result suggests good overall performance, as it represents only a 0.03% difference over the entire acquisition period. Standard deviations ranged from 7.8 ms to 37.4 ms, indicating varying consistency in synchronization. Notably, windows 3 and 4 exhibited lower standard deviations, suggesting more consistent synchronization during those periods. These inter-window differences in R-peak occurrence times can likely be attributed to variations in device acquisition mechanisms. Each device’s hardware may introduce minor delays, and automatic internal clock adjustments during acquisition can lead to uncontrollable divergences in absolute time values. Importantly, because the Equivital device was not synchronized with the system but operated independently, the observed agreement demonstrates the robustness of the proposed methodology when validated against an unbiased gold standard. To provide a more robust evaluation of local precision, the frame-by-frame analysis demonstrated a local Mean Absolute Error (MAE) of 58 ms across the recording session. Looking at individual windows, tracking the local variability across these individual windows yielded standard deviations ranging from 7.8 ms to 37.4 ms, reflecting the localized operational stability of the mapping over time. Table 2Results of R-peak synchronization analysis, comparing the CGX AIM system and the Equivital reference system. Analysis was performed across six stationary windows, evaluating differences in R-peak occurrenceWindow Length [s]Mean of R-peaks occurrence time differences[ms]STD of R-peaks occurrence time differences[ms]Mean Absolute ErrorUpper limit [ms]Lower limit [ms]Window 1558-87,922,20,054-44,4-131,4Window 2 542 -12,115,00,03717,2-41,5Window 3 575 70,87,80,01986,155,5Window 4 491 36,014,10,12763,78,4Window 5 560 -3,420,10,05535,9-42,8Window 6 569 -124,137,40,054-50,8-197,4 Table 2Results of R-peak synchronization analysis, comparing the CGX AIM system and the Equivital reference system. Analysis was performed across six stationary windows, evaluating differences in R-peak occurrenceWindow Length [s]Mean of R-peaks occurrence time differences[ms]STD of R-peaks occurrence time differences[ms]Mean Absolute ErrorUpper limit [ms]Lower limit [ms]Window 1558-87,922,20,054-44,4-131,4Window 2 542 -12,115,00,03717,2-41,5Window 3 575 70,87,80,01986,155,5Window 4 491 36,014,10,12763,78,4Window 5 560 -3,420,10,05535,9-42,8Window 6 569 -124,137,40,054-50,8-197,4 Results of R-peak synchronization analysis, comparing the CGX AIM system and the Equivital reference system. Analysis was performed across six stationary windows, evaluating differences in R-peak occurrence Discussion A novel software-based synchronization methodology for multi-device data acquisition was developed and validated, enabling the precise temporal alignment of distributed sensors within a common PC time base. By utilizing a specific data acquisition procedure followed by offline linear regression, this approach effectively addresses the heterogeneous device challenge in biomedical engineering, where varying signal morphologies and sampling rates often complicate temporal alignement [8]. This is particularly relevant given the increasing need to validate emerging wearable technologies against gold standard systems [30, 31] and the requirement for multi-signal fusion to improve classification performance in physiological signal processing [13, 32]. While commercially available systems like Empatica [33] or Biopac [34] offer internal synchronization, their closed software architectures often prevent the integration of other external devices. Our methodology demonstrates significant improvements over existing techniques reported in the literature. Simple manual alignment [35] lacks reproducibility, while existing software-level cross-correlation methods often report mean delays around 130 ms under optimal conditions [11, 36]. In contrast, our results showed an overall mean difference of only 20.1 ms (a 0.03% difference over one hour) when validated against an independent gold standard. Furthermore, unlike hardware-level synchronization that requires dedicated input channels and signal generators [10, 22], our framework requires no additional hardware. It offers superior generalizability by allowing devices to collect entirely different signal types (such as ECG and acceleration) without requiring a redundant common signal across all sensors to perform alignment. The mapping accuracy of device timestamps to the PC domain was further validated by low RMSE values: 7.5 ms for the CGX AIM and 83.2 ms for the Movella DOT. Quantitatively, the timing errors handled by our framework stem from three primary sources: operating system context-switching overhead, API/SDK execution latency and wireless protocol transmission jitter. This variation explains the behavioral differences captured in our results: the CGX AIM demonstrated an RMSE of 7.5 ms, whereas the Movella DOT (bound by a lower 60 Hz streaming rate and extensive SDK buffering) exhibited a higher mapping error of 83.2 ms. The regression framework effectively treats these combined latencies as zero-mean statistical noise, extracting the true underlying hardware clock behavior from the transmission jitter. This is consistent with findings by Gilbert et al. [37], who noted that low-frequency IMUs exhibit inherent synchronization uncertainties (e.g., up to 40 ms at 25 Hz). Despite these hardware-imposed limits, the achieved temporal accuracy remains clinically relevant for ECG-based applications and reliable signal comparison. A deeper analysis of the results (Table 2) revealed fluctuations in agreement across the six resting windows. While Windows 2 and 5 exhibited the highest degree of synchronization (mean differences as low as -12.1 ms and − 3.4 ms), Windows 1 and 6 showed larger discrepancies (up to -124.1 ms). As noted by Kot et al. [38], this inter-window variability is typical in real-world multimodal acquisition and likely reflects minor hardware delays, automatic internal clock adjustments, and detection variability within the Pan-Tompkins algorithm. The consistency varied across windows, with standard deviations ranging from 7.8 ms to 37.4 ms, suggesting that synchronization stability can be influenced by time-varying device behavior over long-duration recordings. Regarding the specific implications for high-frequency signal correlation, such as R-peak alignment in heart rate variability studies, this local variability remains well within acceptable operational boundaries. Because the regression framework effectively isolates and compensates for the macro-level linear hardware clock drift over the entire session, these micro-level local fluctuations do not degrade or compromise the structural alignment of individual physiological cycles, confirming that the achieved temporal accuracy remains clinically relevant for reliable cross-modal signal comparison. The primary constraint of this methodology is its strict reliance on accessible device SDKs and the availability of raw internal timestamps or hardware-level sequential counters. If a manufacturer’s software ecosystem is entirely closed, or if a device completely lacks internal timing information, this synchronization framework cannot distinguish between hardware clock drift and variable network transmission latency, rendering alignment unfeasible. Furthermore, even with SDK access, this framework cannot detect internal, hardware-level “sensor-to-timestamp” processing latencies. If these onboard buffering or execution delays scale non-linearly under heavy processor loads, they introduce variable temporal offsets that cannot be recovered via software regression. This represents a fundamental hardware-enforced constraint inherent to proprietary biomedical devices. However, when compared to alternative open-source approaches such as the LSL protocol, our framework offers a more universal and lightweight solution for low-level hardware integration. While LSL requires devices to either natively support its communication protocol or rely on an architectural layer capable of streaming data through specific LSL functions, our method only requires access to raw data and any basic internal timestamp accessible via a standard SDK. This significantly simplifies the integration of a wider variety of commercial, low-level sensors that are not inherently LSL-ready. It is also valuable to contextualize this methodology against established packet-based network synchronization standards, such as the NTP or IEEE 1588 PTP. While network-layer protocols like PTP can achieve sub-microsecond synchronization precision, they require low-level access to the operating system network stack and specialized hardware-level timestamping support within the network interface card. Such requirements are inherently incompatible with commercial, closed-source wearable biomedical sensors, which lack Wi-Fi or Ethernet interfaces and communicate strictly via proprietary Bluetooth stacks. Rather than attempting to force real-time clock synchronization at the network or hardware layer, our framework operates entirely at the application software layer. By allowing individual device clocks to drift naturally during data collection and reconstructing their timelines post-acquisition via software regression, our approach bypasses the architectural barriers of closed commercial SDKs. Furthermore, the current framework is fundamentally designed for offline post-processing synchronization. This architectural choice was made because linear regression utilizes the global dataset across the entire duration of the session to compute an optimal, stable fit for clock skew and offset. Crucially, this fundamental assumption of linear clock skew is optimized and highly stable for standard recording durations, such as the one-hour session evaluated in this study. For continuous, long-term monitoring applications extending across multi-day protocols, potential non-linear thermal drift driven by ambient temperature fluctuations or device heat dissipation could manifest, which may necessitate more complex modeling paradigms, such as higher-order polynomial regressions or adaptive filtering. For closed loop or real-time systems, extending this methodology would introduce distinct challenges, predominantly the computational lag of continuous regression updating and the lack of future temporal data points. To transition this framework to real-time applications, future iterations could implement a sliding-window algorithm or an adaptive filter to update clock coefficients dynamically as data packets arrive. Another limitation of the current validation scope is its execution on a single healthy subject within a controlled environment. While this setup served as an essential technical proof-of-concept to isolate algorithmic precision from biological and inter-subject variability, real-world deployments involving multiple subjects and highly dynamic movements could introduce additional challenges, including higher rates of wireless packet dropouts and prolonged transmission latencies. These aspects will need to be specifically investigated in future validation studies. Nevertheless, the high accuracy achieved between these multimodal devices establishes a robust foundation for scalable, software-defined sensor networks that eliminate the need for external hardware or software synchronization references, simplifying device integration for future biomedical applications. Conclusion This study successfully developed and validated a high-level software based synchronization methodology for multimodal sensor networks. By mapping heterogeneous device timestamps to a common PC time domain using linear regression, the framework achieves high temporal accuracy without requiring specialized synchronization hardware or shared reference signals between devices. Validation against an independent gold standard demonstrated an overall mean synchronization delay of 20.1 ms over a continuous one-hour acquisition period. While some inter window variability was observed, likely due to internal device clock adjustments and sampling rate constraints, the results remain well within the acceptable range for clinical physiological monitoring and motion analysis. The implications of this work are significant for the advancement of wearable health technology, multimodal signal monitoring, and the management of distributed sensor networks within the bioengineering field. By eliminating the requirement for shared signal morphologies or physical trigger cables, this methodology enables the integration of multimodal sensors into a unified acquisition platform. This scalability paves the way for more complex frameworks, including Body Area Network applications and the validation of novel biomedical devices. Future work will focus on testing the framework robustness across larger populations and more dynamic environmental conditions. Additionally, by adding more multimodal devices to the network, it will be possible to further assess the project validity and utility in real world clinical monitoring.