Accelerating Machine Learning in Healthcare: Addressing the Labelling Bottleneck

Accelerating Machine Learning in Healthcare: Addressing the Labelling Bottleneck | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Accelerating Machine Learning in Healthcare: Addressing the Labelling Bottleneck Spencer Vecile, William Dixon, Azadeh Assadi, Michael Kim, Robert Greer, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8194963/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 11 You are reading this latest preprint version Abstract Timely detection of postoperative arrhythmias after cardiac surgery is essential for preventing hemodynamic compromise. Machine learning models remain constrained by the scarcity of labeled datasets that reflect real-world monitoring practices. Existing approaches rely on adult 12-lead electrocardiograms which are rarely used continuously in pediatric ICUs and fail to capture age-dependent waveform variability. We present a clinically integrated labeling framework designed to overcome this bottleneck. Leveraging a physiologic waveform repository comprising over 1.6 million hours of unlabeled, continuous lead II ECG from more than 9,000 pediatric patients, we implemented a multi-phase strategy combining retrospective data mining, clinician-in-the-loop annotation, and active learning techniques, including uncertainty sampling and embedding-based retrieval. Initial labeling from MUSE (GE Healthcare) studies and ICU observations produced 154.9 hours of annotated ECG waveforms but required extensive clinician effort and yielded limited inter-patient variability. These two strategies provided sufficient coverage to train a preliminary classifier, enabling representation-aware sampling that dramatically improved efficiency. Embedding-guided retrieval achieved a precision of 60.2% for junctional arrhythmias and increased patient diversity compared to clinician in the loop-based labeling, while reducing annotation time per positive segment. Using this approach, we curated 189.2 hours of expert-labeled ECG from 1,447 unique patients, enriched for junctional arrhythmias, the primary modeling target. This work addresses a critical barrier to pediatric machine learning development and establishes a scalable methodology for creating clinically relevant datasets at scale, paving the way for real-time, clinician-augmented decision support systems in pediatric critical care. Arrhythmia detection Machine Learning Dataset Labeling Electrocardiogram (ECG) Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Introduction Arrhythmias are a common complication following pediatric congenital heart surgery and are associated with increased morbidity and mortality[ 1 ]. Certain arrhythmias, such as junctional ectopic tachycardia (JET), can be particularly harmful, potentially resulting in hemodynamic instability and circulatory collapse if not identified and treated in a timely manner[ 1 ]. Delays in diagnosis may arise from a range of factors, including limited experience among front-line providers, systemic constraints such as provider availability, and other operational challenges. Machine learning (ML)-based clinical decision support (CDS) systems have emerged as a promising tool to improve the efficiency of arrhythmia detection to facilitate prompt diagnosis and intervention. Most existing ML models have been developed using adult datasets and rely heavily on 12-lead electrocardiogram (ECG) recordings[ 2 ] which has limited utility in pediatric populations. These models are not well-suited to the pediatric intensive care unit (ICU) setting, where arrhythmia detection is typically initiated by clinicians using bedside monitors that display only one to three ECG leads. The 12-lead ECG is generally reserved for confirmation after an arrhythmia is suspected. Furthermore, age-dependent variations in ECG waveform characteristics, including heart rate, R-R intervals, and amplitude limit the applicability of adult data for pediatric ML-based CDS[ 3 ]. Existing research often depends on publicly available datasets like MIT-BIH[ 4 ], PTB-XL[ 5 ], PhysioNet 2021[ 6 ], and MIMIC-IV-ECG[ 7 ], which focus on adult populations and 12-lead recordings, making them unsuitable for detecting JET in children. Prior studies[ 8 – 10 ] have trained models using 12-lead or multi-lead ECG data, which typically requires prior clinical suspicion and manual acquisition. In contrast, our approach utilizes continuous lead II ECG monitoring, which reflects standard ICU practice where pediatric patients rarely have all 12 leads attached. This strategy also enables real-time arrhythmia detection without relying on clinician-initiated ECG acquisition. A critical barrier to developing effective ML models for pediatric arrhythmia detection is the lack of large, labeled datasets that reflect the clinical reality of continuous 3-lead monitoring[ 11 ]. The reliance on 12-lead ECGs in research is often driven by their association with confirmed diagnoses, which simplifies retrospective labeling. However, this approach does not capture the continuous data streams used in practice. To address this gap, the present study outlines a novel methodology for constructing a labeled pediatric ECG dataset suitable for training ML models. This paper describes a scalable and clinically informed strategy to extract meaningful labels from an extensive, unlabelled AtriumDB[ 12 ] dataset, with the goal of enabling the development of robust and deployable ML models for pediatric arrhythmia detection. Methods AtriumDB: Physiologic Database We developed two integrated platforms, AtriumDB and AtriumStreams (see Fig. 1 ), to enable both large-scale physiologic data capture and real-time model deployment. AtriumDB is a high-performance timeseries database specifically designed for use with physiological waveforms[ 12 ] and has been operational in 42 bedspaces at The Hospital for Sick Children (SickKids) ICU since 2016. It has supported the collection of over 1.6 million hours of continuous lead II ECG data from more than 9,000 critically ill children in the ICU and operating rooms leading to one of the largest pediatric physiologic waveform datasets in the world. AtriumDB includes a dedicated python library that enables rapid data extraction at scale, specifically designed to support high-throughput deep learning GPU training workflows. Complementing this, AtriumStreams is a real-time ingestion and deployment engine that ingests real time waveforms directly from bedside monitors, enabling real-time inference at the bedside. Together, AtriumDB and AtriumStreams function as the foundation of our broader translational pipeline. This system has now been deployed at 10 international hospitals in both pediatric and adult settings (as of Sept 2025). Label Schema To support dataset labeling, we developed a hierarchical label schema (Fig. 2 ) grounded in pediatric electrophysiology practice and iteratively refined through expert input. Labels were applied to lead II ECG strips from the bedside monitor as described in more detail below. The schema organizes rhythm types into five clinically meaningful parent categories: Sinus Rhythms, Blocks, Paced Rhythms, Ventricular Arrhythmias, and Supraventricular (SVT) Arrhythmias. Each parent class includes a set of child labels representing specific rhythm diagnoses encountered in pediatric critical care. The “unknown” class is excluded from the training data and is available for when samples aren't exactly noise but clinicians also can't confidently determine the rhythm class. Initial Labelling Workflow Ground Truth Labels All ECG labels used for training and evaluation were annotated by a team of pediatric intensive care clinicians with specialized cardiac training operating under the direction of a senior pediatric electrophysiologist. Given the limited availability of expert clinicians, we adopted a distributed annotation strategy to maximize efficiency, limiting labelling of each individual ECG segment to a single labeller to avoid overlapping labels requiring adjudication. When uncertainties arose, annotators maintained active communication with the senior electrophysiologist to resolve ambiguities and uphold labelling integrity. As such, the resulting annotations from each labeller were accepted as ground truth. Labelling Software To train our arrhythmia classifier, which assigns rhythm labels to 5-second segments of lead II ECG, we developed a clinically integrated labelling pipeline built around Label Studio, an open-source annotation platform (Fig. 3 ). The interface was custom-designed with direct input from pediatric electrophysiologists, enabling users to securely log in and label de-identified waveform data through a web-based environment. Label Studio’s native support for time-series data, combined with its customizable HTML framework, allowed us to tailor the interface for efficient ECG annotation at scale. Our interface presented both a 2–5 minute lead II ECG waveform strip and synchronized heart rate signal (sampled at 1 Hz). If a 12-lead ECG was obtained with a confirmed diagnosis in the MUSE system with a timestamp corresponding to the telemetry strip in question, the annotation was included as contextual metadata on the labelling interface. Label selection occurs through a structured dropdown menu derived from the hierarchical schema (Fig. 2 ). The hierarchical schema supports both coarse and fine-grained annotation. Annotators can apply a parent label (e.g., “Ventricular Arrhythmia”) or designate more specific subtypes via child labels. This structure enabled flexible model training strategies (e.g. grouping rare subtypes under a parent class to mitigate class imbalance). During annotation, users first selected a predominant rhythm for the full waveform and then mark specific regions (e.g., noise artifacts or other arrhythmias) where applicable. For instance, a 2-minute segment may be labelled globally as Sinus Rhythm, with one or more short X-second regions identified as another rhythm, which significantly reduces annotation time compared to fully segmented labelling. To mitigate performance constraints inherent to Label Studio, which exhibits latency and interaction delays beyond approximately 300,000 data points, or 10 minutes of ECG sampled at 500 Hz, we adopted a segmentation strategy that partitioned longer waveform regions into contiguous 2–5 minute chunks. These segments were grouped as sequential tasks, with consistent annotations in the user interface denoting patient identity and temporal ordering. This approach preserved critical clinical context while maintaining interface responsiveness. To maintain synchronization with our physiologic data platform, we implemented webhook functionality within Label Studio. These webhooks notify a custom Flask server each time a label is created, modified, or deleted, enabling real-time updates to an AtriumDB dataset and supporting rapid retraining cycles as new labelled data becomes available. This continuous feedback loop between expert labelling and model development is central to our clinician-in-the-loop workflow and ensures that improvements in the dataset immediately translate into model refinement. Summary of Different Label Sources To construct a high-quality training dataset for the arrhythmia classifier, we implemented a multi-pronged strategy to source labels from both retrospective and prospective data streams. Given the rarity and subtlety of JET, no single labelling approach was sufficient. Instead, we leveraged four complementary techniques to enrich the dataset with arrhythmias of interest: MUSE: Structured diagnoses associated with 12-lead ECG studies. ICU Observations: Bedside annotations triggered in real time by clinical staff. Uncertainty Sampling: Model-guided selection of ambiguous or low-confidence windows. Embedding Search: Retrieval of morphologically similar ECG segments via model-derived feature embeddings. Each of these strategies yielded different levels of labelling efficiency, patient coverage, and arrhythmia density. The general principle was that we used a “known” anchor (e.g., from MUSE or an ICU observation) to pull surrounding “unknown” ECG waveforms likely to have a similar label. Sourcing Labels From MUSE Our initial strategy for sourcing arrhythmia labels used the presence of a formal 12-lead ECG study as a proxy indicator for an arrhythmia event. Specifically, we extracted all 12-lead studies from the MUSE system at SickKids and aligned them with continuous ECG waveform data stored in an AtriumDB dataset by matching timestamps. The diagnostic interpretation associated with each 12-lead study, typically authored by an electrophysiologist or generated via rule-based algorithms, was attached to the corresponding ECG segment and presented to expert labellers in Label Studio for verification. However, this approach introduced two critical challenges. First, we frequently observed discrepancies between the diagnostic interpretation derived from the 12-lead ECG and the rhythm classification provided by expert annotators based on the corresponding waveform segments. Combined with our prior findings that ICU clocks are often significantly desynchronized[ 13 ], this suggests that the MUSE system timestamps are misaligned with those of the Philips monitoring infrastructure to a degree that precludes reliable temporal matching. Consequently, the ECG strips presented to annotators often did not correspond precisely to the diagnostic moment captured by the 12-lead system, thereby compromising the validity of the associated diagnostic labels. These misalignments necessitated manual re-labeling and rigorous quality assurance procedures, substantially increasing the annotation burden. Second, even when correctly aligned, the segments sourced via 12-lead studies were predominantly composed of normal sinus rhythm (85% in our cohort). While sinus rhythm examples are essential for training a balanced classifier, the disproportionate prevalence of non-informative segments significantly limited the efficiency of expert labelling efforts. Given the constraints on clinician time, it became clear that this method was not viable as a primary mechanism for surfacing high-value arrhythmia examples. Sourcing Labels From ICU Observations To surface clinically relevant arrhythmia examples in real time, we implemented a clinician-in-the-loop annotation workflow that leverages the situational awareness of ICU providers. Initially deployed via Slack, a secure mobile messaging platform, the system allowed nurses and physicians to rapidly report observed arrhythmias through a structured form (Fig. 4 ). Designed for minimal workflow disruption, the form required only two fields, patient room number and event timestamp, while offering optional fields such as arrhythmia category (restricted to parent-level classes) and free-text clinical notes. These optional fields were later used to filter segments in Label Studio by rhythm class. The form also included a “From admission start” checkbox, enabling clinicians to estimate timing without needing to retrieve the precise encounter start time, further reducing friction. Upon submission, the input was automatically transmitted to a Flask-based server, which queried the dataset and extracted two segments: a one hour window following the reported start time and a 20-minute window at the end of the encounter, which typically contained Sinus Rhythm. This dual-segment retrieval strategy was designed to provide the model not only with positive examples of arrhythmias but also with patient-specific sinus rhythm morphology, facilitating finer-grained contrastive learning. In cases where the patient encounter had not yet ended at the time of submission, a background scheduler re-queried the dataset every five minutes until the terminal segment became available and could be uploaded for labelling. This flexible, low-friction reporting mechanism enabled bedside clinicians to contribute high-yield labelling candidates with minimal interruption to care, significantly increasing the number of rare rhythm examples, particularly junctional arrhythmias, compared to passive approaches. The integration of annotation triggers into tools already used during clinical care not only enhanced data relevance but also improved labeller efficiency by reducing the proportion of sinus rhythm segments. Following completion of ICU-based labeling, the resulting dataset contained approximately 80 minutes of annotated ECG data per patient but was limited by a relatively small cohort size. Figure 5 illustrates the clustering of all labeled segments, where embeddings from the same patient exhibit tight grouping, indicating minimal intra-patient variability within rhythm classes. While labeling one hour per patient provided valuable initial coverage, this approach is highly time-consuming for clinicians and yields diminishing returns because ECG signals are inherently repetitive within a single patient. Nevertheless, it was essential for training an initial model capable of supporting model-guided labeling. To improve generalization and reduce annotation burden moving forward, we aim to increase inter-patient diversity by labeling shorter segments (approximately two minutes) from a larger number of unique patients. This strategy maximizes variability while minimizing clinician workload. To operationalize this approach, we adopted a model-guided labeling framework. Model Guided Labelling Workflow The final two labeling strategies leverage a machine learning model trained on data obtained through the initial labeling methods to identify additional candidate segments within the dataset. Although the model does not yet demonstrate sufficient performance for clinical deployment, it is capable of highlighting regions within the unlabeled data that are likely to contain arrhythmias. This enables a targeted labelling approach, focusing on rhythm classes the model finds challenging or are underrepresented in the current dataset. To contextualize these strategies, we first provide a summary of the model architecture and training process. Further technical details and model performance metrics will be presented in a subsequent publication. Foundational Model The foundation model was trained on continuous lead II ECG data collected from the SickKids ICU. It uses a Wave2Vec 2.0-based[ 15 ] architecture, which includes a 4-layer convolutional encoder followed by 12 transformer blocks configured similarly to BERT-Base[ 16 ]. Training was performed using Contrastive Masked Segment Coding (CMSC)[ 9 ], a self-supervised learning approach that combines local and global contrastive objectives. Locally, the model learns to reconstruct masked ECG segments from the surrounding context, while globally it learns patient-level rhythm patterns by comparing adjacent segments from the same patient against segments from others. This enables the model to learn temporally-rich and morphology-aware representations from unlabelled data. Arrhythmia Classifier To adapt the foundation model for supervised classification, self-supervised components such as masking and contrastive losses were removed. A lightweight classification head was added, consisting of temporal average pooling and three fully connected layers, followed by a softmax output over four rhythm classes: Junctional Arrhythmia, Sinus Rhythm, Other Rhythm, and Noise (see Fig. 6 ). These classes were selected to reflect clinically relevant distinctions and to support the detection of junctional arrhythmias in real-world ICU settings. Training was conducted using the expert-labelled lead II ECG segments obtained from the previously described methods. The model was optimized using weighted binary cross-entropy loss to address class imbalance. Physiologically realistic augmentations were employed such as powerline interference, EMG noise, and baseline wander; to improve robustness under ICU conditions. Standard normalization and interpolation techniques were applied to insure signal consistency and handle missing data. The classifier’s softmax outputs were used to identify low-confidence predictions for uncertainty sampling, enabling the selection of ambiguous segments near the decision boundary. Additionally, the 256-dimensional embeddings from the final hidden layer were used for embedding-based sampling, allowing the retrieval of morphologically similar but diverse examples from the unlabeled dataset. These capabilities enabled targeted mining of an AtriumDB dataset, improving label efficiency and enhancing model generalization. Sourcing Labels From Uncertainty Sampling Uncertainty sampling is a widely used technique in active learning, wherein a model is used to identify samples for which its predictions are least confident, those that lie near the decision boundary. These cases are often the most informative for training, as they expose the model’s limitations and guide learning toward harder examples. In our context, the combination of labelled data from MUSE-derived 12-lead studies and clinician-reported ICU observations provided enough coverage to train a preliminary arrhythmia classifier. While limited in performance, this initial model was sufficient to enable active mining of the unlabeled ECG data dated back to 2016. We applied the preliminary model to the full Lead II ECG dataset and selected ECG segments based on model uncertainty. For the initial round, we extracted 2-minute segments from 25 unique patients where the model's softmax score for the Junctional Arrhythmia class fell between 40% and 65%, a range indicative of ambiguous, borderline predictions. These cases were then pushed to Label Studio for expert labelling. This process was repeated iteratively, with each round expanding the number of sampled patients and enriching the training dataset with ECG morphologies the model found most difficult to classify. As the model improved across successive iterations, the segments retrieved through uncertainty sampling shifted in character. Increasingly, the uncertain windows contained low signal-to-noise ratios or atypical morphologies that were difficult to classify, even for human experts. This inflection point marked the diminishing returns of uncertainty-based sampling. Sourcing Labels From Embeddings In contrast to uncertainty sampling, embedding-based sampling is a form of representation-aware active learning that leverages the model’s internal structure to surface underrepresented but high-confidence samples. Rather than querying the decision boundary, embedding search targets the interior of a known class, retrieving morphologically similar rhythms from new patients. This strategy is particularly effective in the later stages of model development, when the classifier has already learned a stable representation of key rhythm classes but requires additional diversity to generalize across patients. Here, our goal was to further enrich the junctional arrhythmia class by identifying previously unseen morphologic variants across the population, while reducing annotation burden and avoiding redundant examples. Figure 7 represents a high-level overview of the embedding search technique. To achieve this, we applied our arrhythmia classifier to the entire dataset and extracted 256-dimensional embeddings from the final hidden layer of the network for every 5-second ECG segment. Using the set of previously labelled junctional arrhythmia segments, we computed a global centroid in this high-dimensional embedding space, as well as per-patient centroids for each labelled junctional arrhythmia patient. We then defined a hypersphere centred at the global centroid, with a radius equal to the mean plus one standard deviation of the distances between each patient’s centroid and the global centre. This hypersphere captured the typical distribution of junctional arrhythmia morphology as represented by the model and defined our search region for new candidate segments. Within this embedding-defined region, we searched for 2-minute segments of ECG that exhibited low “embedding variability”, that is, whose 24 constituent 5-second windows were tightly clustered in embedding space. This constraint was empirically found to increase the likelihood that a given 2-minute segment contained only one class, thereby reducing the cognitive load on annotators and increasing labelling throughput. For each selected patient, we also extracted an additional 1-minute segment from the end of their encounter to serve as a likely sinus baseline, supporting the model’s ability to learn intra-patient morphology contrasts. Our first batch of embedding-sampled segments yielded data from 294 new patients. After expert annotation, 60.2% of segments were confirmed as junctional arrhythmias, with 18.7% sinus, 3.4% noise, and 17.7% representing other arrhythmias, highlighting the high precision and label efficiency of the approach. Importantly, the remaining 39.8% of non-junctional segments were also valuable, as they represent cases where the model positioned embeddings close to the junctional cluster. Incorporating these examples during retraining enhanced the model’s ability to distinguish junctional arrhythmias from morphologically similar rhythms. As shown in Fig. 8 , the UMAP projection of the model’s embedding space illustrates the structure of the learned representation. This visualization confirms that junctional arrhythmias occupy a coherent and distinct region in embedding space, separable from other classes. The right plot focuses on that junctional region, showing a subset of previously labelled junctional arrhythmia examples alongside unlabeled segments sampled from within the search hypersphere. These are the candidate samples identified for expert annotation via Label Studio. Six representative ECG strips from six unique patients are also shown in Fig. 5 , highlighting the substantial inter-patient variability in junctional arrhythmia morphology. Results Labeled Dataset The final labelled dataset created from the above techniques consists of 189.2 hours of expert-annotated lead II ECG from 1,447 unique pediatric patients, encompassing four high-level rhythm classes: Junctional Arrhythmia, Sinus Rhythm, Other Rhythm, and Noise. Table 1 summarizes the total number of labelled hours and unique patients obtained through each method, as well as the subset of those labels containing Junctional arrhythmias, our primary modelling target. Table 1 Summary of label yield by sourcing strategy All Labels Junctional Arrhythmia Labels Labeling Method Hours Patients Hours Patients MUSE 75.72 991 5.75 97 ICU Observations 79.21 47 11.23 18 Uncertainty Sampling 12.91 96 3.14 37 Embedding Search 21.36 408 7.63 214 Total 189.2 1447 27.77 358 To ensure demographic and physiologic diversity, the dataset was stratified by both sex and age group, capturing a representative distribution across the pediatric population from neonates to late adolescents. As shown in Table 2 , the dataset includes contributions from both male and female patients, while Table 3 further details the label distribution across age groups. Table 2 Distribution of labelled ECG hours and patient counts across rhythm classes, stratified by sex Male Female Hours Patients Hours Patients Hours Patients Junctional Arrhythmia 27.77 358 14.86 (53.5%) 203 12.91 (46.5%) 155 Sinus Rhythm 96.63 1080 59.49 (61.6%) 594 37.14 (38.4%) 486 Other Rhythm 46.17 315 25.84 (56%) 176 20.33 (44%) 139 Noise 18.63 647 12.21 (65.5%) 360 6.42 (34.5%) 287 Total 189.2 1447 112.39 (59.4%) 797 76.81 (40.6%) 650 Table 3 Distribution of labelled ECG hours and patient counts across rhythm classes, stratified by pediatric age group Neonate (0–28 Days) Infant (29 Days-1 Year Toddler (1–3 Years) Child (4–12 Years) Adolescent (13–18 Years) Hours Patients Hours Patients Hours Patients Hours Patients Hours Patients Junctional Arrhythmia 3.09 (11.1%) 51 13.15 (47.5%) 144 1.87 (6.7%) 31 9.61 (34.7%) 134 3.09 (11.1%) 51 Sinus Rhythm 19.44 (20.2%) 205 36.86 (38.4%) 318 9.26 (9.6%) 113 30.47 (31.7%) 483 19.44 (20.2%) 205 Other Rhythm 15.04 (32.6%) 78 22.14 (48%) 116 2.44 (5.3%) 37 6.45 (14%) 92 15.04 (32.6%) 78 Noise 3.63 (19.5%) 168 4.25 (22.9%) 189 5.99 (32.2%) 64 4.74 (25.5%) 249 3.63 (19.5%) 168 Table 4 Comparison of labeling time requirements by label acquisition strategy. The results are presented using two key time constructs: Lead time and Task duration . Lead time refers to the annotator’s active labeling time per task. Outliers in lead time were removed because they typically occur when an annotator steps away and leaves a task open, which would otherwise inflate time estimates and misrepresent efficiency. Task duration represents the length of ECG content presented in each task, while lead time per minute of task normalizes effort to the amount of ECG reviewed. Labeling method Tasks analyzed Task duration (s) Median lead time (s) Lead time per minute of task (s/min) Tasks per annotator hour ECG minutes per annotator hour Total annotator hours Total ECG hours labeled MUSE 1772 120 54 27.07 37 73 48.43 59 ICU observations 598 300 33 6.57 33 165 18.29 50.23 Uncertainty Sampling 207 300 65 13.1 31 124 6.59 13.62 Embedding Search 550 120 41 24.34 58 149 9.53 23.67 Discussion Developing clinically deployable machine learning models for pediatric arrhythmia detection requires overcoming the labeling bottleneck inherent to large-scale physiologic datasets. This study demonstrates that a multi-strategy labeling pipeline combining retrospective data mining, clinician-in-the-loop annotation, and active learning can accelerate dataset creation while preserving clinical fidelity. We began with two initial strategies: extracting labels from MUSE 12-lead electrocardiogram studies and real-time ICU observations. While MUSE provided broad coverage it was inefficient; it yielded predominantly sinus rhythm data and required extensive manual verification. ICU observations, although limited in patient count, produced high-yield arrhythmia segments with minimal workflow disruption. With ICU observations we took one hour of arrhythmia data from each patient and 20 minutes of possible normal sinus rhythm. This was an arbitrary amount of time we selected in the beginning however we have now changed that to 20 minutes of arrhythmia and 10 minutes of sinus. Labelers spent 58% of the total time labeling on MUSE tasks which was our lowest yield method of finding arrhythmias. If repeated, we would prioritize ICU observations from the outset, as they offered the most clinically relevant data with the least annotation overhead helping us get to model-guided annotation faster. The MUSE and ICU observations provided sufficient labeled data to train a preliminary classifier that could generate embeddings for active learning. This model enabled uncertainty sampling and embedding-based retrieval, which dramatically improved label efficiency and inter-patient variability. Embedding-guided sampling, in particular, enriched rare arrhythmia classes such as junctional arrhythmias while reducing annotation burden. These findings highlight the importance of representation-aware approaches in pediatric machine learning development, where physiologic variability across age groups and clinical contexts limits the utility of larger adult datasets. By leveraging the scale of the physiological waveform dataset accumulated at SickKids and integrating expert feedback loops, we created one of the largest labeled pediatric ECG corpora to date, enabling robust model training under real-world ICU conditions. Table 4 shows each strategy offered distinct advantages and trade-offs. MUSE and ICU observations required the greatest total annotator time, at approximately 48.4 hours and 18.3 hours, respectively. In contrast, uncertainty sampling and embedding-guided selection were far more efficient, producing the most diverse set of patients and arrhythmias while requiring 6.6 hours and 9.5 hours, respectively. These figures underscore that most labeling effort occurs before reaching model-guided techniques. Overall, approximately 80 hours of clinician time was utilized in dataset creation, which is relatively modest given the scale and clinical impact of this use case. However, clinician feedback emphasized that concentration declines after about an hour of continuous labeling, and we observed a corresponding drop in label quality. To mitigate fatigue, it was more effective to provide smaller batches (20 minutes to one hour of labeling) rather than assigning hundreds of tasks at once, improving engagement and consistency. When normalized by ECG content, ICU observations and uncertainty sampling were the most efficient, reflecting their 5-minute task durations. In contrast, MUSE data and embedding-guided selection, which used 2-minute windows, were slower to complete on a per-minute basis. Median lead times and tasks per annotator hour were broadly similar across methods, but longer windows provided richer context, making annotations easier. Based on these findings, we recommend task durations in the 2–5 minute range to balance efficiency with annotation quality. Conclusion This study addresses a key barrier to machine learning adoption in pediatric critical care by introducing a scalable, clinically informed labeling framework. Our approach accelerates data curation and establishes a generalizable methodology for domains where expert annotation is time-intensive and rare events dominate. Future work will focus on expanding label diversity and validating model performance in silent trials prior to bedside deployment. Ultimately, the combination of clinician in the loop and model-guided active learning represents a transformative step toward real-time, clinician-augmented machine learning systems capable of improving patient safety and outcomes. Declarations Ethics Approval and Consent to Participate The ECG dataset used in this study is part of a clinical AtriumDB-based dataset, collected prospectively from a 42-bed ICU at The Hospital for Sick Children (SickKids) in Toronto, ON since 2016. Secondary use of this dataset is governed by REB approval #1000068499 and has a waiver of consent. Consent for publication Not applicable. Availability of Data and Material Due to patient privacy and institutional data governance policies, the ECG waveform data used in this study cannot be publicly released. However, access to de-identified subsets of the dataset may be made available to academic researchers upon request and subject to ethics approval and data use agreements with The Hospital for Sick Children (SickKids). The code used to train the arrhythmia classifier is not publicly available. Qualified academic collaborators may request access to specific components for non-commercial research under a data use agreement. Competing Interests The authors declare no competing interests. Funding This work was supported by the Canadian Institutes of Health Research (CIHR) through a Project Grant (Application Number: 480831), The University of Toronto Data Science Institute Catalyst Grant, and The University of Toronto EMH Seed Grant. Authors' Contributions All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by S. Vecile and W. Dixon. ECG labeling was done by A. Bulic, A. Assadi, M. Kim, D. Ehrmann, D. Eytan and M. Mazwi. The first draft of the manuscript was written by S. Vecile. The manuscript was commented on and edited by S. D. Goodfellow, M. Kim, A. Goodwin, W. Dixon, A. Siddiqui, A. Assadi, D. Ehrmann and R. Greer. All authors read and approved the final manuscript. Acknowledgements Not Applicable. Clinical Trial Number Clinical trial number: not applicable. References Mayaya P-C, Tinică G, Chistol RO, Moraru L, Damian SI, Frăsinariu OE, et al. Incidence and Risk Factors for Early Postoperative Arrhythmias in Congenital Heart Disease – Systematic Review. J Cardiovasc Emergencies. Târgu Mureș, Poland: De Gruyter Brill Sp. z o.o., Paradigm Publishing Services; 2025;11:43–53. https://doi.org/10.2478/jce-2025-0005 Pantelidis P, Bampa M, Oikonomou E, Papapetrou P. Machine learning models for automated interpretation of 12-lead electrocardiographic signals: a narrative review of techniques, challenges, achievements and clinical relevance. J Med Artif Intell [Internet]. AME Publishing Company; 2023 [cited 2025 Nov 10];6. https://doi.org/10.21037/jmai-22-94 Skippen PW, Sanatani S, Gow RM, Froese N. Diagnosis of postoperative arrhythmias following paediatric cardiac surgery. Anaesth Intensive Care. 2009;37:705–19. https://doi.org/10.1177/0310057X0903700506 Moody GB, Mark RG. The impact of the MIT-BIH Arrhythmia Database. IEEE Eng Med Biol Mag. 2001;20:45–50. https://doi.org/10.1109/51.932724 Wagner P, Strodthoff N, Bousseljot R-D, Kreiseler D, Lunze FI, Samek W, et al. PTB-XL, a large publicly available electrocardiography dataset. Sci Data. Nature Publishing Group; 2020;7:154. https://doi.org/10.1038/s41597-020-0495-6 Reyna MA, Sadr N, Alday EAP, Gu A, Shah AJ, Robichaux C, et al. Will Two Do? Varying Dimensions in Electrocardiography: The PhysioNet/Computing in Cardiology Challenge 202. Brno, Czech Republic; 2021. p. 1–4. https://doi.org/10.23919/CinC53138.2021.9662687 Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. Nature Publishing Group; 2023;10:1. https://doi.org/10.1038/s41597-022-01899-x Oh J, Chung H, Kwon J, Hong D, Choi E. Lead-agnostic Self-supervised Learning for Local and Global Representations of Electrocardiogram [Internet]. arXiv; 2022 [cited 2025 Oct 2]. https://doi.org/10.48550/arXiv.2203.06889 Kiyasseh D, Zhu T, Clifton DA. CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients [Internet]. arXiv; 2021 [cited 2025 Oct 2]. https://doi.org/10.48550/arXiv.2005.13249 McKeen K, Masood S, Toma A, Rubin B, Wang B. ECG-FM: An Open Electrocardiogram Foundation Model [Internet]. arXiv; 2025 [cited 2025 Oct 2]. https://doi.org/10.48550/arXiv.2408.05178 Shani C, Zarecki J, Shahaf D. The Lean Data Scientist: Recent Advances Toward Overcoming the Data Bottleneck. Commun ACM. 2023;66:92–102. https://doi.org/10.1145/3551635 Goodwin AJ, Eytan D, Greer RW, Mazwi M, Thommandram A, Goodfellow SD, et al. A practical approach to storage and retrieval of high-frequency physiological signals. Physiol Meas. IOP Publishing; 2020;41:035008. https://doi.org/10.1088/1361-6579/ab7cb5 Goodwin AJ, Dixon W, Mazwi M, Hahn CD, Meir T, Goodfellow SD, et al. The truth Hertz-synchronization of electroencephalogram signals with physiological waveforms recorded in an intensive care unit. Physiol Meas. 2023;44. https://doi.org/10.1088/1361-6579/ace49e McInnes L, Healy J, Saul N, Grossberger L. UMAP: Uniform Manifold Approximation and Projection. J Open Source Softw. 2018;3:861. https://doi.org/10.21105/joss.00861 Baevski A, Zhou Y, Mohamed A, Auli M. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. In: Larochelle H, Ranzato M, Hadsell R, Balcan MF, Lin H, editors. Adv Neural Inf Process Syst [Internet]. Curran Associates, Inc.; 2020. p. 12449–60. https://proceedings.neurips.cc/paper_files/paper/2020/file/92d1e1eb 1cd6f9fba3227870bb6d7f07-Paper.pdf Devlin J, Chang M-W, Lee K, Toutanova K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In: Burstein J, Doran C, Solorio T, editors. Proc 2019 Conf North Am Chapter Assoc Comput Linguist Hum Lang Technol Vol 1 Long Short Pap [Internet]. Minneapolis, Minnesota: Association for Computational Linguistics; 2019 [cited 2025 Oct 2]. p. 4171–86. https://doi.org/10.18653/v1/N19-1423 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 23 Feb, 2026 Reviews received at journal 22 Feb, 2026 Reviews received at journal 06 Feb, 2026 Reviewers agreed at journal 02 Feb, 2026 Reviewers agreed at journal 02 Feb, 2026 Reviews received at journal 24 Jan, 2026 Reviewers agreed at journal 29 Dec, 2025 Reviewers invited by journal 04 Dec, 2025 Editor assigned by journal 04 Dec, 2025 Submission checks completed at journal 02 Dec, 2025 First submitted to journal 24 Nov, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8194963","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":555700728,"identity":"a6a3ffb6-392a-4e72-8f0e-df800a6206c3","order_by":0,"name":"Spencer Vecile","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABEUlEQVRIie2QMUvDQBTH/6FwWQ6zXgg0X+FE6CTmq7wQiEvAQSgdSigE0qXgWnDwqxwEziXUtdAiEcHJoZNkKp7oZK+im8P9hrvHO3689z/A4fjvqA7nYGADEPutQsj/qIDQmNMo+EGJbyvd9ZNHyPvqSdHkITnxdSa7cYlgrqyK3OrL00V7DdlqqajdpDXPG6JVA9GSXRHFKPJqQrgsoNJ6Q0z4lSkUJOxKvLx6i7y9Ue5ejbJfJV9KCRl09jDrgkXejBAIbpSZ8mrBGkrrgVngyGJbPQoXmnjAc5NFZx9ZMmmycLG2TzE/9iL6KQ2Z3zzvdtOLJJ7rs7Afl8Pgxj4F4vPiBw+HnW+Kw+FwOI7zDm4wW0voxctfAAAAAElFTkSuQmCC","orcid":"","institution":"Hospital for Sick Children","correspondingAuthor":true,"prefix":"","firstName":"Spencer","middleName":"","lastName":"Vecile","suffix":""},{"id":555700729,"identity":"bf94f511-6ce2-4ac3-8337-ec670f7b05bf","order_by":1,"name":"William Dixon","email":"","orcid":"","institution":"Hospital for Sick Children","correspondingAuthor":false,"prefix":"","firstName":"William","middleName":"","lastName":"Dixon","suffix":""},{"id":555700730,"identity":"c02b0fd8-1225-4e41-897b-1513ce356838","order_by":2,"name":"Azadeh Assadi","email":"","orcid":"","institution":"Hospital for Sick Children","correspondingAuthor":false,"prefix":"","firstName":"Azadeh","middleName":"","lastName":"Assadi","suffix":""},{"id":555700732,"identity":"b2bd0598-be92-45a2-b51a-bd13c072e83d","order_by":3,"name":"Michael Kim","email":"","orcid":"","institution":"Seattle Children's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Michael","middleName":"","lastName":"Kim","suffix":""},{"id":555700734,"identity":"6a94f661-7c9f-4c56-9b48-f3caadb403b8","order_by":4,"name":"Robert Greer","email":"","orcid":"","institution":"Hospital for Sick Children","correspondingAuthor":false,"prefix":"","firstName":"Robert","middleName":"","lastName":"Greer","suffix":""},{"id":555700736,"identity":"ee7eb1de-7005-48e4-bfca-7f01701fea68","order_by":5,"name":"Asad Siddiqui","email":"","orcid":"","institution":"Hospital for Sick Children","correspondingAuthor":false,"prefix":"","firstName":"Asad","middleName":"","lastName":"Siddiqui","suffix":""},{"id":555700741,"identity":"7a2dfe76-754c-42fe-b77c-bb43300da227","order_by":6,"name":"Daniel Ehrmann","email":"","orcid":"","institution":"University of Michigan Medical School","correspondingAuthor":false,"prefix":"","firstName":"Daniel","middleName":"","lastName":"Ehrmann","suffix":""},{"id":555700746,"identity":"52be70eb-f80d-4700-8c55-417a1e6b90b5","order_by":7,"name":"Andrew Goodwin","email":"","orcid":"","institution":"Hospital for Sick Children","correspondingAuthor":false,"prefix":"","firstName":"Andrew","middleName":"","lastName":"Goodwin","suffix":""},{"id":555700749,"identity":"4c6cf74a-4b90-4853-a38f-0ac6d989488b","order_by":8,"name":"Danny Eytan","email":"","orcid":"","institution":"Rambam Health Care Campus","correspondingAuthor":false,"prefix":"","firstName":"Danny","middleName":"","lastName":"Eytan","suffix":""},{"id":555700752,"identity":"77aaacbe-dee5-41ea-ac26-4e931ded9648","order_by":9,"name":"Mjaye Mazwi","email":"","orcid":"","institution":"Seattle Children's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Mjaye","middleName":"","lastName":"Mazwi","suffix":""},{"id":555700754,"identity":"6a7f42a3-a9ef-4434-9e10-f6060fe7f324","order_by":10,"name":"Anica Bulic","email":"","orcid":"","institution":"Hospital for Sick Children","correspondingAuthor":false,"prefix":"","firstName":"Anica","middleName":"","lastName":"Bulic","suffix":""},{"id":555700762,"identity":"6d1cead8-b5e0-4a69-9949-99c3731506b5","order_by":11,"name":"Sebastian D. Goodfellow","email":"","orcid":"","institution":"Hospital for Sick Children","correspondingAuthor":false,"prefix":"","firstName":"Sebastian","middleName":"D.","lastName":"Goodfellow","suffix":""}],"badges":[],"createdAt":"2025-11-24 15:39:00","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8194963/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8194963/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97894701,"identity":"5a84c683-0dbf-44dd-9708-d0066c76a2e6","added_by":"auto","created_at":"2025-12-10 15:32:54","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5980275,"visible":true,"origin":"","legend":"","description":"","filename":"Labelingpaperfinal.docx","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/78149e8c967e24e65f46da71.docx"},{"id":97894356,"identity":"2c20a78a-540b-47b8-bc2f-011abad11e20","added_by":"auto","created_at":"2025-12-10 15:32:23","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12905,"visible":true,"origin":"","legend":"","description":"","filename":"4492c3023b7e40cd98ee0a41dd2af251.json","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/dccf8cc6667f83a5165edc1b.json"},{"id":97894637,"identity":"70e78231-79a7-42e5-9046-2eebee485a74","added_by":"auto","created_at":"2025-12-10 15:32:49","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":97812,"visible":true,"origin":"","legend":"","description":"","filename":"4492c3023b7e40cd98ee0a41dd2af2511enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/4c447ebc835ad5adc9bc1e68.xml"},{"id":97894898,"identity":"7d159f39-effe-418f-b3ee-3d3d4f3ffde3","added_by":"auto","created_at":"2025-12-10 15:33:11","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":385137,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/fc8eed0a4e0aa3a1d2a47c2f.png"},{"id":97702449,"identity":"8f1896b1-36af-4713-b9ef-fb84d3b53f94","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":286297,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/764326e1cc18abac7f06db57.png"},{"id":97892855,"identity":"5cc5fa81-698e-4c8f-903a-f23db997b28b","added_by":"auto","created_at":"2025-12-10 15:23:31","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":396079,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/41c295d2932dc1aa0eafd4df.png"},{"id":97892976,"identity":"c03ea015-5d59-4f26-a39f-002bdcd9088e","added_by":"auto","created_at":"2025-12-10 15:25:10","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":631679,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/fc6b814adc25d76600cc8959.png"},{"id":97702451,"identity":"e3447d10-8858-4808-b0cd-fa5ee5355e23","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":617428,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/ee8e39e9ccda21df1fa9db0c.png"},{"id":97895148,"identity":"923c9f68-8072-4a49-93c4-b299dd430e85","added_by":"auto","created_at":"2025-12-10 15:33:39","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":773736,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/07506b0b5220a50a1a1ce0f0.png"},{"id":97702458,"identity":"b39b220e-5130-49a1-8c4b-cb217cd25c73","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1198391,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/c2a0a7343153b691ef74468a.png"},{"id":97895207,"identity":"c81b612a-ebb5-4c74-968e-0ef552762418","added_by":"auto","created_at":"2025-12-10 15:33:48","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62934,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/b4c62d776a1ba168058ef287.png"},{"id":97702455,"identity":"89502aad-ec55-4a0f-b9fe-bb075ad7d81a","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61682,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/25e5d3d21414c058a3319d4e.png"},{"id":97702466,"identity":"72dcd1cb-7900-445d-bef3-2836bd731553","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":76805,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/07ded7a9a6195f9483e4a537.png"},{"id":97702464,"identity":"fc855c78-1fc2-410c-8f49-479c5b46587f","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":70635,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/ab9a99ab7c7884846ad8e6e3.png"},{"id":97702465,"identity":"9053a20a-9b7a-425f-82c5-df0aefef2702","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":130100,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/ca80ce2ae73899ec280cc64b.png"},{"id":97702461,"identity":"469f8f28-625f-4481-a53a-c3bfd92429ba","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":35090,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/bb883064cced0b28e448b5e0.png"},{"id":97894538,"identity":"d20967fe-02fe-4785-b47f-64b0c2bc44ac","added_by":"auto","created_at":"2025-12-10 15:32:41","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":142377,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/568e9b40603b6c71f443407a.png"},{"id":97893532,"identity":"dc81a9c7-78c8-4abe-96a2-22c808c98b44","added_by":"auto","created_at":"2025-12-10 15:30:40","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":214389,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/29be146df17da545c3594328.png"},{"id":97893163,"identity":"df2aa0ce-125a-4251-88ff-fc11defbb871","added_by":"auto","created_at":"2025-12-10 15:28:06","extension":"xml","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":96847,"visible":true,"origin":"","legend":"","description":"","filename":"4492c3023b7e40cd98ee0a41dd2af2511structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/f0dcda872a0933fc1f623015.xml"},{"id":97702468,"identity":"b43918d5-fa76-4ca2-aef9-ba59f101ec81","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"html","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":107868,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/05c810bc032dd74fdbfc61af.html"},{"id":97702441,"identity":"275ef9f3-45b0-4e0b-a56b-5dc6a1e10878","added_by":"auto","created_at":"2025-12-08 12:33:05","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":562429,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the AtriumDB Platform.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/c2f3287c935c9cc753489f5c.png"},{"id":97893355,"identity":"2a8e983c-8504-456a-aede-bd82cb3b67a9","added_by":"auto","created_at":"2025-12-10 15:30:12","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":566168,"visible":true,"origin":"","legend":"\u003cp\u003eHierarchical arrhythmia label schema used for dataset labeling\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/94ad30c61a5ad023151a2e93.png"},{"id":97702444,"identity":"9cb2b3f0-776f-4724-beab-69448f5f709c","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":653319,"visible":true,"origin":"","legend":"\u003cp\u003eCustom Label Studio interface for ECG annotation\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/c9ad8020fc9f040c05b3d8c8.png"},{"id":97702445,"identity":"f9ada361-2743-4f7c-877f-d84058be3e50","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":629837,"visible":true,"origin":"","legend":"\u003cp\u003eSlack-based clinician workflow for reporting arrhythmia events in real time\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/754b281b5b70a261cee3a741.png"},{"id":97702448,"identity":"21987cd2-a8fe-4b25-8178-25f7f356bf02","added_by":"auto","created_at":"2025-12-08 12:33:06","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":651450,"visible":true,"origin":"","legend":"\u003cp\u003eUMAP\u003ca href=\"https://www.zotero.org/google-docs/?wtZrX2\"\u003e[14]\u003c/a\u003e clustering of patient embeddings. The colours on the left side of the image represent the different label classes. On the right, six representative ECG segments from six unique patients illustrate the morphological diversity of junctional arrhythmias\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/522fe9b09178eaa70165361d.png"},{"id":97892844,"identity":"ffd09e3d-b92e-4712-8a19-200712b91f0b","added_by":"auto","created_at":"2025-12-10 15:23:26","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":140256,"visible":true,"origin":"","legend":"\u003cp\u003eHigh-level architecture of the arrhythmia classifier\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/2abae7ba690b4acee2cfde7d.png"},{"id":97894570,"identity":"69187c59-e9b5-4813-926f-6f5f45943e0a","added_by":"auto","created_at":"2025-12-10 15:32:43","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":981781,"visible":true,"origin":"","legend":"\u003cp\u003eHigh-level workflow for embedding-based label selection\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/90858067fb9d3b96f4eba9fd.png"},{"id":97894976,"identity":"16cf9277-3135-494d-833e-09e8fe261da0","added_by":"auto","created_at":"2025-12-10 15:33:18","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":996102,"visible":true,"origin":"","legend":"\u003cp\u003eUMAP projection of the model’s embedding space, highlighting the clustering of junctional arrhythmias. The left subplot displays the full set of labelled ECG segments colored by their parent rhythm class.\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/4fb94d7c8fed988a9de68bda.png"},{"id":97902586,"identity":"e47d82e2-7430-4f9b-9730-bc04b4ba065b","added_by":"auto","created_at":"2025-12-10 15:53:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5895926,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8194963/v1/5d17a21d-36d3-456a-98bd-29e83e3538d4.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Accelerating Machine Learning in Healthcare: Addressing the Labelling Bottleneck","fulltext":[{"header":"Introduction","content":"\u003cp\u003eArrhythmias are a common complication following pediatric congenital heart surgery and are associated with increased morbidity and mortality[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Certain arrhythmias, such as junctional ectopic tachycardia (JET), can be particularly harmful, potentially resulting in hemodynamic instability and circulatory collapse if not identified and treated in a timely manner[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Delays in diagnosis may arise from a range of factors, including limited experience among front-line providers, systemic constraints such as provider availability, and other operational challenges.\u003c/p\u003e\u003cp\u003eMachine learning (ML)-based clinical decision support (CDS) systems have emerged as a promising tool to improve the efficiency of arrhythmia detection to facilitate prompt diagnosis and intervention. Most existing ML models have been developed using adult datasets and rely heavily on 12-lead electrocardiogram (ECG) recordings[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] which has limited utility in pediatric populations. These models are not well-suited to the pediatric intensive care unit (ICU) setting, where arrhythmia detection is typically initiated by clinicians using bedside monitors that display only one to three ECG leads. The 12-lead ECG is generally reserved for confirmation after an arrhythmia is suspected. Furthermore, age-dependent variations in ECG waveform characteristics, including heart rate, R-R intervals, and amplitude limit the applicability of adult data for pediatric ML-based CDS[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eExisting research often depends on publicly available datasets like MIT-BIH[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], PTB-XL[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], PhysioNet 2021[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], and MIMIC-IV-ECG[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], which focus on adult populations and 12-lead recordings, making them unsuitable for detecting JET in children. Prior studies[\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e–\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] have trained models using 12-lead or multi-lead ECG data, which typically requires prior clinical suspicion and manual acquisition. In contrast, our approach utilizes continuous lead II ECG monitoring, which reflects standard ICU practice where pediatric patients rarely have all 12 leads attached. This strategy also enables real-time arrhythmia detection without relying on clinician-initiated ECG acquisition.\u003c/p\u003e\u003cp\u003eA critical barrier to developing effective ML models for pediatric arrhythmia detection is the lack of large, labeled datasets that reflect the clinical reality of continuous 3-lead monitoring[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. The reliance on 12-lead ECGs in research is often driven by their association with confirmed diagnoses, which simplifies retrospective labeling. However, this approach does not capture the continuous data streams used in practice. To address this gap, the present study outlines a novel methodology for constructing a labeled pediatric ECG dataset suitable for training ML models. This paper describes a scalable and clinically informed strategy to extract meaningful labels from an extensive, unlabelled AtriumDB[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] dataset, with the goal of enabling the development of robust and deployable ML models for pediatric arrhythmia detection.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eAtriumDB: Physiologic Database\u003c/p\u003e\u003cp\u003eWe developed two integrated platforms, AtriumDB and AtriumStreams (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), to enable both large-scale physiologic data capture and real-time model deployment. AtriumDB is a high-performance timeseries database specifically designed for use with physiological waveforms[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] and has been operational in 42 bedspaces at The Hospital for Sick Children (SickKids) ICU since 2016. It has supported the collection of over 1.6\u0026nbsp;million hours of continuous lead II ECG data from more than 9,000 critically ill children in the ICU and operating rooms leading to one of the largest pediatric physiologic waveform datasets in the world. AtriumDB includes a dedicated python library that enables rapid data extraction at scale, specifically designed to support high-throughput deep learning GPU training workflows. Complementing this, AtriumStreams is a real-time ingestion and deployment engine that ingests real time waveforms directly from bedside monitors, enabling real-time inference at the bedside. Together, AtriumDB and AtriumStreams function as the foundation of our broader translational pipeline. This system has now been deployed at 10 international hospitals in both pediatric and adult settings (as of Sept 2025).\u003c/p\u003e\u003cp\u003eLabel Schema\u003c/p\u003e\u003cp\u003eTo support dataset labeling, we developed a hierarchical label schema (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) grounded in pediatric electrophysiology practice and iteratively refined through expert input. Labels were applied to lead II ECG strips from the bedside monitor as described in more detail below. The schema organizes rhythm types into five clinically meaningful parent categories: Sinus Rhythms, Blocks, Paced Rhythms, Ventricular Arrhythmias, and Supraventricular (SVT) Arrhythmias. Each parent class includes a set of child labels representing specific rhythm diagnoses encountered in pediatric critical care. The “unknown” class is excluded from the training data and is available for when samples aren't exactly noise but clinicians also can't confidently determine the rhythm class.\u003c/p\u003e\u003cp\u003eInitial Labelling Workflow\u003c/p\u003e\u003cp\u003eGround Truth Labels\u003c/p\u003e\u003cp\u003eAll ECG labels used for training and evaluation were annotated by a team of pediatric intensive care clinicians with specialized cardiac training operating under the direction of a senior pediatric electrophysiologist. Given the limited availability of expert clinicians, we adopted a distributed annotation strategy to maximize efficiency, limiting labelling of each individual ECG segment to a single labeller to avoid overlapping labels requiring adjudication. When uncertainties arose, annotators maintained active communication with the senior electrophysiologist to resolve ambiguities and uphold labelling integrity. As such, the resulting annotations from each labeller were accepted as ground truth.\u003c/p\u003e\u003cp\u003eLabelling Software\u003c/p\u003e\u003cp\u003eTo train our arrhythmia classifier, which assigns rhythm labels to 5-second segments of lead II ECG, we developed a clinically integrated labelling pipeline built around Label Studio, an open-source annotation platform (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The interface was custom-designed with direct input from pediatric electrophysiologists, enabling users to securely log in and label de-identified waveform data through a web-based environment. Label Studio’s native support for time-series data, combined with its customizable HTML framework, allowed us to tailor the interface for efficient ECG annotation at scale.\u003c/p\u003e\u003cp\u003eOur interface presented both a 2–5 minute lead II ECG waveform strip and synchronized heart rate signal (sampled at 1 Hz). If a 12-lead ECG was obtained with a confirmed diagnosis in the MUSE system with a timestamp corresponding to the telemetry strip in question, the annotation was included as contextual metadata on the labelling interface. Label selection occurs through a structured dropdown menu derived from the hierarchical schema (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe hierarchical schema supports both coarse and fine-grained annotation. Annotators can apply a parent label (e.g., “Ventricular Arrhythmia”) or designate more specific subtypes via child labels. This structure enabled flexible model training strategies (e.g. grouping rare subtypes under a parent class to mitigate class imbalance). During annotation, users first selected a predominant rhythm for the full waveform and then mark specific regions (e.g., noise artifacts or other arrhythmias) where applicable. For instance, a 2-minute segment may be labelled globally as Sinus Rhythm, with one or more short X-second regions identified as another rhythm, which significantly reduces annotation time compared to fully segmented labelling.\u003c/p\u003e\u003cp\u003eTo mitigate performance constraints inherent to Label Studio, which exhibits latency and interaction delays beyond approximately 300,000 data points, or 10 minutes of ECG sampled at 500 Hz, we adopted a segmentation strategy that partitioned longer waveform regions into contiguous 2–5 minute chunks. These segments were grouped as sequential tasks, with consistent annotations in the user interface denoting patient identity and temporal ordering. This approach preserved critical clinical context while maintaining interface responsiveness.\u003c/p\u003e\u003cp\u003eTo maintain synchronization with our physiologic data platform, we implemented webhook functionality within Label Studio. These webhooks notify a custom Flask server each time a label is created, modified, or deleted, enabling real-time updates to an AtriumDB dataset and supporting rapid retraining cycles as new labelled data becomes available. This continuous feedback loop between expert labelling and model development is central to our clinician-in-the-loop workflow and ensures that improvements in the dataset immediately translate into model refinement.\u003c/p\u003e\u003cp\u003eSummary of Different Label Sources\u003c/p\u003e\u003cp\u003eTo construct a high-quality training dataset for the arrhythmia classifier, we implemented a multi-pronged strategy to source labels from both retrospective and prospective data streams. Given the rarity and subtlety of JET, no single labelling approach was sufficient. Instead, we leveraged four complementary techniques to enrich the dataset with arrhythmias of interest:\u003c/p\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eMUSE: Structured diagnoses associated with 12-lead ECG studies.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eICU Observations: Bedside annotations triggered in real time by clinical staff.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eUncertainty Sampling: Model-guided selection of ambiguous or low-confidence windows.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eEmbedding Search: Retrieval of morphologically similar ECG segments via model-derived feature embeddings.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cp\u003eEach of these strategies yielded different levels of labelling efficiency, patient coverage, and arrhythmia density. The general principle was that we used a “known” anchor (e.g., from MUSE or an ICU observation) to pull surrounding “unknown” ECG waveforms likely to have a similar label.\u003c/p\u003e\u003cp\u003eSourcing Labels From MUSE\u003c/p\u003e\u003cp\u003eOur initial strategy for sourcing arrhythmia labels used the presence of a formal 12-lead ECG study as a proxy indicator for an arrhythmia event. Specifically, we extracted all 12-lead studies from the MUSE system at SickKids and aligned them with continuous ECG waveform data stored in an AtriumDB dataset by matching timestamps. The diagnostic interpretation associated with each 12-lead study, typically authored by an electrophysiologist or generated via rule-based algorithms, was attached to the corresponding ECG segment and presented to expert labellers in Label Studio for verification.\u003c/p\u003e\u003cp\u003eHowever, this approach introduced two critical challenges. First, we frequently observed discrepancies between the diagnostic interpretation derived from the 12-lead ECG and the rhythm classification provided by expert annotators based on the corresponding waveform segments. Combined with our prior findings that ICU clocks are often significantly desynchronized[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], this suggests that the MUSE system timestamps are misaligned with those of the Philips monitoring infrastructure to a degree that precludes reliable temporal matching. Consequently, the ECG strips presented to annotators often did not correspond precisely to the diagnostic moment captured by the 12-lead system, thereby compromising the validity of the associated diagnostic labels. These misalignments necessitated manual re-labeling and rigorous quality assurance procedures, substantially increasing the annotation burden.\u003c/p\u003e\u003cp\u003eSecond, even when correctly aligned, the segments sourced via 12-lead studies were predominantly composed of normal sinus rhythm (85% in our cohort). While sinus rhythm examples are essential for training a balanced classifier, the disproportionate prevalence of non-informative segments significantly limited the efficiency of expert labelling efforts. Given the constraints on clinician time, it became clear that this method was not viable as a primary mechanism for surfacing high-value arrhythmia examples.\u003c/p\u003e\u003cp\u003eSourcing Labels From ICU Observations\u003c/p\u003e\u003cp\u003eTo surface clinically relevant arrhythmia examples in real time, we implemented a clinician-in-the-loop annotation workflow that leverages the situational awareness of ICU providers. Initially deployed via Slack, a secure mobile messaging platform, the system allowed nurses and physicians to rapidly report observed arrhythmias through a structured form (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Designed for minimal workflow disruption, the form required only two fields, patient room number and event timestamp, while offering optional fields such as arrhythmia category (restricted to parent-level classes) and free-text clinical notes. These optional fields were later used to filter segments in Label Studio by rhythm class. The form also included a “From admission start” checkbox, enabling clinicians to estimate timing without needing to retrieve the precise encounter start time, further reducing friction. Upon submission, the input was automatically transmitted to a Flask-based server, which queried the dataset and extracted two segments: a one hour window following the reported start time and a 20-minute window at the end of the encounter, which typically contained Sinus Rhythm.\u003c/p\u003e\u003cp\u003eThis dual-segment retrieval strategy was designed to provide the model not only with positive examples of arrhythmias but also with patient-specific sinus rhythm morphology, facilitating finer-grained contrastive learning. In cases where the patient encounter had not yet ended at the time of submission, a background scheduler re-queried the dataset every five minutes until the terminal segment became available and could be uploaded for labelling. This flexible, low-friction reporting mechanism enabled bedside clinicians to contribute high-yield labelling candidates with minimal interruption to care, significantly increasing the number of rare rhythm examples, particularly junctional arrhythmias, compared to passive approaches. The integration of annotation triggers into tools already used during clinical care not only enhanced data relevance but also improved labeller efficiency by reducing the proportion of sinus rhythm segments.\u003c/p\u003e\u003cp\u003eFollowing completion of ICU-based labeling, the resulting dataset contained approximately 80 minutes of annotated ECG data per patient but was limited by a relatively small cohort size. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e illustrates the clustering of all labeled segments, where embeddings from the same patient exhibit tight grouping, indicating minimal intra-patient variability within rhythm classes. While labeling one hour per patient provided valuable initial coverage, this approach is highly time-consuming for clinicians and yields diminishing returns because ECG signals are inherently repetitive within a single patient. Nevertheless, it was essential for training an initial model capable of supporting model-guided labeling. To improve generalization and reduce annotation burden moving forward, we aim to increase inter-patient diversity by labeling shorter segments (approximately two minutes) from a larger number of unique patients. This strategy maximizes variability while minimizing clinician workload. To operationalize this approach, we adopted a model-guided labeling framework.\u003c/p\u003e\u003cp\u003eModel Guided Labelling Workflow\u003c/p\u003e\u003cp\u003eThe final two labeling strategies leverage a machine learning model trained on data obtained through the initial labeling methods to identify additional candidate segments within the dataset. Although the model does not yet demonstrate sufficient performance for clinical deployment, it is capable of highlighting regions within the unlabeled data that are likely to contain arrhythmias. This enables a targeted labelling approach, focusing on rhythm classes the model finds challenging or are underrepresented in the current dataset. To contextualize these strategies, we first provide a summary of the model architecture and training process. Further technical details and model performance metrics will be presented in a subsequent publication.\u003c/p\u003e\u003cp\u003eFoundational Model\u003c/p\u003e\u003cp\u003eThe foundation model was trained on continuous lead II ECG data collected from the SickKids ICU. It uses a Wave2Vec 2.0-based[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] architecture, which includes a 4-layer convolutional encoder followed by 12 transformer blocks configured similarly to BERT-Base[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Training was performed using Contrastive Masked Segment Coding (CMSC)[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], a self-supervised learning approach that combines local and global contrastive objectives. Locally, the model learns to reconstruct masked ECG segments from the surrounding context, while globally it learns patient-level rhythm patterns by comparing adjacent segments from the same patient against segments from others. This enables the model to learn temporally-rich and morphology-aware representations from unlabelled data.\u003c/p\u003e\u003cp\u003eArrhythmia Classifier\u003c/p\u003e\u003cp\u003eTo adapt the foundation model for supervised classification, self-supervised components such as masking and contrastive losses were removed. A lightweight classification head was added, consisting of temporal average pooling and three fully connected layers, followed by a softmax output over four rhythm classes: Junctional Arrhythmia, Sinus Rhythm, Other Rhythm, and Noise (see Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). These classes were selected to reflect clinically relevant distinctions and to support the detection of junctional arrhythmias in real-world ICU settings.\u003c/p\u003e\u003cp\u003eTraining was conducted using the expert-labelled lead II ECG segments obtained from the previously described methods. The model was optimized using weighted binary cross-entropy loss to address class imbalance. Physiologically realistic augmentations were employed such as powerline interference, EMG noise, and baseline wander; to improve robustness under ICU conditions. Standard normalization and interpolation techniques were applied to insure signal consistency and handle missing data.\u003c/p\u003e\u003cp\u003eThe classifier’s softmax outputs were used to identify low-confidence predictions for uncertainty sampling, enabling the selection of ambiguous segments near the decision boundary. Additionally, the 256-dimensional embeddings from the final hidden layer were used for embedding-based sampling, allowing the retrieval of morphologically similar but diverse examples from the unlabeled dataset. These capabilities enabled targeted mining of an AtriumDB dataset, improving label efficiency and enhancing model generalization.\u003c/p\u003e\u003cp\u003eSourcing Labels From Uncertainty Sampling\u003c/p\u003e\u003cp\u003eUncertainty sampling is a widely used technique in active learning, wherein a model is used to identify samples for which its predictions are least confident, those that lie near the decision boundary. These cases are often the most informative for training, as they expose the model’s limitations and guide learning toward harder examples. In our context, the combination of labelled data from MUSE-derived 12-lead studies and clinician-reported ICU observations provided enough coverage to train a preliminary arrhythmia classifier. While limited in performance, this initial model was sufficient to enable active mining of the unlabeled ECG data dated back to 2016.\u003c/p\u003e\u003cp\u003eWe applied the preliminary model to the full Lead II ECG dataset and selected ECG segments based on model uncertainty. For the initial round, we extracted 2-minute segments from 25 unique patients where the model's softmax score for the Junctional Arrhythmia class fell between 40% and 65%, a range indicative of ambiguous, borderline predictions. These cases were then pushed to Label Studio for expert labelling. This process was repeated iteratively, with each round expanding the number of sampled patients and enriching the training dataset with ECG morphologies the model found most difficult to classify. As the model improved across successive iterations, the segments retrieved through uncertainty sampling shifted in character. Increasingly, the uncertain windows contained low signal-to-noise ratios or atypical morphologies that were difficult to classify, even for human experts. This inflection point marked the diminishing returns of uncertainty-based sampling.\u003c/p\u003e\u003cp\u003eSourcing Labels From Embeddings\u003c/p\u003e\u003cp\u003eIn contrast to uncertainty sampling, embedding-based sampling is a form of representation-aware active learning that leverages the model’s internal structure to surface underrepresented but high-confidence samples. Rather than querying the decision boundary, embedding search targets the interior of a known class, retrieving morphologically similar rhythms from new patients. This strategy is particularly effective in the later stages of model development, when the classifier has already learned a stable representation of key rhythm classes but requires additional diversity to generalize across patients. Here, our goal was to further enrich the junctional arrhythmia class by identifying previously unseen morphologic variants across the population, while reducing annotation burden and avoiding redundant examples. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e represents a high-level overview of the embedding search technique.\u003c/p\u003e\u003cp\u003eTo achieve this, we applied our arrhythmia classifier to the entire dataset and extracted 256-dimensional embeddings from the final hidden layer of the network for every 5-second ECG segment. Using the set of previously labelled junctional arrhythmia segments, we computed a global centroid in this high-dimensional embedding space, as well as per-patient centroids for each labelled junctional arrhythmia patient. We then defined a hypersphere centred at the global centroid, with a radius equal to the mean plus one standard deviation of the distances between each patient’s centroid and the global centre. This hypersphere captured the typical distribution of junctional arrhythmia morphology as represented by the model and defined our search region for new candidate segments.\u003c/p\u003e\u003cp\u003eWithin this embedding-defined region, we searched for 2-minute segments of ECG that exhibited low “embedding variability”, that is, whose 24 constituent 5-second windows were tightly clustered in embedding space. This constraint was empirically found to increase the likelihood that a given 2-minute segment contained only one class, thereby reducing the cognitive load on annotators and increasing labelling throughput. For each selected patient, we also extracted an additional 1-minute segment from the end of their encounter to serve as a likely sinus baseline, supporting the model’s ability to learn intra-patient morphology contrasts.\u003c/p\u003e\u003cp\u003eOur first batch of embedding-sampled segments yielded data from 294 new patients. After expert annotation, 60.2% of segments were confirmed as junctional arrhythmias, with 18.7% sinus, 3.4% noise, and 17.7% representing other arrhythmias, highlighting the high precision and label efficiency of the approach. Importantly, the remaining 39.8% of non-junctional segments were also valuable, as they represent cases where the model positioned embeddings close to the junctional cluster. Incorporating these examples during retraining enhanced the model’s ability to distinguish junctional arrhythmias from morphologically similar rhythms.\u003c/p\u003e\u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e, the UMAP projection of the model’s embedding space illustrates the structure of the learned representation. This visualization confirms that junctional arrhythmias occupy a coherent and distinct region in embedding space, separable from other classes. The right plot focuses on that junctional region, showing a subset of previously labelled junctional arrhythmia examples alongside unlabeled segments sampled from within the search hypersphere. These are the candidate samples identified for expert annotation via Label Studio. Six representative ECG strips from six unique patients are also shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, highlighting the substantial inter-patient variability in junctional arrhythmia morphology.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eLabeled Dataset\u003c/p\u003e\u003cp\u003eThe final labelled dataset created from the above techniques consists of 189.2 hours of expert-annotated lead II ECG from 1,447 unique pediatric patients, encompassing four high-level rhythm classes: Junctional Arrhythmia, Sinus Rhythm, Other Rhythm, and Noise. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the total number of labelled hours and unique patients obtained through each method, as well as the subset of those labels containing Junctional arrhythmias, our primary modelling target.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eSummary of label yield by sourcing strategy\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003eAll Labels\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e\u003cp\u003eJunctional Arrhythmia Labels\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLabeling Method\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMUSE\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e75.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e991\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e5.75\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eICU Observations\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e79.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e47\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e11.23\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e18\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eUncertainty Sampling\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e12.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e3.14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e37\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eEmbedding Search\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e21.36\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e408\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e7.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e214\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTotal\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e189.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1447\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e27.77\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e358\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eTo ensure demographic and physiologic diversity, the dataset was stratified by both sex and age group, capturing a representative distribution across the pediatric population from neonates to late adolescents. As shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the dataset includes contributions from both male and female patients, while Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e further details the label distribution across age groups.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eDistribution of labelled ECG hours and patient counts across rhythm classes, stratified by sex\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e\u003cp\u003eMale\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003eFemale\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eJunctional Arrhythmia\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e27.77\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e358\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e14.86 (53.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e203\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e12.91 (46.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e155\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSinus Rhythm\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e96.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1080\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e59.49 (61.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e594\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e37.14 (38.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e486\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eOther Rhythm\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e46.17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e315\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e25.84 (56%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e176\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e20.33 (44%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e139\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eNoise\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e18.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e647\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e12.21 (65.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e360\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e6.42 (34.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e287\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTotal\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e189.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1447\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e112.39 (59.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e797\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e76.81 (40.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e650\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eDistribution of labelled ECG hours and patient counts across rhythm classes, stratified by pediatric age group\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"11\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003eNeonate (0\u0026ndash;28 Days)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e\u003cp\u003eInfant (29 Days-1 Year\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003eToddler (1\u0026ndash;3 Years)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003eChild (4\u0026ndash;12 Years)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003eAdolescent (13\u0026ndash;18 Years)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eHours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c11\"\u003e\u003cp\u003ePatients\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eJunctional\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eArrhythmia\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3.09 (11.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e51\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e13.15 (47.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e144\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.87 (6.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e31\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e9.61 (34.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e134\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e3.09 (11.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e51\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSinus Rhythm\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e19.44 (20.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e205\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e36.86 (38.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e318\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e9.26 (9.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e113\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e30.47 (31.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e483\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e19.44 (20.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e205\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eOther Rhythm\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e15.04 (32.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e22.14 (48%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e116\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e2.44 (5.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e37\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e6.45 (14%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e15.04 (32.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e78\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eNoise\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3.63 (19.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e168\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e4.25 (22.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e189\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e5.99 (32.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e64\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e4.74 (25.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e249\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e3.63 (19.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e168\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of labeling time requirements by label acquisition strategy. The results are presented using two key time constructs: \u003cem\u003eLead time\u003c/em\u003e and \u003cem\u003eTask duration\u003c/em\u003e. \u003cem\u003eLead time\u003c/em\u003e refers to the annotator\u0026rsquo;s active labeling time per task. Outliers in lead time were removed because they typically occur when an annotator steps away and leaves a task open, which would otherwise inflate time estimates and misrepresent efficiency. \u003cem\u003eTask duration\u003c/em\u003e represents the length of ECG content presented in each task, while \u003cem\u003elead time per minute of task\u003c/em\u003e normalizes effort to the amount of ECG reviewed.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLabeling method\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTasks analyzed\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTask duration (s)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMedian lead time (s)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eLead time per minute of task (s/min)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eTasks per annotator hour\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eECG minutes per annotator hour\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eTotal annotator hours\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTotal ECG hours labeled\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMUSE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1772\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e120\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e27.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e37\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e48.43\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e59\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eICU observations\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e598\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e300\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e33\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e6.57\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e33\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e165\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e18.29\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e50.23\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eUncertainty Sampling\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e207\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e300\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e65\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e13.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e31\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e124\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e6.59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e13.62\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eEmbedding Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e550\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e120\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e41\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e24.34\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e58\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e149\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e9.53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e23.67\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eDeveloping clinically deployable machine learning models for pediatric arrhythmia detection requires overcoming the labeling bottleneck inherent to large-scale physiologic datasets. This study demonstrates that a multi-strategy labeling pipeline combining retrospective data mining, clinician-in-the-loop annotation, and active learning can accelerate dataset creation while preserving clinical fidelity.\u003c/p\u003e\u003cp\u003eWe began with two initial strategies: extracting labels from MUSE 12-lead electrocardiogram studies and real-time ICU observations. While MUSE provided broad coverage it was inefficient; it yielded predominantly sinus rhythm data and required extensive manual verification. ICU observations, although limited in patient count, produced high-yield arrhythmia segments with minimal workflow disruption. With ICU observations we took one hour of arrhythmia data from each patient and 20 minutes of possible normal sinus rhythm. This was an arbitrary amount of time we selected in the beginning however we have now changed that to 20 minutes of arrhythmia and 10 minutes of sinus. Labelers spent 58% of the total time labeling on MUSE tasks which was our lowest yield method of finding arrhythmias. If repeated, we would prioritize ICU observations from the outset, as they offered the most clinically relevant data with the least annotation overhead helping us get to model-guided annotation faster.\u003c/p\u003e\u003cp\u003eThe MUSE and ICU observations provided sufficient labeled data to train a preliminary classifier that could generate embeddings for active learning. This model enabled uncertainty sampling and embedding-based retrieval, which dramatically improved label efficiency and inter-patient variability. Embedding-guided sampling, in particular, enriched rare arrhythmia classes such as junctional arrhythmias while reducing annotation burden. These findings highlight the importance of representation-aware approaches in pediatric machine learning development, where physiologic variability across age groups and clinical contexts limits the utility of larger adult datasets. By leveraging the scale of the physiological waveform dataset accumulated at SickKids and integrating expert feedback loops, we created one of the largest labeled pediatric ECG corpora to date, enabling robust model training under real-world ICU conditions.\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e shows each strategy offered distinct advantages and trade-offs. MUSE and ICU observations required the greatest total annotator time, at approximately 48.4 hours and 18.3 hours, respectively. In contrast, uncertainty sampling and embedding-guided selection were far more efficient, producing the most diverse set of patients and arrhythmias while requiring 6.6 hours and 9.5 hours, respectively. These figures underscore that most labeling effort occurs before reaching model-guided techniques. Overall, approximately 80 hours of clinician time was utilized in dataset creation, which is relatively modest given the scale and clinical impact of this use case.\u003c/p\u003e\u003cp\u003eHowever, clinician feedback emphasized that concentration declines after about an hour of continuous labeling, and we observed a corresponding drop in label quality. To mitigate fatigue, it was more effective to provide smaller batches (20 minutes to one hour of labeling) rather than assigning hundreds of tasks at once, improving engagement and consistency.\u003c/p\u003e\u003cp\u003eWhen normalized by ECG content, ICU observations and uncertainty sampling were the most efficient, reflecting their 5-minute task durations. In contrast, MUSE data and embedding-guided selection, which used 2-minute windows, were slower to complete on a per-minute basis. Median lead times and tasks per annotator hour were broadly similar across methods, but longer windows provided richer context, making annotations easier. Based on these findings, we recommend task durations in the 2\u0026ndash;5 minute range to balance efficiency with annotation quality.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study addresses a key barrier to machine learning adoption in pediatric critical care by introducing a scalable, clinically informed labeling framework. Our approach accelerates data curation and establishes a generalizable methodology for domains where expert annotation is time-intensive and rare events dominate. Future work will focus on expanding label diversity and validating model performance in silent trials prior to bedside deployment. Ultimately, the combination of clinician in the loop and model-guided active learning represents a transformative step toward real-time, clinician-augmented machine learning systems capable of improving patient safety and outcomes.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics Approval and Consent to Participate\u003c/p\u003e\n\u003cp\u003eThe ECG dataset used in this study is part of a clinical AtriumDB-based dataset, collected prospectively from a 42-bed ICU at The Hospital for Sick Children (SickKids) in Toronto, ON since 2016. Secondary use of this dataset is governed by REB approval #1000068499 and has a waiver of consent.\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003eAvailability of Data and Material\u003c/p\u003e\n\u003cp\u003eDue to patient privacy and institutional data governance policies, the ECG waveform data used in this study cannot be publicly released. However, access to de-identified subsets of the dataset may be made available to academic researchers upon request and subject to ethics approval and data use agreements with The Hospital for Sick Children (SickKids). The code used to train the arrhythmia classifier is not publicly available. Qualified academic collaborators may request access to specific components for non-commercial research under a data use agreement.\u003c/p\u003e\n\u003cp\u003eCompeting Interests\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Canadian Institutes of Health Research (CIHR) through a Project Grant (Application Number: 480831), The University of Toronto Data Science Institute Catalyst Grant, and The University of Toronto EMH Seed Grant.\u003c/p\u003e\n\u003cp\u003eAuthors\u0026apos; Contributions\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by S. Vecile and W. Dixon. ECG labeling was done by A. Bulic, A. Assadi, M. Kim, D. Ehrmann, D. Eytan and M. Mazwi. The first draft of the manuscript was written by S. Vecile. The manuscript was commented on and edited by S. D. Goodfellow, M. Kim, A. Goodwin, W. Dixon, A. Siddiqui, A. Assadi, D. Ehrmann and R. Greer. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eNot Applicable.\u003c/p\u003e\n\u003cp\u003eClinical Trial Number\u003c/p\u003e\n\u003cp\u003eClinical trial number: not applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMayaya P-C, Tinică G, Chistol RO, Moraru L, Damian SI, Frăsinariu OE, et al. Incidence and Risk Factors for Early Postoperative Arrhythmias in Congenital Heart Disease \u0026ndash; Systematic Review. J Cardiovasc Emergencies. T\u0026acirc;rgu Mureș, Poland: De Gruyter Brill Sp. z o.o., Paradigm Publishing Services; 2025;11:43\u0026ndash;53. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2478/jce-2025-0005\u003c/span\u003e\u003cspan address=\"10.2478/jce-2025-0005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePantelidis P, Bampa M, Oikonomou E, Papapetrou P. Machine learning models for automated interpretation of 12-lead electrocardiographic signals: a narrative review of techniques, challenges, achievements and clinical relevance. J Med Artif Intell [Internet]. AME Publishing Company; 2023 [cited 2025 Nov 10];6. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.21037/jmai-22-94\u003c/span\u003e\u003cspan address=\"10.21037/jmai-22-94\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSkippen PW, Sanatani S, Gow RM, Froese N. Diagnosis of postoperative arrhythmias following paediatric cardiac surgery. Anaesth Intensive Care. 2009;37:705\u0026ndash;19. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0310057X0903700506\u003c/span\u003e\u003cspan address=\"10.1177/0310057X0903700506\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMoody GB, Mark RG. The impact of the MIT-BIH Arrhythmia Database. IEEE Eng Med Biol Mag. 2001;20:45\u0026ndash;50. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/51.932724\u003c/span\u003e\u003cspan address=\"10.1109/51.932724\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWagner P, Strodthoff N, Bousseljot R-D, Kreiseler D, Lunze FI, Samek W, et al. PTB-XL, a large publicly available electrocardiography dataset. Sci Data. Nature Publishing Group; 2020;7:154. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41597-020-0495-6\u003c/span\u003e\u003cspan address=\"10.1038/s41597-020-0495-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eReyna MA, Sadr N, Alday EAP, Gu A, Shah AJ, Robichaux C, et al. Will Two Do? Varying Dimensions in Electrocardiography: The PhysioNet/Computing in Cardiology Challenge 202. Brno, Czech Republic; 2021. p. 1\u0026ndash;4. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.23919/CinC53138.2021.9662687\u003c/span\u003e\u003cspan address=\"10.23919/CinC53138.2021.9662687\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJohnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. Nature Publishing Group; 2023;10:1. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41597-022-01899-x\u003c/span\u003e\u003cspan address=\"10.1038/s41597-022-01899-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOh J, Chung H, Kwon J, Hong D, Choi E. Lead-agnostic Self-supervised Learning for Local and Global Representations of Electrocardiogram [Internet]. arXiv; 2022 [cited 2025 Oct 2]. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2203.06889\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2203.06889\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKiyasseh D, Zhu T, Clifton DA. CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients [Internet]. arXiv; 2021 [cited 2025 Oct 2]. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2005.13249\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2005.13249\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMcKeen K, Masood S, Toma A, Rubin B, Wang B. ECG-FM: An Open Electrocardiogram Foundation Model [Internet]. arXiv; 2025 [cited 2025 Oct 2]. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2408.05178\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2408.05178\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShani C, Zarecki J, Shahaf D. The Lean Data Scientist: Recent Advances Toward Overcoming the Data Bottleneck. Commun ACM. 2023;66:92\u0026ndash;102. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/3551635\u003c/span\u003e\u003cspan address=\"10.1145/3551635\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGoodwin AJ, Eytan D, Greer RW, Mazwi M, Thommandram A, Goodfellow SD, et al. A practical approach to storage and retrieval of high-frequency physiological signals. Physiol Meas. IOP Publishing; 2020;41:035008. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1361-6579/ab7cb5\u003c/span\u003e\u003cspan address=\"10.1088/1361-6579/ab7cb5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGoodwin AJ, Dixon W, Mazwi M, Hahn CD, Meir T, Goodfellow SD, et al. The truth Hertz-synchronization of electroencephalogram signals with physiological waveforms recorded in an intensive care unit. Physiol Meas. 2023;44. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1361-6579/ace49e\u003c/span\u003e\u003cspan address=\"10.1088/1361-6579/ace49e\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMcInnes L, Healy J, Saul N, Grossberger L. UMAP: Uniform Manifold Approximation and Projection. J Open Source Softw. 2018;3:861. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.21105/joss.00861\u003c/span\u003e\u003cspan address=\"10.21105/joss.00861\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBaevski A, Zhou Y, Mohamed A, Auli M. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. In: Larochelle H, Ranzato M, Hadsell R, Balcan MF, Lin H, editors. Adv Neural Inf Process Syst [Internet]. Curran Associates, Inc.; 2020. p. 12449\u0026ndash;60. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://proceedings.neurips.cc/paper_files/paper/2020/file/92d1e1eb\u003c/span\u003e\u003cspan address=\"https://proceedings.neurips.cc/paper_files/paper/2020/file/92d1e1eb\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e1cd6f9fba3227870bb6d7f07-Paper.pdf\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDevlin J, Chang M-W, Lee K, Toutanova K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In: Burstein J, Doran C, Solorio T, editors. Proc 2019 Conf North Am Chapter Assoc Comput Linguist Hum Lang Technol Vol 1 Long Short Pap [Internet]. Minneapolis, Minnesota: Association for Computational Linguistics; 2019 [cited 2025 Oct 2]. p. 4171\u0026ndash;86. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.18653/v1/N19-1423\u003c/span\u003e\u003cspan address=\"10.18653/v1/N19-1423\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-medical-systems","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Journal of Medical Systems](https://www.springer.com/journal/10916)","snPcode":"10916","submissionUrl":"https://submission.nature.com/new-submission/10916/3","title":"Journal of Medical Systems","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Arrhythmia detection, Machine Learning, Dataset Labeling, Electrocardiogram (ECG)","lastPublishedDoi":"10.21203/rs.3.rs-8194963/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8194963/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTimely detection of postoperative arrhythmias after cardiac surgery is essential for preventing hemodynamic compromise. Machine learning models remain constrained by the scarcity of labeled datasets that reflect real-world monitoring practices. Existing approaches rely on adult 12-lead electrocardiograms which are rarely used continuously in pediatric ICUs and fail to capture age-dependent waveform variability. We present a clinically integrated labeling framework designed to overcome this bottleneck. Leveraging a physiologic waveform repository comprising over 1.6\u0026nbsp;million hours of unlabeled, continuous lead II ECG from more than 9,000 pediatric patients, we implemented a multi-phase strategy combining retrospective data mining, clinician-in-the-loop annotation, and active learning techniques, including uncertainty sampling and embedding-based retrieval.\u003c/p\u003e\u003cp\u003eInitial labeling from MUSE (GE Healthcare) studies and ICU observations produced 154.9 hours of annotated ECG waveforms but required extensive clinician effort and yielded limited inter-patient variability. These two strategies provided sufficient coverage to train a preliminary classifier, enabling representation-aware sampling that dramatically improved efficiency. Embedding-guided retrieval achieved a precision of 60.2% for junctional arrhythmias and increased patient diversity compared to clinician in the loop-based labeling, while reducing annotation time per positive segment. Using this approach, we curated 189.2 hours of expert-labeled ECG from 1,447 unique patients, enriched for junctional arrhythmias, the primary modeling target.\u003c/p\u003e\u003cp\u003eThis work addresses a critical barrier to pediatric machine learning development and establishes a scalable methodology for creating clinically relevant datasets at scale, paving the way for real-time, clinician-augmented decision support systems in pediatric critical care.\u003c/p\u003e","manuscriptTitle":"Accelerating Machine Learning in Healthcare: Addressing the Labelling Bottleneck","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-08 12:33:01","doi":"10.21203/rs.3.rs-8194963/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-23T08:32:03+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-23T04:58:43+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-06T20:01:33+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"113309968249341503564006040994468515334","date":"2026-02-02T23:43:41+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"98278440790775455407078149015966419921","date":"2026-02-02T10:28:11+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-25T04:35:20+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"205255714227170592173629417622586583110","date":"2025-12-30T01:24:14+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-05T03:48:26+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-12-04T20:07:20+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-12-02T13:21:03+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Medical Systems","date":"2025-11-24T14:31:33+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-medical-systems","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Journal of Medical Systems](https://www.springer.com/journal/10916)","snPcode":"10916","submissionUrl":"https://submission.nature.com/new-submission/10916/3","title":"Journal of Medical Systems","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"d7cd1ca1-2010-44b6-95b1-172ddf51d33d","owner":[],"postedDate":"December 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[],"tags":[],"updatedAt":"2026-04-28T22:23:10+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-08 12:33:01","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8194963","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8194963","identity":"rs-8194963","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}