AIx-Neo-Guard is a collaborative project of the neonatal and pediatric ICU of Uniklinik RWTH Aachen, Germany, and the Chair of Embedded Software of RWTH Aachen, Germany. The project is conducted as an observational cohort study and was prospectively registered at the German Clinical Trials register (DRKS00033228). The study is conducted in accordance with the Declaration of Helsinki [28] and the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement guidelines for reporting observational studies [29]. Data collection started in September 2024 and has no scheduled end date.

The study was approved by the local ethics committee (Ethik-Kommission an der Medizinischen Fakultät der Rheinisch-Westfälischen Technischen Hochschule Aachen [RWTH Aachen]; approval EK 24-213). The German Health Data Utilization Act grants an exemption from individual parental informed consent, as only data that are regularly used for diagnostic and therapeutic purposes are collected. Openly available information about ongoing studies at an institution needs to be provided to enable the exemption, and parents can actively revoke study participation.

Data are collected at the neonatal and pediatric tertiary care ICU of Uniklinik RWTH Aachen, Germany. The ward offers 9 beds, each equipped with the capability to provide respiratory support. Annually, approximately 350 patients aged 0 to 17 years are treated on the ward.

All eligible patients treated in the ICU are included in the dataset. Patients are excluded if their parents object to data processing or study participation. Furthermore, technical issues can lead to unintended exclusion.

The dataset incorporates monitoring and ventilation data, laboratory results, patient characteristics, selected diagnoses and interventions, and manual annotations. An overview is given in Table 1. Waveform data are stored at 125 Hz, except for electrocardiogram (ECG) data, which are stored at 500 Hz. Numerical data are recorded with a sampling rate of 1 Hz.

The bedside monitoring is conducted using a Philips IntelliVue X2 monitor (all Philips devices provided by Philips GmbH Market DACH) connected to an IntelliVue MP70 monitor. A Sentec Digital Monitor with V-Sign sensor (SENTEC AG) is used for transcutaneous CO₂ monitoring. Ventilatory support is supplied by a Leoni plus or Elisa ventilator (Löwenstein Medical). The ventilators and CO₂ monitoring are connected to the bedside monitoring using Philips IntelliBridge EC5 ID modules connected via a local area network cable to IntelliBridge EC10 modules. EC10 modules are connected to the monitor by a module rack M8048A. All monitoring and ventilation data are retrieved after storage from the Philips Patient Information Center iX using Philips Data Warehouse Connect, version C.03.

Parameters derived from the PDMS (CGM Medico [CompuGroup Medical SE and Co KGaA]) are extracted manually and retrospectively added to the dataset. These comprise patient characteristics (eg, admission weight and sex), diagnoses, laboratory results, body positioning, and interventions (imaging, intubations and extubations, operations, and nursing care). Results from manual data analyses are annotated using a project-intern visualization and annotation tool (refer to the IntensiView Tool section). These annotations included patient-ventilator asynchronies, apnea of prematurity, exact time of blood sampling, disconnection of ventilators or monitoring devices, or suctioning.

All data, including timelines, are stored in a coded form in a database hosted by the Chair of Embedded Software. The hospital intern patient identifiers are encoded using a randomly generated 32-character universally unique identifier. The encoding table is stored in an access-protected file at Uniklinik RWTH Aachen and only accessible to clinical investigators. A data flowchart is provided in Figure 1.

The IntensiView tool (RWTH Aachen University) is a project-specific software developed by the Chair of Embedded Software. The tool is developed in C++ using the Qt Framework to easily deploy across operating systems. A PHP-based backend is used to communicate with the database. IntensiView is used for data synchronization, visualization, annotation, and export. Visualization can be adapted to the user’s requirements by changing colors, adapting the number of plots with freely selectable axes, and switching between column and line diagrams. The annotations can then be added manually or automatized using detection algorithms and are inserted into the database. A screenshot is shown in Figure 2.

An annotation search tool enables the user to reload episodes with specific annotations directly. All data can be exported in CSV format. Additionally, external time-series data and annotations can be easily imported and visualized.

A plug-in interface offers standardized access to the data loaded in the IntensiView tool to integrate new functionality, such as detection algorithms that support current research questions. The tool offers a variety of plug-ins for data analyses, for example, an ECG analysis tool, parameter calculation (eg, mean and median, calculation of airway volume curve from the flow curve, and calculation of airway resistance and compliance from ventilator pressure and flow waveform data), and the creation of ventilator loops.

We included all eligible patients to avoid selection bias. All patients were treated according to local standards. The study participation did not alter diagnostics and treatment. Monitoring and measurements were conducted by the same type of measuring devices in all patients to minimize detection bias.

The aim was to include all neonatal and pediatric intensive care patients at Uniklinik RWTH Aachen. Due to limited staff resources, short-stay patients may sometimes not be included because they are discharged before a clinical investigator can link their patient identifiers with the corresponding patient data.

The following variables were stored in a categorized fashion: sex, age or gestational age, admission weight, admission weight percentile, and type of annotation.

Descriptive statistics, including medians, IQRs, and frequency distributions, were computed to characterize the dataset. No inferential statistical analyses were conducted. Statistical analyses were performed using SPSS (version 29.0.0.0; IBM Corp).

Only the clinical investigators who were directly involved in patient treatment had access to the linking file with the identifiers and were able to access the decoded data at all times. The other project partners had unlimited access to the coded dataset.

After synchronization, the data were reviewed for completeness and consistency. Data cleaning involved identifying missing or incomplete values, ensuring consistency across variables, and resolving any discrepancies. Data anomalies, such as outliers (eg, z scores >3), were detected manually or by statistical means. Where appropriate, missing data, duplicate entries, and outliers were removed prior to analyses.

Data were not linked across multiple databases.

In this collaborative project, we successfully established a monocentric high-frequency database of neonatal and pediatric intensive care patients. The dataset incorporates 400 admissions with a total of 3600 treatment days and more than 20,000 manual annotations. We store data from bedside monitoring, ventilation, laboratory results, patient characteristics, and a variety of manual annotations. There is currently no published description of comparable European high-frequency datasets for neonatal and pediatric intensive care patients.

There are various existing databases incorporating neonatal or pediatric ICU data. Many of these databases contain patient-level data but do not provide high-frequency data [36]. The Pediatrix BabySteps Data Warehouse is a large database of more than 1 million neonatal patients [37]. It contains patient data and characteristics recorded on a daily basis. Many of these data are also included in our database. We collect a variety of additional variables, many of which are captured with high temporal resolution at the second or subsecond sampling levels.

Previously published intensive care databases have paved the way for secondary data analyses. The most commonly used datasets are the Medical Information Mart for Intensive Care-III (MIMIC-III) and MIMIC-IV databases [38,39]. These databases contain medical and demographic data from more than 350,000 ICU patients and have been used in more than 2000 secondary data analyses [40]. However, pediatric patients only constitute a small proportion of the MIMIC-III cohort and were completely excluded from MIMIC-IV. The MIMIC-III Waveform Database [41] extends the clinical data from MIMIC-III by waveform data from approximately 30,000 ICU patients, including 7870 neonates. However, only a subset of these recordings has been matched with the clinical data from MIMIC-III [42]. Comparing the IDs from the matched subset with the original one, we found only 6 neonatal patients with both clinical and waveform data.

Specific pediatric datasets exist: the Pediatric Intensive Care database, for example, incorporates data from more than 12,000 neonatal and pediatric patients [43]. Among others, it provides vital signs, medications, laboratory measurements, and lengths of stay. However, these are stored at low recording frequencies (ie, at 1-minute intervals). The investigation of pathophysiologic conditions often benefits from much higher data resolutions [44].

An example of such a high-frequency dataset is the Preterm Infant Cardio-Respiratory Signals Database [45]. It incorporates 1.6 GB of high-resolution data of ECG and chest impedance from 10 spontaneously breathing premature infants. However, the small sample size and restricted parameter set limit its potential applications.

For pediatric patients, a Canadian group established a database of approximately 1200 patients treated in a pediatric ICU [46]. The researchers collected data at higher frequencies compared to other available databases. The data were recorded every 5 seconds from the bedside monitors and every 30 seconds from the ventilators. This represents a major advancement for research involving high-frequency data. However, this time resolution does not allow a reconstruction of waveform data such as ECG or the airway pressure waveform. In previous analyses, we developed algorithms for detection of ineffective efforts during expiration. They constitute one of the most prevalent forms of patient-ventilator asynchronies [47] and result in prolonged mechanical ventilation and length of stay [48]. The detection algorithms are based on synchronous breath-by-breath waveform analyses of airway flow, which requires a very high time resolution. To the best of our knowledge, there are no descriptions of neonatal or pediatric datasets that can provide high-frequency data with an equivalent parameter set for a comparable sample size in Europe.

An approach for collecting high-frequency data was developed for The Hospital for Sick Children in Toronto, Canada, establishing a new data collection system for waveform storage and retrieval [49]. The database used (AtriumDB) holds waveforms from more than 5300 patients. The paper focuses on the technical development and compression techniques, giving only a short overview of the available data. Furthermore, only waveform data were stored through the described setup, without additional data relevant for intensive care.

Besides the high-resolution monitoring and ventilation data, the AIx-Neo-Guard dataset incorporates important diagnoses, laboratory results, patient characteristics, and manual annotations. This enables a detailed reconstruction of patient states and dynamic physiological changes. In particular, machine learning approaches benefit from such a dataset. The large number of manual annotations can be used to train detection algorithms for highly relevant complications. Furthermore, the unlabeled data can be used for unsupervised or self-supervised learning approaches to detect anomalies, identify clusters, or train foundation models. The integration of high-frequency physiological signals with clinical variables offers an important platform for clinically meaningful and translational research. The dataset opens important opportunities toward outcome-oriented research and enhances pathophysiologic understanding [50]. For example, using detection methods for ineffective efforts based on waveforms together with information on the patient’s state can deepen our understanding of the effects of patient-ventilator asynchrony on the patient’s outcome.

The most important limitation is the single-center study design. First, it limits the available study population. Furthermore, diagnostic and treatment practices are based on local standards, which may introduce institutional and performance bias. We are preparing to collaborate with other centers to increase the available data and reduce the influence of local practices. Joint research across institutions is also desirable to allow the validation of algorithms (eg, for clinical decision support) on external data. This can contribute to enhancing the generalizability of results derived from a particular dataset.

So far, the dataset does not incorporate all relevant information on diagnostics and treatment. Fluid management and medication, for instance, are not yet integrated into the parameter set. Furthermore, the duration, for example, of nursing care is not labeled. This, in turn, is a current limitation for the possible use cases of the dataset. Predictive modeling and causal inference often require this valuable information to provide reliable estimates. The legal framework and ethics approval permit the storage of all variables related to diagnostics and treatment in the ICU. We extend the number of collected parameters continuously and plan to incorporate fluid management and medication data in the future.

Due to the rapid growth of data, there are insufficient staff resources to manually annotate the full dataset. The annotations are continuously extended and improved to enhance the quality of data. The set of annotations may furthermore also be enhanced by automated annotation using machine learning approaches. As previous analyses specifically focused on respiratory events, respiratory annotations are overrepresented compared to nonrespiratory annotations. These annotations will have to be made prior to use for nonrespiratory use cases. The data quality is furthermore hampered by technical issues. Occasionally, the ventilator fails to connect to the central monitoring interface. Resolving this issue would require restarting the ventilator. However, during mechanical ventilation, such an intervention would not be ethically justifiable because it could jeopardize the safety of critically ill patients. Similarly, the patient identifiers must be linked to the patient data on the bedside monitor to include the patient in the dataset. The limited staff resources do not allow a clinical investigator to be present at all times to carry out this step. This problem therefore systematically affects short-stay patients who require only short-term treatment or observation (such as those with transient tachypnea of the newborn, those recovering from operative procedures, or those admitted for intoxications). These episodes result in missing valuable data. This introduces a relevant selection bias toward more severe and longer-stay patients and has an impact on the representativeness of the dataset. Specifically, the overrepresentation of infants with longer or more severe courses will influence descriptive analyses and downstream modeling based on this cohort. As a result, findings derived from this dataset should be interpreted as reflecting the subset of neonatal ICU patients for whom complete data capture was feasible, rather than the entire admitted population.

Finally, the dataset was established to promote secondary data analyses. Secondary data research is always prone to bias due to unmeasured confounding and missing data. As the data were collected for a different purpose, the available data do not always meet the requirements of a specific research question, for example, if parameters of interest are missing. In these cases, other data sources must be sought. We hope to reduce this source of bias by pooling data with other centers in the future, increasing the set of available parameters, and continuing data collection and annotation. While secondary data research can support evidence-based medicine, interventional trials are sometimes still necessary.

We are currently working on machine learning algorithms to automatically detect and correct time stamps for nursing rounds and blood sampling. These are the first steps toward an automated improvement of data quality, as these interventions inevitably lead to data anomalies, such as noisy ECG.

To further extend the dataset, information from pregnancy medical history is currently being integrated into the database for all neonates. In the future, the dataset will be further enriched with microbiological data (eg, results from blood cultures) and more comprehensive information on fluid management and medication administration. This will help improve the interpretability and increase the number of research questions that can be addressed using the dataset.

By sharing these database descriptions, we hope to encourage cross-institutional cooperation and invite all interested researchers to contact us, for example, to conduct external validation of algorithms. We hope that collaborations will promote the multicenter pooling of data and thereby enhance the available data. Larger study cohorts and datasets can improve statistical power and are particularly important for investigating rare diseases, for which adequate sample sizes are often difficult to achieve in single-center studies.

Using federated learning could also leverage the described dataset without the need for direct data sharing. Federated data access enables privacy-preserving development of AI algorithms across institutions and could create a pathway toward more robust foundation models for neonatal and pediatric intensive care.

We successfully established a monocentric high-frequency neonatal and pediatric intensive care dataset. The available data have been used for various use cases, with further analyses ongoing. Secondary data research can substantially reduce the need for experimental studies and contribute to evidence-based medicine. We hope that ongoing European Union legislation (European Health Data Space) will create a legal framework for the multicenter pooling of data. This would represent a major advance, as it could rapidly increase the available data, counteract bias, and thereby expand research opportunities and improve generalizability.