Data harmonization framework for neonatal hypoxic-ischemic encephalopathy studies

Abstract

Objectives

To develop a data harmonization framework for neonatal hypoxic-ischemic encephalopathy (HIE) studies and demonstrate its suitability for prognostic biomarker development.

Materials and Methods

Variables were first categorized by chronological stages and then by medical topics. We created a dictionary to harmonize variable names and value coding. We began by merging comprehensive data from 2 landmark nationwide therapeutic hypothermia for HIE trials (2008-2016, 21 sites) in the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) Neonatal Research Network (NRN). The 2 datasets differ in available variables, variable naming and coding, necessitating harmonization. To demonstrate the utility of this data harmonization framework, we computed the distributions of variables and ranked them by the strength of associations with 18- to 22-month outcomes. Associations were measured using Pearson’s correlation analysis. Outcomes were defined as (a) a 5-class variable: survivors with normal, mild, moderate, severe disability, or death and (b) the Bayley-III Scales.

Results

We created a dictionary of 1181 variables on 532 patients across 5 chronologic categories and 60 medical subcategories. The distribution of major predictive and outcome variables, and the variables strongly associated with neurodevelopmental outcomes at 18-22 months were presented. The modified Sarnat scores at the Post-intervention and NICU-discharge stage, and the NRN pattern of MRI injury score showed strong associations with outcome variables.

Conclusion

We designed a data harmonization framework specifically for HIE. Our initial effort in merging 2 iconic clinical trials shows strong predictor-outcome associations, allowing subsequent development of advanced prognostic biomarkers of neonatal HIE.

Background and significance

Hypoxic-ischemic encephalopathy (HIE) is a brain condition occurring in 1-5 out of every 1000 term-born neonates, characterized by insufficient blood flow and oxygen to the brain, potentially resulting in lifelong brain functional disability.^1–4 In 2022, the Consortium Of MRI Biomarkers In Neonatal Encephalopathy (COMBINE) was established by a group of neonatologists, pediatric neurologists, pediatric neuroradiologists, along with computer and data scientists. Its mission is to develop artificial intelligence (AI)-driven prognostic clinical and imaging biomarkers that can predict, during the neonatal stage, whether a specific HIE patient is at risk of developing adverse neurologic outcomes by 2 years of age. Although the introduction of therapeutic hypothermia in 2005 for moderate or severe HIE has reduced morbidity and mortality in high-income countries,^5–8 the 2-year adverse outcome, defined as moderate/severe neurodevelopmental disability or death, continues to negatively impact around one third of moderate or severe HIE patients.^9,10 An accurate prognostic biomarker could serve as the basis for identifying high-risk patients, evaluating treatment effects early, and expediting therapeutic innovations in targeted sub-cohorts, toward eventually improving neurodevelopmental outcomes in HIE.^11–14

One fundamental step in COMBINE is to amass a large-scale and comprehensive database for AI-driven biomarker development. Most HIE biomarker studies to date have used data from only dozens or a few hundreds of patients, with clinical variables often considered in isolation. The lack of comprehensive large-scale diverse data hinders the accuracy and application of AI approaches for biomarker development. The Neonatal Research Network (NRN) within the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) has authorized the COMBINE consortium to use existing data from 2 large-scale nationwide HIE trials. The primary goal of the Late Hypothermia (LH, 2008-2016, NCT00614744, 21 sites) trial was to compare whether the initiation of cooling within 6-24 hours was neuroprotective compared to normothermia,¹⁵ while the Optimizing Cooling (OC, 2010-2014, NCT01192776, 18 sites) trial examined whether longer or deeper cooling was neuroprotective.¹⁰ A total of 532 patients were enrolled from 21 hospitals in the United States (see sites in Figure 1). Each trial collected over 1000 variables, covering pregnancy, delivery, maternal and infant demographics, treatment, infant neuroimaging, follow-up outcome measurements, and other comprehensive information.

Figure 1.

The 21 sites in the LH and 18 sites in the OC trials. Locations are based on Google Maps.

Merging comprehensive and diverse data into a coherent database is, however, not a trivial task.^16,17 Challenges stem from at least 3 aspects: (a) variable availability: some variables are present in only one trial; (b) variable naming: the same variable may have different names across trials when deposited into a central data repository (eg, the variable for maternal race was named MRACE in the LH dataset but RACE in the OC dataset); and (c) coding of variable values: the values for the same variable may differ among different trials (eg, sex coded as M/F in one dataset and 0/1 in another).

This paper presents our benchmark work in the COMBINE consortium to merge and harmonize existing data from these 2 NICHD NRN trials on HIE. The Methods section details (1) the data harmonization framework that creates the dictionary of common data elements (CDE) to harmonize the 2 HIE datasets. The Results section covers (2) the characteristics of the merged database and the dictionary of HIE variables; and (3) results from mass-univariate analysis demonstrating the utility of the merged database.

Methods

Fortier et al. noted that, starting with defining the questions and objectives, at least 6 key steps are required to disseminate and preserve final harmonization products.¹⁶ Cheng et al, citing Stuckenschmidt, highlighted 3 essential components: file format, categorization of variables, and variable naming and value coding.^17,18 The F.A.I.R. principles, denoting findability, accessibility, interoperability, and reusability, are widely recognized as foundational guidelines for data management and exchange.¹⁹ Our development of the data harmonization framework aligns with these 3 guidelines.

Harmonization framework

Variable categorization

In our data harmonization framework, we categorized the variables into 5 categories, reflecting the progression of HIE care: (1) Pre-intervention—screening and baseline information before starting the intervention, such as maternal demographics information, pregnancy history, labor and delivery details, and infant’s condition at birth; (2) Intervention—repeated measurements related to various body systems during the intervention, such as temperature, cardiovascular function, and blood values; (3) Post-intervention—measurements after the intervention, such as continuous blood measurements, brain MRI, and neurological exams; (4) Neonatal Intensive Care Unit (NICU) Discharge—records of clinical diagnosis at death or discharge; and (5) Follow-Up—study-specific measurements at the follow-up visit and interim medical history. Categories 1-4 occur during the neonatal period. In addition to this chronological categorization, we further subdivided the variables based on specific medical topics within each category.

Names and coding of variables

To ensure clarity, we used English-comprehensive naming conventions for the variables. Optionally, for variables with the same meaning across different categories, prefixes for categories may be added for further distinction. To precisely describe a variable, additional details may be necessary, such as the measurement unit for measurement variables or whether the screening criteria are inclusive or exclusive. This information is added as suffixes to the variable names.

Data types of variables

Data types are important in data analysis. In addition to the typical Boolean, numerical, date, time, and text, there are several ordinal and nominal types for categorical variables that require standardized coding. The coding for these categorical variables was also designed to be English-comprehensive. For ordinal variables, in addition to standardized coding, we had implemented a mechanism that allows for different ranking strategies based on specific contexts.

Missing variables: union-based approach

Some variables may not be included from certain datasets. We took the “union” of all the variables from 2 datasets. The advantage was to keep all variables as long as they existed in at least one dataset. This better prepared our dictionary of variables when future datasets are to be merged. The disadvantage, though, was the existence of missing values if a variable does not exist in a dataset. This is addressed as described next.

Missing values: keep the original data

Missing values were anticipated following the merging of datasets, either due to their absence in the original sources or because certain variables were not present in all datasets after harmonization. To preserve the integrity of the original data, we did not infer the missing values in either scenario. As a result, all patient-level data remained unaltered and reflected actual records. The imputation of missing data was deferred to the data analysis stage.

Derived variables to reconcile different data types

Some variables recorded as different data types might be eligible for inference from one another. In addition to preserving the original data without inference, we created derived variables, where appropriate, and to the best of our knowledge, to infer values that could be further harmonized. For example, the 5-minute Apgar scores were recorded as numerical variables in both the LH and the OC trials. However, only the OC trial included an additional variable indicating whether the 5-minute Apgar score was less or equal to 5. In addition to retaining the OC-specific variable, Apgar5minLessEq5, we created a derived variable, Apgar5minLessEq5_deriv, inferred from the numerical 5-minute Apgar scores.

Repeated variables

Some variables were collected repeatedly over time. One type of such time-series variables was measured at regular time intervals, such as temperature being measured at 15, 60, and 1440 minutes after starting of the intervention. We included the common unit of time as the postfix of the indexing variable, such as temperatureTimeSlot_min. Another type of repeated variables recorded the natural incidences of events, such as serious adverse events. These events occurred randomly, and the indices could only represent the order of the occurrence. We used “number” as the postfix of the indexing variables, such as adverseEventNumber. In spreadsheet presentation, the repeated variables were flattened by appending the indices as the postfix of the variable, such as skinTemperature_C_15min, skinTemperature_C_1440min, and SAEOther_1.

Mass-univariate analysis to show data utility

To demonstrate the utility of our data harmonization framework, we used Pearson’s correlation coefficient analysis to explore the relationships between predictor and outcome variables. In this study, adapted from the LH and OC studies, we defined the outcome variables as the 5-level disability level or death (normal, mild, moderate, severe, and death), and the Bayley-III Composite Cognitive Scales of survivors at 18-22 months of age. Our analysis focused on variables collected at the Pre-intervention, Post-intervention, NICU Discharge, and Derived Data stages. Conventionally, the significance threshold is assigned as P-value < 0.05. However, considering the inclusion of over 1000 variables in the framework, we applied Bonferroni correction and assigned the significance threshold as P-value < 10⁻⁵. For the variables meeting this threshold, we ranked the variables based on their corresponding coefficient of determination R² in Pearson’s correlation coefficient analysis.

Data privacy

Data privacy is always a major concern when dealing with medical data. Besides the usual already-anonymized site names and subject names replaced with anonymized identifiers, the privacy of the medical practitioners should be protected as well. We manually reviewed the text and removed the names and initials of the medical practitioners. Furthermore, information about date and time can be sensitive to privacy issues. The date and time is transformed to the days since birth, with precision to the hour. Only the year and month are revealed in the birth date.

Versioning

We recognize that different studies have varying aims. It is foreseeable that other studies may include variables not present in our framework, requiring expanding the current variables. Moreover, the medical subcategories may also need to be extended to account for the uniqueness of these new variables. In addition, the coding of existing ordinal and nominal variables, such as antibiotics, may need expansion as well to accommodate new values in the future. Lastly, it is possible that the chronological categories may be different for specific needs in the future. We use semantic versioning for different versions of the data harmonization framework.²⁰

Compliance to the F.A.I.R. principles

The F.A.I.R. principles stand for findability, accessibility, interoperability, and reusability.¹⁹ Our data harmonization framework explicitly addresses each principle as follows: (a) Findability: We assigned persistent identifiers in metadata and data dictionary, and publicly release them in GitHub version control and Supplementary Material; (b) Accessibility: All metadata, data dictionary, and software are retrievable through open GitHub repositories using standard HTTPS protocols without authentication barriers; (c) Interoperability: Our framework employed standardized naming conventions and controlled vocabularies for categorical variables, facilitating integration with other HIE datasets. The hierarchical organization by chronological stages and medical topics creates a formal, shared knowledge representation structure; and (d) Reusability: We described variables with detailed attributes in our data dictionary (Supplementary Material). All released materials are provided under permissive licensing (CC BY 4.0 for results and MIT license for software). The comprehensive documentation of our harmonization process establishes detailed provenance. To comply with both the F.A.I.R. principles and the intellectual property of the original datasets, we adopted the “separation of the Datasets, Software, and Results” strategy. This strategy is described in the Data Availability Section and Code Availability sections.

Results

Variable categorization and dictionary construction

Figure 2 illustrates the categorization of variables in our data harmonization framework, grouped into 5 major categories during the clinical course and 60 subcategories based on medical topics from the LH and OC trials. The 5 major categories, along with their respective color codes, are as follows: (in the neonatal period) Pre-intervention: light peach, Intervention: light blue, Post-intervention: yellow, and Neonatal Intensive Care Unit (NICU) Discharge: pink; and (after the neonatal period) Follow-Up: green. A total of 1181 variables remain after excluding those with conceptually similar content.

Figure 2.

The categorization of variables in the LH and OC trials. The left-most column represents the 5 color-coded categories based on the clinical course. The 60 subcategories are presented with icons and descriptions. The numbers in parentheses represent the number of variables in each category and subcategory. There are 1181 variables in total. MRI, magnetic resonance imaging; GMFCS, Gross Motor Function Classification System.

As depicted in Figure 3, our framework includes 956 variables from the LH trial and 1091 variables from the OC trial, and 14 derived variables, such as derived total modified Sarnat scales and averaged MRI NRN pattern of injury scores. After taking a union of these variables, our merged database contains a total of 1181 variables.

Figure 3.

Distributions of the variables in the LH and OC trials.

Data type plays a crucial role in data analysis. Table 1 shows the number of variables for each data type. In addition to common types such as Boolean, number, date, time, or text, various ordinal and nominal types are available for harmonizing categorical variables.

Table 1.

Distributions of different data types.

Data type	Count	Example	Description
Boolean	401	Apgar5minLessEq5	5-minute Apgar score ≤ 5
number	207	birthWeight_g	birth weight
ordinal	202	disabilityLevelDeath	disability level or death
nominal	119	anticonvulsants	anticonvulsants
text	69	screenComment	screening comments
date	123	birthDate	birth date
time	60	birthTime	birth time
Total	1181

Table 2 presents excerpts from our constructed data dictionary in our data harmonization framework. The full table is in Supplementary Material S1, S2, and S4. Variable names that are different between the LH and OC datasets (the last 2 columns in Table 1) are harmonized into the same variable name in this dictionary (the second column in the table).

Table 2.

Excerpts of the data dictionary.

Subcategory		Harmonized variable	Data type	Description
	Screening	birthWeightLessEq1800g_e	Boolean	Exclusion: birth weight ≤ 1800 g
		first60MinAnyBloodGasPHLessEq7_i	Boolean	Inclusion: pH ≤ 7 in any blood gas (cord, postnatal) within the first 60 minutes
	Pregnancy history	Gravida	Number	Gravida
		antepartumHemorrhage	Boolean	Antepartum hemorrhage
	Labor and delivery	laborOnsetDate	Date	Date of labor onset
		laborOnsetTime	Time	Time of labor onset
		deliveryMode	Nominal	Final mode of delivery
	Birth	birthweight_g	Number	Birth weight
		birthHeadCircumference_cm	Number	Birth head circumference
	Temperature	temperatureTimeSlot_min	Number	Time slot of the measurement (unit: minute)
		esophagealTemperature_C	Number	Esophageal temperatures
	Infection	positiveCultureNumber	Number	Infection incidence number
		positiveCultureDate	Date	Date of positive culture
		positiveCultureTime	Time	Time of positive culture
	Neuro exam	post_NeuroExamPosture	Number	Posture in neuro exam in Post-intervention
		post_NeuroExamMoro	Number	Moro reflex in neuro exam in Post-intervention
	Imaging report	headSonogramResult1	Nominal	The first result in the head sonogram report
		headCTResult1	Nominal	The first result in the head computer tomography report
		brainMRIResult1	Nominal	The first result in the brain MRI report
	MRI	MRIAvailable	Boolean	MRI available
		MRINRNPatternOfInjury	Ordinal	NRN pattern of injury score
	Status	status	Nominal	Status of the infant
		deathCause	Nominal	Cause of death
	Bayley-III	BayleyIIIInEnglish	Boolean	Bayley–III exam was conducted in English
		BayleyIIICognitiveComposite	Number	Bayley–III Composite Cognitive Scale
	Outcome	moderateSevereDisabilityOrDeath	Boolean	Moderate severe disability or death
		disabilityLevelDeath	Ordinal	Disability level or death (normal, mild, moderate, severe, death)

Subcategory		Harmonized variable	Original LH variables	Original OC variables
	Screening	birthWeightLessEq1800g_e	LH2WGHT	OC2WGHT
		first60MinAnyBloodGasPHLessEq7_i	LH2PH	OC2PH
	Pregnancy history	gravida	LH4GRAV	OC4FRAV
		antepartumHemorrhage	LH4HMRG	OC4ANTE
	Labor and delivery	laborOnsetDate	LH4LBDT	N/A
		laborOnsetTime	LH4LBTM	N/A
		deliveryMode	LH4MODE	OC4MODE
	Birth	birthweight_g	LH5BTWGT	OC5BWHGT
		birthHeadCircumference_cm	LH5HC	OC5HCIRC
	Temperature	temperatureTimeSlot_min	L6ATMPRD	OC6TINTV
		esophagealTemperature_C	L6AESPHT	OC6TESOT
	Infection	positiveCultureNumber		OC9IPCNU
		positiveCultureDate	L6FPADT	OC9IDATE
		positiveCultureTime	L6FPATM	OC9ITIME
	Neuro exam	post_NeuroExamPosture	LH11PO_1	OC11CPOS
		post_NeuroExamMoro	LH11MR_1	OC11CPRM
	Imaging report	headSonogramResult1	LH9HSREA	OC12HSRA
		headCTResult1	LH9HCREA	OC12HCRA
		brainMRIResult1	LH9BMREA	OC12BMRA
	MRI	MRIAvailable	LM1AVAI	OM1MRIA
		MRINRNPatternOfInjury	LM3PTINJ	OM3PATINJ
	Status	Status	LH12STAT	OC13STAT
		deathCause	LH12DTCA	OC13COD
	Bayley-III	BayleyIIIInEnglish	NF9ABSEN	NF9ABSEN
		BayleyIIICognitiveComposite	NF9ABSCC	NF9ABSCC
	Outcome	moderateSevereDisabilityOrDeath	disab_die	disab_die
		disabilityLevelDeath	N/A	N/A

Subcategories in the leftmost column are examples from Figure 2. The original LH/OC variable names in the rightmost 2 columns were harmonized in the second left column. MRI, magnetic resonance imaging; N/A: not available; NRN, Neonatal Research Network.

Characteristics of the merged database

Figure 4 presents the distributions of 20 variables in the merged database. In the context of prognostic biomarker development, the selected 20 variables include variables on basic demographics and birth conditions, clinician’s evaluation of severity, and outcome measures.

Figure 4.

Distributions of 20 variables in the merged database from the LH and OC trials.

Utility of the merged database

Tables 3 and 4 highlight the results of the mass-univariate analyses. A total of 341 variables from Pre-intervention, Post-intervention, and NICU Discharge were analyzed using Pearson’s correlation coefficient analysis. These results demonstrate the utility of the merged database for the future development of prognostic biomarkers. For clarity, variables with definitions too similar to the listed ones, such as the scores from the subcategories of the modified Sarnat grading scales or specific regions described in the MRI reports, were excluded. The full results of all the 341 variables are in Supplementary Material S3. The color coding indicates the corresponding categories of the variables in Figure 2.

Table 3.

Association with disability level or death (normal, mild, moderate, severe, and death).

Category	Variable	r	R 2	P-value
DC	dischargeTotalModifiedSarnatScore	0.722	0.522	<10−74
DC	dischargeFullNippleFeed	−0.585	0.342	<10−31
Post	MRINRNPatternOfInjuryAvg	0.568	0.322	<10−37
Post	post_TotalModifiedSarnatScore	0.557	0.310	<10−35
Pre	pre_TotalModifiedSarnatScore	0.455	0.207	<10−25
Pre	encephalopathyLevel	0.422	0.178	<10−22
DC	homeTherapyStatus	0.391	0.153	<10-11
DC	dischargeAnticonvulsants	0.375	0.141	<10−16
Pre	Apgar5min	−0.352	0.124	<10−15
Pre	Apgar10min	−0.346	0.120	<10−12
DC	dischargeHomeTherapyGastrostomyTubeFeed	0.343	0.118	<10-7
DC	dischargeHearingTestNormal	−0.314	0.099	<10-10
Pre	firstPostnatalBloodGasPH	−0.303	0.092	<10−10
Post	post_BloodValueASTSGOT_UPerL	0.297	0.088	<10-5
Pre	at10MinChestCompression	0.289	0.084	<10−8
DC	dischargeEEGAbnormalBackgroundActivity	0.285	0.081	<10-8
Pre	Apgar1min	−0.283	0.080	<10−9
Post	post_BloodValueALTSGPT_UPerL	0.283	0.080	<10-5
DC	dischargeSeizure	0.275	0.076	<10−9
DC	dischargeVentilator_day	0.258	0.066	<10−8

The abbreviations in the Category column represent the different chronological categories of the variables in Figure 2 (Pre: Pre-intervention, Post: Post-intervention, DC: NICU discharge). r: Pearson’s correlation coefficient; R²: coefficient of determination; P-value: P-value from each univariate Pearson’s correlation coefficient analysis. The P-value was adjusted using the Bonferroni correction, with the significance threshold set at P-value <10⁻⁵.

Table 4.

Association with the Bayley-III Composite Cognitive Scale.

Category	Variable	r	R 2	P-value
DC	dischargeHomeTherapyGastrostomyTubeFeed	−0.364	0.133	<10−7
Post	MRINRNPatternOfInjuryAvg	−0.342	0.117	<10−11
DC	dischargeTubeFeedingDuration_day	−0.338	0.114	<10−11
DC	dischargeTotalModifiedSarnatScore	−0.333	0.111	<10−11
Post	post_TotalModifiedSarnatScore	−0.297	0.088	<10−8
DC	homeTherapyStatus	−0.283	0.080	<10-6
DC	dischargeFullNippleFeed	0.269	0.073	<10−5

The color coding and heading are the same as Table 3.

As shown in Table 3, the total modified Sarnat grading scales from all 3 stages, MRI NRN Pattern of Injury derived from the Post-intervention stage, Apgar scores from the Pre-intervention stage, and several other variables were significantly associated with the primary outcome, defined as 5-level disability level or death (normal, mild, moderate, severe, and death).

Additionally, Table 4 displays correlations between the variables and the numerical Bayley Scales of Infant Development, Third Edition—Composite Cognitive Scale (Bayley–III Cognitive) (see Supplementary Material S3 for Bayley–III Language and Motor scales).²¹ The total modified Sarnat grading scales in the Post-intervention and NICU-discharge stages, the MRI NRN Pattern of Injury derived from Post-intervention stage, and feeding difficulty variables were among the top-ranked factors associated with Bayley–III Cognitive scales.

Discussion

The goal of COMBINE is to develop artificial intelligence (AI)-driven prognostic clinical and imaging biomarkers that can predict, during the neonatal stage, whether a specific HIE patient is at risk of developing adverse neurological outcomes by 2 years of age. Large-scale, comprehensive, diverse, and representative data are essential for developing quantitative, objective, and accurate prognostic biomarkers for HIE. When multiple trials or studies exist, a natural extension, as presented in this paper, is to harmonize and merge existing datasets. This approach facilitates the creation of a unified database with a larger sample size, covering not only more sites but also more studies, more comprehensive data elements, and increasing the diversity and heterogeneity of the data. This harmonized database can also enhance the generalizability of results to real-world applications.

Selecting representative datasets is an essential step before starting the construction of a data harmonization framework.¹⁶ The LH and OC trials are the follow-ups to the landmark trial that established therapeutic hypothermia as the current norm for HIE treatment.⁵ They aimed to further explore the potential of therapeutic hypothermia for a broader group of neonatal HIE patients who cannot be treated within the typical 6 hours of life (LH), as well as to optimize the intervention parameters (OC). As illustrated in Figure 1, the 21 sites reflect the geographic diversity of the population across the United States in these 2 studies. Furthermore, the 60 medical subcategories, encompassing a total of 1181 variables across the 5 major chronological categories, demonstrate the comprehensive considerations involved in conducting the HIE-related clinical trials, including the rigorous design of the trials, data definitions, and the training of neuro examiners, psychologists, and psychometricians. The representativeness and comprehensiveness of the LH and OC trials make them the ideal datasets to create the proposed harmonization framework.

An important component of our data harmonization framework is the creation of a dictionary of variables, as presented in Table 2. We recognized that while some variables may exist in only one trial, they remain crucial due to the uniqueness of each clinical study. Therefore, we retained a union of the variables from both trials. We resolved the differences in variable names by creating standardized data element names in this dictionary, followed by standardizing the coding of each variable across trials. This effort aligns with similar dedications to merge and harmonize heterogeneous datasets, such as medication in the intensive care units,²² cancer,²³ stroke,²⁴ pediatric sepsis,²⁵ and SARS-CoV-2.^26,27 The creation of common data elements (CDE) or standardized dictionaries for various diseases is also a key NIH effort.²⁸

The harmonized dictionary of HIE variables we created in this paper can be used in at least 2 scenarios. First, it can serve as a reference to merge additional existing HIE datasets, further expanding our unified database. Potential large-scale datasets include the BONBID-HIE dataset (237 patients from 2 sites and 20+ variables),²⁹ the HEAL-HIE data (500 patients from 17 sites),³⁰ etc. These datasets may offer variables not present in the LH and OC trials, so the dictionary will be continuously expanded beyond the current 1181 variables as new datasets are merged. In addition, the dictionary of HIE variables can guide the future design of HIE clinical trials. By using the same variable names, definitions, and coding conventions in the design phase, newly acquired data can be easily merged into the dynamically expanding database.

Currently, there is no widely accepted standard dictionary for HIE clinical studies. A parallel work by Newborn Brain Society Guidelines and Publications Committee found that only 4 variables (birth weight, gestational age, 1- and 5-minute Apgar scores) were common across 22 HIE registries, with inconsistent naming and coding, highlighting the urgent need for harmonization.⁴ Our dictionary (Figure 2, Table 2, and Supplementary Material S1, S2, and S4) aligns with the National Institute of Health’s Common Data Elements (CDE) Initiative, which supports data standardization to facilitate cross-study comparison and data sharing.²⁸ Integrating into the NIH CDE Repository will promote the interoperability and broad adoption of the framework. In addition, several frameworks and applications, such as the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) by Observational Health Data Sciences and Informatics (OHDSI),³¹ can be integrated with electronic healthcare records (EHR) systems, such as Epic’s analytic database storing as a Microsoft SQL database (Clarity) or a cloud-based database (Snowflake),^32–34 to extract EHR and transform them into harmonized variables for further utilization of the integrated data. It is anticipated that our data harmonization framework will yield real-world impact through collaboration with relevant stakeholders and seamless integration with their applications.

In the literature, there are studies exploring the ability of clinical variables to predict HIE outcomes. Broadly speaking, our results (Figure 4; Tables 3 and 4) are consistent with previous reports on neuroimaging scores,^{1,11,13,35–38} Sarnat grading scales,^39–46 and Apgar scores at 5 and 10 minutes after birth,⁴⁷ in that the associations between these variables and the primary outcome (disability level or death) are significant.

We envision several directions for future exploration. First, other HIE trials may include additional variables, such as fine-grained electrocardiogram (ECG) and electroencephalogram (EEG) monitoring, different timing of assessments or different expert interpretations of neuroimaging. We will continuously expand our dictionary as we merge new datasets into our database. In addition, we conducted only a basic mass-univariate analysis, as the primary goal of this paper is not to perform an extensive analysis but to present the utility of the merged database. Our ultimate objective in the COMBINE consortium is to develop multivariate AI analyses that can thoroughly examine the combinatorial effects of over 1000 variables and enhance the predictive value for each individual patient. Furthermore, our data harmonization framework has the potential to extend beyond biomarker development. It can also function as a knowledge base or support an HIE-specific chatbot to further enhance clinical training and practice.

Conclusion

In conclusion, we have established a data harmonization framework tailored for HIE-related projects. Utilizing this framework, we successfully combined data from 2 HIE clinical trials from the NICHD NRN into a cohesive database. The resulting dictionary, which contains 1181 variables, serves as a foundation for integrating additional existing HIE datasets and recommending common data elements for future HIE trials or studies. This framework and the merged database will facilitate the subsequent development of prognostic HIE biomarkers.

Author contributions

Chuan-Heng Hsiao (Data curation, Formal analysis, Software, Validation, Visualization, Writing—original draft), Anna N. Foster (Data curation, Writing—review & editing), Scott A. McDonald (Data curation, Writing—review & editing), Rutvi Vyas (Data curation, Visualization, Writing—review & editing), Aseelah Ashraf (Data curation, Writing—review & editing), Rina Bao (Data curation, Writing—review & editing), Lena Tran (Data curation, Writing—review & editing), Ankush Kesri (Data curation, Writing—review & editing), Erfan Darzidehkalani (Data curation, Writing—review & editing), Matheus D. Soldatelli (Data curation, Writing—review & editing), Jeanette O. Auman (Data curation, Writing—review & editing), Janet S. Soul (Supervision, Writing—review & editing), Lina F. Chalak (Supervision, Writing—review & editing), C. Michael Cotten (Conceptualization, Funding acquisition, Methodology, Supervision, Writing—review & editing), Seetha Shankaran (Conceptualization, Funding acquisition, Methodology, Supervision, Writing—review & editing), Abbot Laptook (Conceptualization, Funding acquisition, Methodology, Supervision, Writing—review & editing), P. Ellen Grant (Conceptualization, Funding acquisition, Methodology, Supervision, Writing—review & editing), and Yangming Ou (Conceptualization, Funding acquisition, Methodology, Supervision, Writing—review & editing)

Supplemental material

Supplementary materials are available at JAMIA Open online.

Funding

This work was supported, in part, by the National Institute of Health by grant R61NS126792.

Conflicts of interest

All authors declare that they have no financial or non-financial competing interests in relation to this paper.

Data availability

The definition of the metadata and the data dictionary of the framework is maintained as spreadsheets and REDCap format and can be found in Supplementary Material S1, S2, and S4.⁴⁸ With the separation of Datasets, Software, and Results, the Results, including the definition of the metadata and the data dictionary of the framework, and the statistical analysis results, are licensed under CC BY 4.0 and can be found at the public repository https://github.com/i3-research/COMBINE-harmonizer/tree/v1.1.0/results. The LH dataset and the OC dataset have been released to the public.^49,50

Code availability

With the separation of Datasets, Software, and Results, the Software for this study is available on GitHub under the MIT License and can be accessed via this link https://www.github.com/i3-research/COMBINE-harmonizer/tree/v1.1.0.