Automated phenotyping of congenital heart disease for dynamic patient aggregation and outcome reporting

Abstract

Objectives

Accurate characterization of patients with congenital heart disease is fundamental to research, outcomes reporting, quality improvement, and clinical decision-making. Here we present an approach to computing the anatomy of patients with congenital heart disease based on the whole of their diagnostic and surgical codes.

Materials and Methods

All diagnostic and procedure codes for patients cared for between 1981 and 2020 at Boston Children’s Hospital were extracted from a database containing diagnostic codes from echocardiograms, and procedural codes from surgical and catheterization procedures. The pipeline sequentially (1) mapped each of the 7500 native codes to algorithm codes; (2) computed the parent anatomy for each study using a pre-defined hierarchy; (3) computed the parent anatomy for the patient, based on highest ranking parent anatomy; and (4) computed the subcategories and mandatory co-variate findings for each patient. Thereafter, diagnostic accuracy of 500 unseen patients was adjudicated against clinical documentation by clinical experts.

Results

A total of 514 541 echocardiograms on 161 735 patients were available for this study. Phenotypes of congenital cardiac diseases were assigned in 84 285 patients (52%), and the remainder were computed to have normal anatomy. Clinicians agreed with algorithm assignments in 96.4% (482 of 500 patients), with disagreements most often representing definitional differences. An interactive dashboard enabled by the output of this algorithm is presented.

Conclusions

The computation of detailed congenital heart defect phenotypes from raw diagnostic and procedure codes is possible with a high degree of accuracy and efficiency. This framework may enable tools to support interactive outcomes reporting and clinical decision support.

Introduction

The accurate identification of the primary anatomic diagnosis of a patient is fundamental for clinical decision-making, benchmarking outcomes, research, and quality improvement. A standardized nomenclature for congenital heart defects (CHD) was first introduced by Maude Abbott’s Atlas of Congenital Cardiac Disease in 1936.¹ Don Fyler then established a codified classification system of anatomic findings and interventions, placing codes prominently within clinical reports at Boston Children’s Hospital; this substantively increased clinician involvement in the sophistication, accuracy, and completeness of the schema.^2–4 This approach has been advanced by the International Society for Nomenclature of Paediatric and Congenital Heart Disease, which has iteratively developed and field tested the International Paediatric and Congenital Cardiac Code (IPCCC).⁵ At long last, this schema has been incorporated as the congenital cardiac component of the 11th revision of the World Health Organization’s International Classification of Diseases (ICD-11).⁶ Both IPCCC and ICD-11 provide a hierarchical infrastructure that provides type-subtype relational insights. In this way, more than 17 000 terms, each describing a morphological phenotype, procedure, symptom, or genetic syndrome, are used to provide nuanced details of 318 “parent” diseases.⁷ Synonyms are also described, allowing for the harmonization of diseases and interventions across nomenclature types. The schema also incorporates Extension Codes, used to describe the size, degree, or subtype of a specific defect. For example, a common atrioventricular canal (CAVC) can be described as “right dominant,” “left dominant,” or “balanced” or a ventricular septal defect as “small,” “moderate,” or “large” using extension codes. Linkages to other standard terminologies, such as Systematised Nomenclature of Medicine Clinical Terms (SNOMED-CT), are also provided.⁸ Thus, the Fyler coding system, the IPCCC, and ICD-11 represent robust, scalable information-bearing infrastructures on which to categorize and compare patients.

In CHD, the assignment of primary diagnosis is an essential first step for deciding management and assessing prognosis. In clinical practice, a patient is described according to their primary diagnosis (eg, atrioventricular canal) with modifiers (eg, balanced, complete CAVC with pulmonary stenosis), which characterizes the patient in clinical documents and on problem lists. However, existing coding systems for congenital heart disease and its treatments, including Fyler, IPCCC, and ICD-11, are not optimized to compute the primary diagnosis that is singular, detailed, and able to handle discrepancies. Multimodal automated phenotyping has been successfully performed from diagnostic and billing codes,⁹ clinical documentation, and laboratory information in cardiology,¹⁰ sepsis,¹¹ pulmonary hypertension,¹² and others,¹³ with variable accuracy and specificity.

The purpose of this work was to develop and validate a system to compute the detailed phenotype (ie, parent anatomic disease, relevant subcategories, and co-variates) of the native anatomy and first procedure in a large cohort of patients at a single institution. Following its validation, we present an interface demonstrating its utility in a clinical decision support tool.

Methods

Data collection and software

This study was approved by the Institutional Review Board at Boston Children’s Hospital (P00035157). All patients undergoing at least one echocardiogram at Boston Children’s Hospital between January 1981 and March 2020 were included. For each patient, all diagnostic and procedural codes were electronically extracted from an institutional database containing codes from echocardiograms, surgeries, and catheterizations and analyzed as below. We note that although these are listed as “Fyler codes,” this coding system has been adopted as ICD-11. Data from 500 random patients were not used during the development of the algorithm and were used for validation of the algorithm. The algorithm was developed using Python and its libraries, including NumPy and Pandas.

Anatomic phenotyping

We created a 4-phase algorithm for phenotyping the cardiovascular anatomy of a patient (Figure 1). The algorithm was repeatedly modified through internal validation using data from random patients other than the 500 patients for final evaluation. Data from 1500 patients were used in the internal validation process.

Figure 1.

Schematic of coding approach. Detailed Fyler codes (column 1, left) were first mapped to one or more algorithm codes (Phase 1). The highest ranking (ie, most defining) algorithm code was then computed for each event (eg, echocardigram) according to a pre-defined hierarchy (Phase 2). The defining anatomy codes were then analyzed at a patient level, where conflicts were also resolved (Phase 3), resulting in a single patient-level defining anatomy. Following assignment of the patient level anatomic diagnosis, subcategories and covariates were assigned (Phase 4). Figure by KaiOu Tang.

Phase 1. Feature extraction from existing codes to algorithm codes

First, in order to accommodate the flat structure of the Fyler coding system (ie, that one Fyler code may contain more than one anatomic finding), as well as the evolution of Fyler coding concepts over time, each of the 7865 Fyler codes was assigned one or more algorithm codes (Table S1). For example, “0606—Double outlet right ventricle (DORV) with atrioventricular canal defect” was assigned 2 algorithm codes, one representing DORV and a second representing CAVC. As the Fyler coding system also natively describes some findings with modifiers such as “ruled out,” algorithm codes were each assigned a ruled-in or ruled-out modifier.

Each algorithm code exists in a relational schema of parent–child relationships that inherited the IPCCC hierarchy⁶ with minor modifications (Table S2). Based on its position within this schema, each algorithm code inherited properties as either a diagnosis or procedure, an anatomic or functional finding, a congenital or acquired disease, and an abnormal or normal finding.

Phase 2. Establish the study-level defining anatomy

Among the 7865 Fyler codes, 129 defining anatomies were identified (see selected subset Table 1 and full listing in Table S3), hierarchically related according to IPCCC classification. To identify a defining anatomy in each study (eg, echocardiogram), all algorithm diagnostic codes were considered. Because of their misleading effect on clinical phenotyping, findings of patent ductus arteriosus or patent foramen ovale were disregarded on fetal echocardiograms; codes describing trivial valvar stenosis were similarly disregarded. Thereafter, the highest ranked defining anatomy (of the 129) was computed for each study. When none of the 129 defining anatomies were coded for, a defining anatomy of “Normal” was assigned to the study.

Table 1.

Selected subset of schema of defining anatomies, their subcategories, and mandatory covariates.

Major anatomy	Value	Subcategory	Mandatory covariate	Comments
Double-outlet right ventricle	127	Ventricular septal defect type Subaortic Subpulmonary Uncommitted Doubly committed Intact ventricular septum	Common atrioventricular canal Total anomalous pulmonary venous connection Mitral atresia Pulmonary atresia Aortic atresia
Hypoplastic left heart syndrome	125	Aortic valve Atresia Stenosis Mitral valve Atresia Stenosis	Intact atrial septum
Aortic atresia	123		Ventricular septal defect
Pulmonary atresia with intact ventricular septum	117		Coronary artery fistula Hypoplastic right ventricle
Common atrioventricular canal	110	Ventricular balance Balanced Unbalanced Ventricle dominance N/A Right Left	Overriding aortic valve Straddling mitral valve, Straddling tricuspid valve	Partial AVC and canal type VSD are not included
Interrupted aortic arch	109	Type A, B, or C	Ventricular septal defect
D-looped transposition of the great arteries	107	Type With intact ventricular septum With ventricular septal defect With ventricular septal defect and pulmonary stenosis	Aortic stenosis Left-ventricular outflow obstruction Unusual coronary pattern
L-looped transposition of the great arteries	106	Type With intact ventricular septum With ventricular septal defect	Pulmonary stenosis Ebstein’s anomaly
Total anomalous pulmonary venous connection	104	Pulmonary venous obstruction Obstructed Unobstructed Type Supracardiac Cardiac Infracardiac Mixed
Coarctation of the aorta, moderate to severe	101	Severity Moderate Severe	Ventricular septal defect Aortic stenosis
Ventricular septal defect	98		Conal ventricular septal defect Conoventricular ventricular septal defect Muscular ventricular septal defect Canal ventricular septal defect Restrictive ventricular septal defect Overriding aortic valve Straddling mitral valve Straddling tricuspid valve
Normal cardiac anatomy	0		Cardiomyopathy Endocarditis Cardiac tumor Cardiovascular trauma Aortic regurgitation Mitral regurgitation Pulmonary regurgitation Tricuspid regurgitation Left ventricular dysfunction Right ventricular dysfunction Arrhythmia Pulmonary hypertension Aortic aneurysm Aortic dissection

The complete schema can be found in Table S3. AVC, atrioventricular canal; VSD, ventricular septal defect.

Phase 3. Establish state-dependent, patient-level defining anatomy

The timing of each study was contextualized into one of 2 states. The first state included all prenatal and post-natal studies up to the date of their first intervention, defining the “native anatomy.” The second state was following the first intervention, defining the “post-intervention anatomy.” Native and post-intervention anatomies were computed as the highest ranked defining anatomy across all studies during each state. When there was a discrepancy between native and post-intervention anatomies, native anatomy was used to categorize the patient for cohort assignment. When a patient’s echo-defined anatomy was “Normal cardiac anatomy,” diagnostic Fyler codes extracted from the surgical or catheter database were used for establishing the defining anatomy. Otherwise, only codes extracted from the echocardiographic database were used.

Phase 4. Assignment of subcategories and mandatory co-variates

We created a matrix of clinically important subcategories and mandatory covariates for each parent diagnosis (Table 1). For example, a patient with heterotaxy, DORV with subaortic ventricular septal defect (VSD) and a balanced CAVC was assigned a parent diagnosis of DORV. The position of the VSD (eg, subaortic) would represent a subcategory of the parent diagnosis. For this parent diagnosis, CAVC and totally anomalous pulmonary venous connection were assigned a priori as mandatory covariates, findings that are clinically important and must be described as ruled-in or ruled-out to wholly describe a patient with a given parent diagnosis. When mandatory covariates were also defining anatomies (eg, CAVC), the subcategories and co-variates of each were also included and nested within the phenotypic description (eg, ventricular dominance of the CAVC would be described). This allowed the algorithm to losslessly describe any permutation of CHD anatomic findings (including combinations that have not yet been described) within a relatively compact schema.

When there was disagreement in subcategory (eg, subaortic vs subpulmonary VSD) between studies, the subcategory was computed as “Unclear” unless only a single echocardiogram was in disagreement. When the subcategory of interest was not mentioned in any study, a subcategory of “Unmentioned” was computed.

Surgical phenotyping

Following the steps above, the first recorded procedure was categorized according to a pre-defined hierarchy, whether operative or catheterization (Table 2). As with anatomic phenotypes, each surgical procedure was assigned relevant subcategories and mandatory covariates.

Table 2.

Subset of schema of procedures, their subcategories. and mandatory covariates.

	Hierarchy	Subcategory	Mandatory covariate
Heart transplant	249
Fontan	248	Intracardiac: Fontan—Intracardiac; Extracardiac: Fontan—Extracardiac; Atrial: Fontan—RA to PA, Fontan—RA to RV	Fontan—Fenestrated
Super Glenn	247
1.5 ventricle repair	246
Cavopulmonary connection	245
Hybrid stage 1 palliation	244
Stage 1 palliation	243	Sano: RV to PA conduit; BTS: Blalock-Taussig shunt
TOF repair	240		Pulmonary transannular patch, RV to PA conduit, Interventricular fenestration
AVC repair	235	Single patch: Complete common atrioventricular canal repair—Patch—Single patch; Two patch: Complete common atrioventricular canal repair—Patch—2 patches; Australian: Complete common atrioventricular canal repair—Direct suture (Australian)	Aortopulmonary anastomosis
Double switch operation	232	Atrial switch Senning procedure; Mustard procedure
Aortic arch repair	218	Procedure [Patch: Aortic arch repair—Patch, Coarctation repair—Patch plasty; Homograft: Aortic arch repair—Homograft arch augmentation; ETE: Coarctation repair—End-to-end anastomosis; EETE: Coarctation repair—Extended end-to-end anastomosis; SCF: Coarctation repair—Subclavian flap plasty; Conduit: Coarctation repair—Conduit interposition]	VSD closure
VSD closure	217	Procedure Patch: VSD closure—Patch; Suture: VSD closure—Suture; Device: VSD closure—Occlusion device Approach Transatrial: VSD closure—Transatrial; RV: VSD closure—RV approach; LV: VSD closure—LV approach; Aorta: VSD closure—Aortic approach; PA: VSD closure—MPA approach	Interventricular fenestration
Pulmonary venous obstruction correction	199		PVO dilation, Pulmonary venous atherectomy, Pulmonary venous stent implantation, Pulmonary venous obstruction correction—Sutureless
Aortic valve replacement	181	Types Bioprosthetic Mechanical Homograft Non-valved aortic graft conduit	Converted from valvuloplasty in the same operation Aortic annulus enlargement or augmentation
ASD closure	115	Procedure Suture; Patch; Device closure	Interatrial fenestration, ASD closure—ASD primum, ASD closure—ASD secundum; Repair or reconstruction of superior vena cava
Arrhythmia procedure	59	Ablation type Radiofrequency ablation Cryoablation Region Atrium Ventricle Bypass tract

Abbreviations: ASD, atrial septal defect; PA, pulmonary atresia; PVO, pulmonary venous obstruction; RA, right atrium; RV, right ventricle; VSD, ventricular septal defect.

Evaluation of reliability of the algorithm

Following development, phenotypes were computed for the entire cohort and the time required for computing was recorded. Thereafter, the computed native anatomies of the random 500 unseen patients were used to validate the algorithm. Five experienced pediatric cardiologists and pediatric cardiac surgeons were each provided with the list of studies, study reports (Fyler codes), and phenotype assigned by the algorithm of 100 patients (ie, they were not blinded to algorithm output), and asked to adjudicate whether the algorithm correctly described the patient. The agreement by the expert clinicians was calculated as a percent in agreement. The performance for identifying the presence of congenital heart disease was evaluated using cross-tabulation.

Interactive real-time outcome reporting

To demonstrate the utility of this construct, we created an interactive dashboard based on parent diagnosis in which subcategories, mandatory covariates, and interventional strategy can be dynamically selected and de-selected. Key variables of interest were ingested (eg, surgical era, birth weight, gestational age, relevant valve dimensions), allowing for dynamic comparisons based on patient risk factors. Furthermore, outcomes of interest were computed for each patient (eg, percent alive) and index hospitalization (eg, lengths of stay). These were manifest on a front-end display that allows for the dynamic comparison of risk factors, interventions, and outcomes, using Kaplan–Meier analysis. Clinical predictor elements (eg, weight, estimated gestational age, cardiac dimensions) and outcomes data (eg, survival) are maintained up-to-date by reading against live-updating tables in the electronic health record.

Results

Among 161 735 patients (48% male, age 6.7 [IQR, 0.72-16] years) cared for between February 1981 and March 2020, Fyler coding data were available for 514 541 echocardiograms, 43 332 operations, and 38 264 cardiac catheterizations. Of all echocardiograms, 34 550 (6.7%) were fetal echocardiograms, 23 099 (4.5%) had been performed in outside hospitals, and 24 720 (4.8%) were transesophageal echocardiograms. The compute time required to phenotype all patients was 25 minutes, and the mean was 9.4 milliseconds per patient. The computed phenotypes of the entire cohort and the validation cohort are shown in Table 3.

Table 3.

Number of patients included in each defining anatomy.

Defining anatomy identified by the algorithm	Total patients, N	%	% in patients with abnormal cardiovascular anatomy	Patients in validation cohort, N	N correct
Normal cardiac anatomy	77 540	47.9		222	221
Ventricular septal defect	13 267	8.2	15.8	47	44
Patent foramen ovale	10 558	6.5	12.5	44	38
Patent ductus arteriosus	8484	5.2	10.1	23	21
Atrial septal defect	7487	4.6	8.9	38	38
Mitral valve prolapse	3410	2.1	4.1	10	8
Pulmonary stenosis	2829	1.7	3.4	7	6
Common atrioventricular canal	2674	1.7	3.2	7	7
D-looped transposition of the great arteries	2584	1.6	3.1	6	6
Aortic stenosis, mild or severity unmentioned	2419	1.5	2.9	4	2
Pulmonary arterial branch stenosis	2137	1.3	2.5	10	10
Coarctation of the aorta, trivial, mild, or severity unmentioned	2060	1.3	2.4	5	5
Hypoplastic left heart syndrome	1919	1.2	2.3	1	1
Aortic commissure abnormality	1898	1.2	2.3	12	11
Double-outlet right ventricle	1744	1.1	2.1	3	3
Tetralogy of Fallot with pulmonary stenosis	1523	0.9	1.8	8	8
Tetralogy of Fallot	1491	0.9	1.8	8	8
Aortic stenosis, moderate to severe	1344	0.8	1.6	3	3
Tetralogy of Fallot with pulmonary atresia	1174	0.7	1.4	5	5
Other ventricular structural abnormality	946	0.6	1.1	2	2
Unknown cardiac anatomy	732	0.5	0.9	0
Total anomalous pulmonary venous connection	698	0.4	0.8	1	1
Tricuspid atresia	646	0.4	0.8	3	3
Coarctation of the aorta, moderate to severe	623	0.4	0.7	2	2
Pulmonary atresia with intact ventricular septum	585	0.4	0.7	4	4
Ebstein’s anomaly	562	0.3	0.7	0
L-looped transposition of the great arteries	562	0.3	0.7	3	2
Coronary arterial fistula	534	0.3	0.6	1	1
Truncus arteriosus	521	0.3	0.6	1	1
Partial anomalous pulmonary venous connection	507	0.3	0.6	1	1
Cleft mitral valve	482	0.3	0.6	2	2
Double-inlet left ventricle	476	0.3	0.6	1	1
Left superior vena cava	383	0.2	0.5	2	2
Other atrial abnormality	343	0.2	0.4	0
Other mitral valve abnormality	326	0.2	0.4	0
Other coronary artery abnormality	304	0.2	0.4	1	1
Interrupted aortic arch	295	0.2	0.4	1	1
Vascular ring	294	0.2	0.3	1	1
Mitral stenosis	277	0.2	0.3	0
Dilated aortic annulus	272	0.2	0.3	1	1
Tricuspid valve prolapse	261	0.2	0.3	0
Other abnormality of inferior vena cava	243	0.2	0.3	0
Aneurysm of septum primum	236	0.1	0.3	1	1
Aortic arch abnormal branching	231	0.1	0.3	0
Single ventricle	224	0.1	0.3	0
Tetralogy of Fallot with common atrioventricular canal	198	0.1	0.2	0
Pulmonary atresia	187	0.1	0.2	0
Pulmonary venous stenosis	186	0.1	0.2	1	1
Anomalous origin of right coronary artery from left	179	0.1	0.2	0
Hypoplastic left ventricle	159	0.1	0.2	0
Tetralogy of Fallot with absent pulmonary valve syndrome	155	0.1	0.2	0
Other pulmonary valve abnormality	151	0.1	0.2	0
Shone’s syndrome	139	0.1	0.2	1	1
Anomalous origin of the coronary artery from pulmonary artery	135	0.1	0.2	1	1
Hypoplastic aortic annulus	122	0.1	0.1	0
Right aortic arch	117	0.1	0.1	0
Aortopulmonary window	109	0.1	0.1	2	2
Hypoplastic aorta	103	0.1	0.1	0
Hypoplastic pulmonary arterial branch	101	0.1	0.1	1	1
Total	161 735			500	482

Phenotypes with fewer than 100 patients assigned are not shown.

In the validation, experts completely agreed with the native phenotype assignments in 482 out of 500 patients (96.4%). The assignment of “normal cardiac anatomy” was correct in 221 of 222 patients. Descriptions of the phenotypic discrepancies are shown in Table 3. In 6 patients, there was disagreement regarding the age at which a patent foramen ovale should be considered normal, and in another an atrial septal defect closed spontaneously and this was not captured in a Fyler code. In 6 other patients, findings noted only in an operative report but not any echocardiogram were not included in the phenotypic description. Six other patients who were assigned a phenotype of congenital heart disease (all were assigned patent foramen ovale [PFO]) were considered normal by the experts. The algorithm therefore exhibited a sensitivity of 99.6% (95% confidence interval, 98.0%-99.9%), specificity 97.4% (94.3%-98.7%), positive predictive value 97.8% (95.4%-99.0%), and negative predictive value 99.5% (97.5%-99.9%) for identifying the presence of congenital heart disease.

A prototypic interactive outcome reporting tool was developed (Figure 2). This tool allows users to select anatomic subtypes and important co-variates of interest and interventional approaches that have been used, dynamically displaying outcomes of interest. A video of the tool is provided as Video S1.

Figure 2.

Interactive outcome reporting system. A 4-quadrant dashboard can be used to explore the data within a single parent diagnosis, in this case hypoplastic left heart syndrome (HLHS). The upper left quadrant lists the subcategories (eg, mitral stenosis and mitral atresia) and mandatory covariates (eg, intact atrial septum, IAS), including the number of patients with each variant. The upper right displays relevant clinical variables that can be customized to each parent diagnosis, such as estimated gestational age (EGA), birth weight (BW), birth year, procedure year, aortic valve (AV) and mitral valve (MV) dimensions, and degree of tricuspid regurgitation. The bottom left quadrant displays the potential initial interventions—in this case, a stage 1 palliation including a Damus-Kaye-Stansel with Sano shunt (DKS/Sano) or modified Blalock-Taussig-Thomas Shunt (BTTS), a hybrid procedure, or a biventricular (BiV) repair. These interventional choices define the color scheme of both the clinical variables (upper right) and outcomes plots (lower left). A demonstration of an example interface is provided in Video S1.

Discussion

We have demonstrated the feasibility of computing the primary diagnosis and procedures of patients with CHD based on the whole of their diagnostic and procedural codes with a high degree of clinical accuracy and computational efficiency. Our approach leveraged the inherent hierarchy of the Fyler, IPCCC and ICD-11 coding systems with the intent to align with clinician cognition and communication in which patients are described with sequential terms in order of clinical import. For example, a patient with hypoplastic left heart syndrome (HLHS), mitral atresia, aortic atresia, and PFO is first described as having HLHS; one would never characterize the patient’s primary disease as a PFO, for example, as it is lower on the hierarchy. We accomplished this computation with 96% accuracy due to several important features. First, each native code was mapped to one or more “algorithm” codes, which allowed us to harmonize differences in approach to coding over the decades and to losslessly extract information in circumstances in which multiple codes were combined into one. This approach is optimal for efficient code growth and for mapping of other existing coding systems into this infrastructure. Also, we note that this infrastructure is immediately portable to ICD-11, which has adopted the Fyler coding infrastructure for the coding of congenital heart disease. Second, codes were iteratively processed, first at the event level (eg, each echocardiogram was independently analyzed for a parent diagnosis) and then at a patient level. This approach accounts for disagreements between interpretations of different echocardiograms (eg, DORV vs tetralogy of Fallot), and the fact that parent diagnoses are often not recorded on all studies, which may be performed for focused evaluations. Among other advantages, this infrastructure creates the flexibility to define stages in a single patient; for example, a description of the circulation may be defined prior to and following a procedure. Third, any parent diagnosis can be modified with any number of subcategories or mandatory covariates, which both simplifies the requirements placed on the parent diagnosis and allows for nearly infinite specificity regarding anatomy, since each covariate that is also a parent diagnosis also inherits its details. In addition, the algorithm’s short computational of 9.4 milliseconds per patients allows its real-time implementation in the clinical setting.

We envision several potential use cases and opportunities for growth as this approach matures. (1) As we demonstrate here, efficient phenotype computation allows for the creation of a real-time institutional, disease-specific interactive dashboard. Interactive dashboarding allows for global assessments of outcomes over time, answering questions such as “how many patients with HLHS have we treated in the past decade and what are their outcomes?” As depicted here, though, it also allows provider selection of known risk factors that might be relevant to a decision at hand (eg, a 2.5 kg newborn with HLHS and severe TR), dynamically identifying the number of patients treated with a specific set of risk factors, treatment choices, and outcomes. In the future, such dashboards could be multiplexed across institutions (adding institution as a variable), enabling visualization of not only which procedures might portend the best outcomes in a specified situation but which centers. (2) The scalability of this approach hinges upon the use of algorithm codes as pass-through entities to which any number of coding systems can be mapped. While we initially mapped Fyler codes to algorithm codes, the application of this approach at other sites is likely to require that other nomenclatures, including ICD-11, SNOMED-CT, and Intelligent Medical Objects (IMO) be mapped to algorithm codes. Given that many EMR-based diagnoses originate from IMO-derived vocabularies, we recognize the importance of building and maintaining a dictionary that enables interoperability between EMR problem lists and algorithm-derived phenotypes. Work is ongoing to develop such mappings, and the modular design of the algorithm readily accommodates additional dictionaries without altering its core logic. In the near future, it is also likely that large language models (LLMs) will be leveraged for such mappings. The active infrastructure provided here will likely greatly improve the accuracy of LLM inferences by providing a structured target. Future iterations of our platform will incorporate logic to flag discrepancies, prioritize sources based on reliability and recency, and allow clinician adjudication when needed. (3) The accurate and detailed assignment of anatomic diagnosis, procedural details, and comorbidities are integral to outcomes reporting. Current constructs, including Society of Thoracic Surgeons Congenital Heart Surgery (STS-CHS) Database, utilize broad categories of anatomic diagnosis and procedures. It has been shown that the incorporation of diagnosis-procedure combination categories, procedure-specific risk factors, and syndromes/abnormalities meaningfully alters hospital performance metrics.^14,15 Given the importance of accurate outcome reporting, the algorithm described here could be used to compute diagnoses and procedures from raw data elements according to a standardized set of criteria and without loss of detail. Outcomes networks such as STS could integrate such an algorithm directly into source health record systems, enabling computation of each patient’s phenotype based on the codes identified in their primary data elements.

Limitations

Several limitations should be acknowledged. First, this effort took place at a single institution with a mature Fyler coding system. The performance of this approach at other sites with less structured or even unstructured anatomic information is unknown. Second, the algorithm was designed to be tolerant of nuances in the coding system used to train it, some of which represented peculiarities of our institutional coding practices. The translation of this technology to other sites is likely to require a similar fine-tuning and validation process. Third, the scope of our effort was intentionally limited to characterization of the phenotype at birth and the initial procedure. The ability of the algorithm to adjudicate phenotype at multiple stages of palliation is unclear, though conceptually no limitations are evident.

Conclusions

The computation of detailed CHD phenotypes from raw diagnostic and procedure codes is possible with a high degree of accuracy and efficiency. The algorithm designed to integrate with other coding systems, such as ICD-11, and with large language models to extract anatomic features. This framework may enable tools to support interactive outcomes reporting and clinical decision support.

Author contributions

Shuhei Toba (Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration), Taylor M. Smith (Formal analysis, Investigation, Methodology), Francesca Sperotto (Conceptualization, Data curation, Formal analysis, Investigation, Methodology), Chrystalle Katte Carreon (Formal analysis, Investigation), Kwannapas Saengsin (Data curation, Investigation, Methodology), Samuel Casella (Data curation, Formal analysis, Investigation, Methodology), Marlon Delgado (Data curation, Investigation, Methodology), Peng Zeng (Data curation, Formal analysis, Investigation), Stephen P. Sanders (Formal analysis, Investigation, Methodology), Audrey Dionne (Formal analysis, Investigation, Methodology), Eric Feins (Data curation, Formal analysis, Investigation), Steven D. Colan (Data curation, Formal analysis, Investigation, Methodology), John E. Mayer (Data curation, Formal analysis, Investigation), and John Kheir (Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration)

Supplementary material

Supplementary material is available at JAMIA Open online.

Funding

This work was funded in part by grants from MilliporeSigma, the Georgia Claire Bowen Foundation, the Joyful Heart Fund, and Japan Society for Promotion of Science, Grants-in-Aid for Scientific Research (20KK0375).

Conflicts of interest

The authors have no financial conflict of interest that relate to the study. There are no relationships with industry.

Data availability

The clinical data underlying this article cannot be shared publicly due to institutional and patient privacy regulations. The diagnostic and procedural code data were extracted from the electronic health record system at Boston Children’s Hospital and contain protected health information that is not eligible for public release.

The algorithms, code hierarchies, and phenotype assignment logic developed in this study are generally included in the manuscript as supplementary tables. Access to the GitHub repository and any derived, de-identified materials will be provided upon reasonable request to the corresponding authors on a case-by-case basis following approval by the Boston Children’s Hospital Institutional Review Board and completion of a data use agreement.