BackgroundA subset of children with congenital heart disease (CHD) travel internationally for definitive cardiac surgery, yet outcome data for selectively referred international patients undergoing complex congenital cardiac procedures remain limited. We describe results for international pediatric cardiac patients operated on within a structured referral model at a US tertiary congenital cardiac center.MethodsFrom September 2020 to July 2025, 1465 patients underwent cardiac surgery at our institution. Sixty-six consecutive international patients underwent 117 procedures, coordinated through Global Services. Demographics, diagnoses, operative variables, and outcomes were analyzed.ResultsMedian age was 53 months (interquartile range [IQR] 12-140) and weight 15.6 kg (IQR 8.5-39.5); 62% (41/66) were male. Referring countries were the United Arab Emirates (37/66, 56%), Kuwait (35%, 23/66), and Qatar (9%, 6/66). Prior cardiac surgery occurred in 74% (49/66), and 58% (38/66) had ≥2 sternotomies; 25% (17/66) had single-ventricle physiology. Society of Thoracic Surgeons - European Association for Cardio-Thoracic Surgery Congenital Heart Surgery category 3 to 5 procedures accounted for 48% (32/66) of the cases. Median cardiopulmonary bypass time was 174 min (IQR 120-255) and cross-clamp time 83 min (IQR 33-137). There was no hospital mortality. Early reoperation occurred in 7.5% (5/66) (3% for bleeding). Median intensive care unit and hospital stay were 5 (IQR 3-12) and 11 days (IQR 7-30). Major complications included extracorporeal membrane oxygenation, stroke, and dialysis in 1 patient each (1.5% each) and vocal cord injury in 6/66 (9%). Over a mean follow up of 312 days, 2 late deaths occurred (6 months and 1 year).ConclusionsInternational patients with CHD can achieve excellent early outcomes despite high complexity when managed at experienced centers with structured referral coordination and multidisciplinary infrastructure.

Background:Digoxin use after the Norwood procedure has been associated with improved interstage survival in hypoplastic left heart syndrome and related conditions. Whether this benefit translates into improved longer-term outcomes through staged palliation remains unknown. We aimed to determine the association of digoxin use at Norwood discharge with transplant-free survival and Fontan completion. Methods:We conducted a retrospective cohort study using the Pediatric Heart Network (PHN) Single Ventricle Reconstruction trial public dataset, including 549 infants enrolled at 15 North American centers between 2005 and 2008. Competing risk analysis was used to evaluate Fontan completion and Cox regression to assess death or transplantation within 6 years after the Norwood procedure. Mixed-effects models compared pre-Fontan hemodynamic and echocardiographic right ventricular indices between patients treated with and without digoxin after accounting for center clustering and adjustment for sex, shunt type, heart failure medications at Norwood discharge, and census block poverty level. Results:The 6-year cumulative incidence of Fontan completion was higher among patients discharged on digoxin than among those not receiving digoxin (82% vs 71%; p = 0.013). Competing-risk analysis accounting for death and transplant demonstrated a greater likelihood of Fontan completion among digoxin users (aHR 1.31; 95%CI 1.09-1.58; p = 0.005), without significant difference in the hazard of death or transplant (aHR 0.78; 95%CI 0.53-1.15; p = 0.208). No significant differences in pre-Fontan hemodynamic or echocardiographic indices were observed between groups. Initiation of digoxin post Stage II procedure was not associated with improved survival or likelihood to complete Fontan. Conclusion:Digoxin use at the time of Norwood discharge was associated with a 30% greater likelihood of Fontan completion by 6 years, without accompanying improvement in transplant-free survival. These findings extend prior observations of improved interstage outcomes associated with digoxin use and suggest that treatment may facilitate progression through staged palliation.

OBJECTIVE:To examine the accuracy of pulse oximetry in neonates by race and ethnicity. STUDY DESIGN:We performed a single-centre retrospective cross-sectional study in neonates aged ≤ 7 days at Children's Healthcare of Atlanta in the neonatal and cardiac ICU between 2010 and 2022. Eligible patients had pulse oximetry with arterial oxygen saturation measurements taken within 10 minutes. We evaluated the mean saturation bias, accuracy root mean square, and the frequency of occult hypoxaemia by race and ethnicity. RESULTS:Of 1645 infants, there were 639 non-Hispanic Black infants (38.8%), 651 non-Hispanic White infants (39.6%), and 215 Hispanic infants (13.1%). The majority had no CHD (71.4%), required respiratory support (81.1%), and had oxygen saturation levels = > 90% (87.6%). The mean pulse oximetry and arterial oxygen measurement time difference was 0.8 minutes. The mean bias and accuracy root mean square were 8.7% and 13.7%, respectively, with no significant difference between the groups (p = 0.91). Occult hypoxaemia was found in 20.5% of infants. Compared to White infants, there were no differences in likelihood of occult hypoxaemia for either Black (adjusted prevalence ratio 1.13 (0.92-1.4), p = 0.24) or Hispanic (1.06 (0.77-1.45, p = 0.72) infants. CONCLUSION:There was not a systemic discrepancy in pulse oximetry between racial and ethnic groups as previously described in older children and adults. However, pulse oximetry significantly overestimated arterial oxygen saturation. Future prospective studies that objectively measure skin pigmentation may be able to overcome some of the limitations of our study.

Extracorporeal life support (ECLS) is required in 8-24% of neonates following the Norwood operation and is associated with high morbidity and mortality. Early identification of patients at risk may enable timely intervention and resource optimization. We developed and internally validated a bedside risk score at arrival to the cardiac intensive care unit (CICU) to predict ECLS requirement in the first 7 days postoperatively. In a retrospective cohort of 322 neonates undergoing the Norwood procedure between January 2010 and December 2023, 71 (22%) required postoperative ECLS. Multivariable logistic regression identified six independent predictors: moderate-severe atrioventricular valve regurgitation (5 points), modified Blalock-Taussig-Thomas (m-BTT) shunt (5 points), cardiopulmonary bypass time greater than 173.5 minutes (5 points), vasoactive inotropic score greater than 19.5 on CICU arrival (7 points), delayed sternal closure (5 points), and postoperative inhaled nitric oxide use (7 points). The resulting score (range 0-29) stratified patients into low- (0-10), moderate- (12-19), and high-risk (20-29) groups, with corresponding ECLS rates of 8%, 31%, and 65%. The model demonstrated good discrimination (area under the receiver operating characteristic curve [AUC]: 0.78). This novel risk tool may support early clinical decision-making and triage in high-risk neonates following Norwood palliation. External validation is warranted to assess generalizability and clinical utility.

Background: The interstage period is a critical phase for single ventricle infants due to their fragile cardiovascular state. Infants often experience medical and feeding challenges during this period, resulting in caregiver stress. We completed a quality improvement project at Children’s Healthcare of Atlanta to understand these challenges to inform targeted interventions. Methods: This single-center project included a medical chart review and a cross-sectional caregiver survey. Data were collected on patient and caregiver demographics and clinical variables. Feeding outcomes were assessed using the Pediatric Functional Oral Intake Scale. Caregiver impact was measured using the Feeding/Swallowing Impact Survey. Results: The project included 15 single ventricle patients with a mean (standard deviation) age of 151.73(25.92) days at the time of the second-stage palliation. Forty percent of patients experienced at least one readmission, primarily due to feeding intolerance (20%) and desaturations (26.7%). Milk protein allergy (26.9%) was the most common medical complication, followed by interstage unplanned reinterventions. Pediatric Functional Oral Intake Scale scores demonstrated that 33% consumed minimal volumes or no oral intake at the time of the bidirectional Glenn, and 93.3% of patients did not receive outpatient feeding services during the interstage. Caregiver stress scores resulted in mean scores (standard deviation) of 2.23(1.54), with the highest impact on daily activities. All caregivers affirmed the need for a dedicated multidisciplinary clinic. Conclusion: The interstage period for single ventricle patients poses significant medical and feeding challenges, resulting in caregiver stress. Comprehensive, multidisciplinary feeding support during the interstage period may improve patient outcomes and alleviate caregiver burden.

Patients with single ventricle heart disease (SVHD) require multiple palliative surgeries. Due to high mortality rates observed during the four to sixth-month interstage period between the first and second surgery, at-home monitoring practices are important. However, modern-day interstage monitoring does not offer comprehensive quantitative tracking of health conditions since it relies on logging the infant's weight, feeding patterns, and oxygen saturation once a day in a journal. Here, we introduce a soft wearable electronic device for the home monitoring of SVHD patients. The ultrathin, imperceptible, wireless system is designed for safe integration with infants' delicate skin. The system features a chest-mounted biopatch and forehead-mounted pulse oximeter, streaming real-time data to a cloud-based platform. This smart device has built-in alert functions to help caregivers recognize when the patients should be returned to the hospital for in-person observation, easing decision-making from non-clinically trained personnel. In addition, the patient-tailored database also improves the communication between the home and hospital, reducing the load on caretakers during this stressful period.

Background: The Norwood operation (NO) for infants with univentricular physiology has high interstage mortality. This study evaluated outcomes and risk factors for mortality following NO. Methods: Retrospective single-center study of patients undergoing NO from 2010 to 2020. Analysis used appropriate statistics. Results: Of 269 patients undergoing NO, 213 (79.2 %) survived to discharge. Non-survivors had longer bypass times, delayed sternal closure, required nitric oxide, higher vasoactive scores, required post-operative catheterization, Extracorporeal Life Support (ECLS), and longer ventilation (p < 0.05). Logistic regression showed moderate-severe atrioventricular valve regurgitation on intraoperative TEE (OR 2.6), requiring nitric oxide (OR 2.63), delayed sternal closure (OR 2.94), post-operative catheterization (OR 10.48), and ECLS (OR 14.54) increased mortality odds (p < 0.05). Multivariable analysis confirmed catheterization (aOR 10.48) and ECLS (aOR 14.54) as significant predictors. Of survivors, 26 (12.3 %) developed new morbidity, 9 (4.2 %) had unfavorable outcomes. Functional status improved from 6.0 to 8.04, mainly in feeding and respiratory domains (p < 0.0001). Conclusions: Norwood survival was 79.2 %. Requiring post-operative catheterization and ECLS significantly increased mortality risk. Multicenter evaluation of these modifiable risk factors is needed to improve outcomes in this high-risk population.

Background: Patients requiring extracorporeal life support (ECLS) support post-Norwood operation constitute an extremely high-risk group. Materials and methods: We retrospectively aimed to evaluate the relationship of hyperoxia with mortality and other clinical outcomes in patients who required ECLS following Norwood operation between January/2010 and December/2020 in a large volume center. Results: During the study period 65 patients required ECLS post-Norwood. Using receiver operating characteristic (ROC) curve analysis, mean PaO2 of 182 mmHg in the first 48-hour on ECLS was determined to have the optimal discriminatory ability for mortality (sensitivity 68%, specificity 70%). Of the 65 patients, 52% had PaO2 > 182 mmHg and were designated as hyperoxia group. Patients in the hyperoxia-group had longer cardiopulmonary bypass time (187 vs. 165 min, p = 0.023), shorter duration from CICU arrival to ECLS-cannulation (13.28 vs. 132.58 h, p = 0.003), higher serum lactate within 2-hours from ECLS-canulation (14.55 vs. 5.80, p = 0.01), higher ECLS flows in the first 4-hours (152.68 vs. 124.14, p = 0.006), and higher mortality (77% vs. 39%, p = 0.005). In the unadjusted-analysis, using a derived cut-point, patients in the hyperoxia-group had 5.15 higher odds of mortality (p = 0.003). However, this association was insignificant when adjusting for confounding variables (p = 0.104). Using a functional status scale, new morbidity (38% vs. 21%), and unfavorable outcomes (13% vs. 5%) were higher in the hyperoxia group. Despite being higher in the hyperoxia group, this did not reach statistical significance. Conclusion: Neonates with hyperoxia (PaO2 > 182 Torr) during the first 48-hour of ECLS post-Norwood operation had 5 times higher odds of mortality in the unadjusted analysis, however, this was insignificant when adjusting for confounding variables. Patients in the hyperoxia group had shorter duration from CICU arrival to ECLS-cannulation, higher serum lactate prior to ECLS-canulation, and higher ECLS flows in the first 4-hours, (p < 0.05). Multicenter evaluation of this modifiable risk factor is imperative to improve the care of this high-risk cohort.

Congenital heart defects (CHDs) are birth defects that change the heart's structure and blood flow. Patients with these conditions may require an artificially placed shunt to allow adequate circulation through the heart and lungs. Evaluating shunt function is paramount to sustaining the functionality and life of the shunt. Effective monitoring of the shunt requires noninvasive approaches to enable frequent monitoring of the shunt's condition. In this work, we propose deep learning architectures for the classification of heart sounds associated with blood flow through shunts, obtained using a digital stethoscope from infants with ductal-dependent physiology. Specifically, we propose a convolutional neural network-long short-term memory model for shunt-type classification. In addition, a variational autoencoder model utilizing latent space representation addresses five clinically relevant tasks: 1) determining flow status during extracorporeal membrane oxygenation (ECMO); 2) discriminating flow patterns before and after shunt or pulmonary artery angioplasty; 3) identifying cyanotic conditions versus noncyanotic states; 4) change of flow dynamics over time; and 5) assessing flow status under elevated pulmonary artery pressure. Experimental results showcase the effectiveness of our approach, with the shunt-type classification achieving an F1 score of 0.88 and an area under the ROC curve (AUC) of 0.95. For the remaining tasks (1)-(5), support vector machine, random forest, and k-nearest neighbors models are trained using the latent space representations, yielding AUC values of 0.77, 0.72, 0.85, 0.60, and 0.81, respectively. This development demonstrates that the acoustic signals captured via a stethoscope may contain valuable information for monitoring shunt flow changes in patients with CHD.

This cross-sectional study investigates perioperative oxygen saturation differences in Black and White infants with single ventricles undergoing stage 1 palliation.

Patients requiring extracorporeal life support (ECLS) post-Norwood operation constitute an extremely high-risk group. We retrospectively described short-term outcomes, functional status, and assessed risk factors for requiring ECLS post-Norwood operation between January 2010 and December 2020 in a high-volume center. During the study period, 269 patients underwent a Norwood procedure of which 65 (24%) required ECLS. Of the 65 patients, 27 (41.5%) survived to hospital discharge. Mean functional status scale (FSS) score at discharge increased from 6.0 on admission to 8.48 (p < 0.0001). This change was primary in feeding (p < 0.0001) and respiratory domains (p = 0.017). Seven survivors (26%) developed new morbidity, and two (7%) developed unfavorable functional outcomes. In the regression analysis, we showed that patients with moderate—severe univentricular dysfunction on pre-Norwood transthoracic echocardiogram (odds ratio [OR] = 6.97), modified Blalock Taussig Thomas (m-BTT) shunt as source of pulmonary blood flow (OR = 2.65), moderate—severe atrioventricular valve regurgitation on transesophageal echocardiogram (OR = 8.50), longer cardiopulmonary bypass time (OR = 1.16), longer circulatory arrest time (OR = 1.20), and delayed sternal closure (OR = 3.86), had higher odds of requiring ECLS (p < 0.05). Careful identification of these risk factors is imperative to improve the care of this high-risk cohort and improve overall outcomes.

Background: There is a paucity of information reported regarding the use of milrinone in patients with hypoplastic left heart syndrome prior to the Norwood procedure. At our institution, milrinone is initiated in the pre-operative setting when over-circulation and elevated serum lactate levels develop. We aimed to review the responses associated with the administration of milrinone in the pre-operative hypoplastic left heart syndrome patient. Second, we compared patients who received high- versus low-dose milrinone prior to Norwood procedure. Methods: Single-centre retrospective study of patients diagnosed with hypoplastic left heart syndrome between January 2000 and December 2019 who underwent Norwood procedure. Patient characteristics and outcomes were compared. Results: During the study period, 375 patients were identified; 79 (21%) received milrinone prior to the Norwood procedure with median lactate 2.55 mmol/l, and SpO2 93%. Patients who received milrinone were older at the time of Norwood procedure (6 vs. 5 days) and were more likely to be intubated and sedated. In a subset analysis stratifying patients to low- versus high-dose milrinone, median lactate decreased from time of initiation (2.39 vs 2.75 to 1.6 vs 1.8 mmol/l) at 12 hours post-initiation, respectively. Repeated measures analysis showed a significant decrease in lactate levels by 4 hours following initiation of milrinone, that persisted over time, with no significant difference in mean arterial pressure. Conclusions: The use of milrinone in the pre-operative over-circulated hypoplastic left heart syndrome patient is well tolerated, is associated with decreased lactate levels, and was not associated with significant hypotension or worsening of excess pulmonary blood flow.

OBJECTIVES:Provision of pulmonary blood flow with a systemic-to-pulmonary artery shunt is essential in some patients with cyanotic congenital heart disease. Traditionally, aspirin (ASA) has been used to prevent thrombosis. We evaluated ASA dosing with 2 separate antiplatelet monitoring tests for accuracy and reliability. METHODS:This is a retrospective, pre-post intervention single center study. Two cohorts were evaluated; the pre-intervention group used thromboelastography platelet mapping (TPM) and post-intervention used VerifyNow aspirin reactivity unit (ARU) monitoring. The primary endpoint was to compare therapeutic effect of TPM and ARU with regard to platelet inhibition. Inadequate platelet inhibition was defined as TPM <50% inhibition and ARU >550. RESULTS:Data from 49 patients were analyzed: 25 in the TPM group and 24 in the ARU group. Baseline characteristics were similar amongst the cohorts. The TPM group had significantly more patients with inadequate platelet inhibition (14 [56%] vs 2 [8%]; p = 0.0006) and required escalation with additional thromboprophylaxis (15 [60%] vs 5 [21%]). There was no difference in shunt thrombosis (1 [2%] vs 0 [0%]; p = 0.32), cyanosis requiring early re-intervention (9 [36%] vs 14 [58%]; p = 0.11), or bleeding (15 [60%] vs 14 [58%]; p = 0.66). CONCLUSION:With similar cohorts and the same ASA-dosing nomogram, ARU monitoring resulted in a reduced need for escalation of care and concomitant thromboprophylaxis with no difference in adverse outcomes. Our study suggests ARU monitoring compared with TPM may be a more reliable therapeutic platelet inhibition test for determining ASA sensitivity in children with congenital heart disease requiring systemic-to-pulmonary artery shunt.

OBJECTIVES: There is an ongoing need for a method of obtaining long-term venous access in critically ill pediatric patients that can be completed at the bedside and results in a durable, highly functional device. We designed a novel technique for tunneled femoral access to address this need. Herein, we describe the procedure and review the outcomes at our institution.DESIGN: A single-center retrospective chart review identifying patients who underwent tunneled femoral central venous catheter (tfCVC) placement between 2017 and 2021 using a two-puncture technique developed by our team.SETTING: Academic, Quaternary Children's Hospital with a dedicated pediatric cardiac ICU (CICU).PATIENTS: Patients in our pediatric CICU who underwent this procedure.INTERVENTIONS: Tunneled femoral central line placement.MEASUREMENTS AND MAIN RESULTS: One hundred eighty-two encounters were identified in 161 patients. The median age and weight at the time of catheter placement was 22 days and 3.2 kg. The median duration of the line was 22 days. The central line-associated bloodstream infection (CLABSI) rate was 0.75 per 1,000-line days. The prevalence rate of thrombi necessitating pharmacologic treatment was 2.0 thrombi per 1,000-line days. There was no significant difference in CLABSI rate per 1,000-line days between the tfCVC and nontunneled peripherally inserted central catheters placed over the same period in a similar population (-0.40 [95% CI, -1.61 to 0.82; p = 0.52]) and no difference in thrombus rates per 1,000-line days (1.37 [95% CI, -0.15 to 2.89; p = 0.081]).CONCLUSIONS: tfCVCs can be placed by the intensivist team using a two-puncture technique at the bedside with a high-rate of procedural success and low rate of complications. Advantages of this novel technique of obtaining vascular access include a low rate of CLABSIs, the ability to place it at the bedside, and preservation of the upper extremity vasculature.

Background: Infants with congenital heart disease who have insufficient systemic or pulmonary blood flow require a connection between the systemic and pulmonary vasculature. There are four types of shunts: right ventricle to pulmonary artery (Sano), Blalock-TaussigThomas (BTTS), central, and ductal arterial stent (DAS). They have a high incidence of mortality and unplanned interventions. Machine learning algorithms evaluate a shunt murmur recorded on a digital stethoscope to differentiate the audio features of various shunts. Methods: A 3M Littman Electronic Stethoscope acquired phonocardiogram (PCG) data from infants with shunt placement on post-operative days 0, 1, 2, 7, 14, discharge and day of shunt take-down, as well as, when imaging (echo, CT, catheterization) was performed. Patient demographics and clinical outcomes noted. A wavelet decomposition algorithm denoised the PCG data with a bandpass filter ranging from 20 to 2000 Hz and then segmented into S1 and S2. One hundred and fifteen audio features, comprising temporal, spectral, statistical, and bandpower domains were extracted from the waveforms. A statistical algorithm including ANOVA, Kruskal Wallis and pairwise comparison tests detected 33 total features showing a statistically significant difference in means (alpha = 0.05). Results: Thirteen subjects (BTTS (n=3), DAS (n=2), and Sano (n=8)) who had second stage palliation were analyzed using the two most recent recordings prior to shunt takedown. Principal component analysis (Figure 1) and tdistributed stochastic neighbor embedding (Figure 2) visualization techniques indicate a separation among shunt types. Conclusion: Audio signatures could exist based on shunt type characteristics and may lead to a novel non-invasive tool to monitor shunt health. Figure 1. Principal Component Analysis (PCA) indicating a separation in BTTS, Sano, and PDA shunt types from selected recordings. Figure 2. t-distributed stochastic neighbor embedding (tSNE) indicating a separation in BTTS, Sano, and PDA shunt types from selected recordings.

AbstractObjective:A standardised multi-site approach to manage paediatric post-operative chylothorax does not exist and leads to unnecessary practice variation. The Chylothorax Work Group utilised the Pediatric Critical Care Consortium infrastructure to address this gap.Methods:Over 60 multi-disciplinary providers representing 22 centres convened virtually as a quality initiative to develop an algorithm to manage paediatric post-operative chylothorax. Agreement was objectively quantified for each recommendation in the algorithm by utilising an anonymous survey. “Consensus” was defined as ≥ 80% of responses as “agree” or “strongly agree” to a recommendation. In order to determine if the algorithm recommendations would be correctly interpreted in the clinical environment, we developed ex vivo simulations and surveyed patients who developed the algorithm and patients who did not.Results:The algorithm is intended for all children (<18 years of age) within 30 days of cardiac surgery. It contains rationale for 11 central chylothorax management recommendations; diagnostic criteria and evaluation, trial of fat-modified diet, stratification by volume of daily output, timing of first-line medical therapy for “low” and “high” volume patients, and timing and duration of fat-modified diet. All recommendations achieved “consensus” (agreement >80%) by the workgroup (range 81–100%). Ex vivo simulations demonstrated good understanding by developers (range 94–100%) and non-developers (73%–100%).Conclusions:The quality improvement effort represents the first multi-site algorithm for the management of paediatric post-operative chylothorax. The algorithm includes transparent and objective measures of agreement and understanding. Agreement to the algorithm recommendations was >80%, and overall understanding was 94%.

Background: Coarctation of the aorta (COA) is one of the most common congenital heart lesions. Most patients undergoing repair extubate without complication. A subset of patients are unable to extubate within 24 h postoperatively. Objective: We examined risk factors determined a priori and their association with delays in extubation. Methods: This is a retrospective single-center study at an academic quaternary children's hospital. All neonates (0-28 days) who underwent isolated coarctation repair via lateral thoracotomy between January 2008 - December 2019 were included. Delay in extubation was defined as duration from cardiac intensive care unit admission to time of extubation >24 h. Risk factors examined were total preoperative prostaglandin dose and total perioperative opiate dose following repair. Logistic regression analyses were performed using appropriate statistics with significance level set at p = 0.05. Results: There were 149 patients who met inclusion criteria, 102 (68.5 %) were intubated for >24 h. We found that patients who were smaller (3.2 kg vs. 3.4 kg, p = 0.012) and patients who were in the intensive care unit preoperatively (p = 0.048), were more likely to remain intubated >24 h. There was also a statistically significant difference between those that were extubated before 24 h postoperatively and those that were not with respect to the total amount of oral morphine milligram equivalent received (1.5 mg vs. 2.82 mg, p = 0.005). Conclusions: Of the 149 patients in our study, the majority (102 patients, 68.5 %) were unable to be extubated within 24 h of surgery following coarctation repair via lateral thoracotomy. Risk factors for delayed extubation were smaller neonates and patients who received greater amounts of opiates.

Background: Postoperative management of the neonate following the Norwood operation is among the most complex and challenging in pediatric critical care. Artificial intelligence is poised to assist in management of this complex population. Methods: We developed a convolutional neural network (CNN) model trained on electrocardiogram (ECG) waveforms collected from 45 neonates after the Norwood procedure. Waveforms from the first two postoperative days (critical) and the day prior to transfer out of the ICU (stable) were used for training. The model was validated on 10 post-Norwood neonates. Models were compared to traditional machine learning algorithms on non-waveform data, heart rate variability, and then combined in a final model to optimize performance. Retrospective clinical observation scoring was completed for comparison. Results: The CNN model on 3-lead ECG yielded an area under the curve of the receiver operating characteristic (AUC-ROC) of 0.97 ( 0.02). The final model combining predictions from the CNN, random forest on vital signs, and logistic regression on pH, lactic acid, and base deficit values achieved 0.98 ( 0.02). Retrospective clinical observations agreed with the final model 78% of the time. Conclusions: Application of this novel combined machine learning model can accurately detect changes in clinical status as patients progress from critically ill to stable following the Norwood procedure, yielding impressive clinical insights from the ECG beyond what is currently possible. This work provides the basis for the development of a novel bedside monitoring tool and suggests new ways artificial intelligence may influence clinical care beyond predicting deterioration events.

Deep learning models have achieved expert-level performance in healthcare with an exclusive focus on training accurate models. However, in many clinical environments such as intensive care unit (ICU), real-time model serving is equally if not more important than accuracy, because in ICU patient care is simultaneously more urgent and more expensive. Clinical decisions and their timeliness, therefore, directly affect both the patient outcome and the cost of care. To make timely decisions, we argue the underlying serving system must be latency-aware. To compound the challenge, health analytic applications often require a combination of models instead of a single model, to better specialize individual models for different targets, multi-modal data, different prediction windows, and potentially personalized predictions. To address these challenges, we propose HOLMES-an online model ensemble serving framework for healthcare applications. HOLMES dynamically identifies the best performing set of models to ensemble for highest accuracy, while also satisfying sub-second latency constraints on end-to-end prediction. We demonstrate that HOLMES is able to navigate the accuracy/latency tradeoff efficiently, compose the ensemble, and serve the model ensemble pipeline, scaling to simultaneously streaming data from 100 patients, each producing waveform data at 250~Hz. HOLMES outperforms the conventional offline batch-processed inference for the same clinical task in terms of accuracy and latency (by order of magnitude). HOLMES is tested on risk prediction task on pediatric cardio ICU data with above 95% prediction accuracy and sub-second latency on 64-bed simulation.