Infants younger than 1 year with cCHD admitted perioperatively to the PICU of the University Medical Centre Utrecht between 2002 and 2018 were included based on the availability of time-synchronous data streams. We collected data from 5 vital parameters in a frequency of 1 measurement per second, namely SpO₂, regional cerebral saturation (rSO₂) in both hemispheres, invasive mean arterial blood pressure (IBP), respiratory rate (RR), and heart rate (HR), as well as current mechanical ventilation status. Patients were excluded if less than 12 hours of complete data were available or due to low birth weight (<2000 g). 

As fully anonymized data were used, the medical ethical review committee of the Wilhelmina’s Children Hospital waived informed consent (application number 22/822). In manuscript preparation, the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) checklist [22] was used (Multimedia Appendix 1).

RR was measured through thoracic movement as a result of electrocardiographic impedance derivation with the electrocardiographic leads from the Philips Intellivue MP70 monitor. Because infants, especially neonates, can have considerable fluctuations of RR within minutes, a trend movement was examined rather than absolute values: a 300-second moving average for RR was implemented preceding each time point t. Cerebral rSO₂ was measured with near-infrared spectroscopy with the Medtronic INVOS 5100 monitor using 2 pediatric cerebral sensors. If both probes recorded a value, their mean was used in model calculations. At our institution, end-tidal carbon dioxide (EtCO₂) is considered in all mechanically ventilated patients to monitor the efficacy of ventilation; therefore, EtCO₂ was extracted to determine the current mechanical ventilation status at each time point (ie, currently mechanically ventilated if EtCO₂>0 at time t). No imputation was performed to account for missing values in these parameters. To account for the underlying varying physiology of CHD, patients were stratified into 2 subsets based on average SpO₂ during admission (ie, <90% versus ≥90%), as measured with oximetry using the Philips Intellivue MP70 monitor (FAST technology with the Nelcor sensor). As SpO₂ is a parameter in the model and therefore directly influences predictive performance, we decided to use data-driven stratification of CHD in order to accurately represent the spectrum of underlying diseases throughout the stratified group regardless of clinical diagnosis.

To facilitate future clinical use, our model was developed using explainable methods (ie, through methods allowing clinicians to understand what features and assets contribute to the output) as opposed to the so-called “black box” models (eg, deep neural networks), where their methodological foundation and feature derivation remains beyond grasp to most clinicians. The model’s internal architecture consisted of 3 separate models, which were integrated to coordinate a classification response of either sensor dysfunction or stable or unstable patient status (Figure 1). Each of the 3 models relied on a specific analysis: sensorial dysfunction (submodel 1 in Figure 1), classification of normal and abnormal vital parameter combinations (submodel 2 in Figure 1), and detection and analysis of significant patient-specific baseline deviations (submodel 3 in Figure 1).

An analyzed time point t was deemed unstable if no sensorial dysfunction was detected, and either submodel 2 or 3—or both—classified the time point t to be unstable. The continuous data points were converted to episodes through a 5-minute moving time frame, where an episode was considered unstable when classified thus in at least 4 minutes (ie, ≥80%) out of any 5-minute time frame. If less than 4 (nonconsecutive) minutes of the episode (ie, <80%) were deemed unstable, the time frame was consequently classified as stable. To allow for baseline build-up (submodel 3), the first hour of admission was analyzed without triggering a classification response.

All models were built using RStudio (version 1.4; R Foundation for Statistical Computing). The packages used in construction, as well as the source code, can be found on the website of our research group [23]. 

To reduce the faulty classification of patient status due to sensor errors, dysfunctions of IBP, SpO₂, rSO₂, and RR were evaluated. HR sensor dysfunction was not included as no reliable distinction between, for example, cardiac arrest (HR=0 beats per minute) and sensor error, could be made. IBP and SpO₂ dysfunction was determined as a difference of >25 points on their respective scales compared to the previously measured value (at time point t–1). The lower and upper limits of the rSO₂ scale (≤15% and ≥95%, respectively) were noted as measurement error as these values are unlikely to be a valid measurement and rather emerge due to escaping sensor-light emission. An RR sensor malfunction was considered to be a rate below 5 breaths per minute. Upon detection, measurements in the minute preceding the first detection (at time point t–60 seconds) up to the minute proceeding (at time point t+60 seconds) the last detection (t) were considered unfit for adequate classification and consequently classified as sensor dysfunction. 

Combinations of parameters were analyzed and classified to either be stable or unstable. Each vector of the parameters (RR, HR, IBP, rSO₂, and SpO₂) was normalized and reduced to a single principal component using the Mahalanobis method [24], with respect to the stratified subset–specific (SpO₂<90% versus SpO₂≥90%) mean, variance, and correlation matrices (Multimedia Appendix 2). Vectors with a corresponding Mahalanobis distance greater than the 80th percentile were deemed unstable and discarded from the subset. The remaining vectors were divided into a random 80:20 train:test partition and used to train a one-class support vector machine (SVM). We used a square-exponential radial basis function kernel with a 5% soft margin (µ) to prevent overfitting. As the SVM was trained using, presumably, stable vectors of parameters, any nonresemblant vector was classified as unstable by the SVM. Additionally, singular parameters were considered unstable when exceeding static cutoff values determined by the consensus of pediatric intensivists (Figure 2).

Baseline variation analysis focused on the detection of abnormal parameter deviations in comparison to the patient’s unique baseline (Figure 3). A vector of parameters (HR, RR, IBP, rSO₂, and SpO₂) was reduced to a single principal component using the earlier introduced Mahalanobis distance [24] and increased by 20% at times of mechanical ventilation (ie, EtCO₂>0) due to the consequent iatrogenically diminished variation in parameters. The current Mahalanobis trend (Z), through a 300-second moving median preceding time point t, was subsequently compared to the patient’s unique baseline (B, median of all Mahalanobis distances preceding t). As Mahalanobis distance is calculated using normalized values, trend movement toward subset mean values (Z–B<0) was assumed to be related to clinical improvement, where a significant trend drifting ≥2 SDs from both the subset mean values, as well as the baseline, was deemed to result from instability (Z–B≥+2SD). SD was calculated after the removal of the upper 20th percentile of the baseline corrected Mahalanobis distance (ie, the SD in supposedly stable time points) with respect to chronologicity.

Novel, unseen data from the 5 parameters (HR, RR, IBP, SpO₂, and rSO₂), along with the model’s classification, were visualized in detail (Multimedia Appendix 3). Two experienced pediatric intensivists (JN and EK) reviewed these charts, each noting independently, being blinded from each other, whether they agreed with the model classification. Any difference in opinion was resolved by an independent third expert. The performance of the algorithm was consequently based on expert opinion, noting both time-percentual correctness as well as episodic performance. Episodes were counted with a maximum duration of 2 consecutive hours to prevent shifting results based on episode length. 

In total, 92 patients were initially identified with time-synchronized parameters in their data sets, of whom 14 (15%) were excluded (<12-hour data: n=11, 79%; birthweight<2000 g: n=3, 21%). The remaining 78 patients were stratified into 2 subgroups based on mean SpO₂ during admission: SpO₂<90% (n=26, 33%) and SpO₂≥90% (n=52, 67%). The group characteristics are shown in Table 1. A list of cardiac diagnoses and performed surgical interventions on included patients is provided in Multimedia Appendix 4.

A total of 209 hours of data from 10 patients across the SpO₂<90% group (n=5, t=98 hours) and SpO₂≥90% group (n=5, t=111 hours) were classified by our algorithm for performance analysis.

In this proof-of-concept study, we have developed and retrospectively evaluated an advanced data science algorithm for PICU patients with cCHD aimed at automated detection of clinical deterioration during their critical perioperative period. Through 2-fold analysis of vital parameters, both in relation to each other and in comparison to the patient’s unique baseline parameters, a tailored approach was demonstrated to monitor complex and hemodynamically challenging patients. Overall, our model accurately detected clinical stability and deterioration in, respectively, 88% and 81% of expert-confirmed episodes. Sensor dysfunction occurred 138 times, of which 94% were rightfully detected.

The population of patients with cCHD has been shown to be at substantial risk of deterioration in their perioperative period, as they are susceptible to a range of hemodynamic and respiratory events, especially in the postoperative period [4-7]. These disturbances in (cerebral) blood flow and oxygenation may eventually result in damage to internal organs, such as the gut and the brain [7,25]. Brain injury, for example, is observed in up to 60% of postoperative patients with cCHD and is known to cause severe neurodevelopmental impairment, significantly impacting quality of life [26,27]. Adequate detection of patient deterioration could facilitate timely intervention and may, eventually, prevent the onset of novel (brain) injury. However, adequate and timely detection of ongoing deterioration is becoming increasingly difficult through the ever-growing amount of complex and dynamically interpretable data inherent to the cCHD population, posing a 24/7 monitoring challenge to the medical team. Additionally, previous research has noted subtle variations in vital parameters to precede adverse events [7] as well as significant phenotype differences in cCHD related to an adverse outcome [4]. Through mixed-effects regression analysis, Nicoll et al [4] described independent associations between elevated HR (P=.003) and elevated systolic BP (P=.02) with novel brain injury in the first 72 hours after surgery. These physiological differences were most significant directly postoperatively and decreased with time, again highlighting the importance of adequate and intensive perioperative monitoring to identify patients at higher risk of deterioration. However, paying attention to these different physiological phenotypes and subtle parameter variations requires 24/7 vigilance from staff, greatly increasing their cognitive load. With algorithmic condensation of clinical data streams toward comprehensible information, the cognitive load on clinicians and nurses will likely be decreased, providing support to both patients and the medical team.

Overall, research classifying current patient status in CHD—rather than predicting a future adverse event—is very scarce. To the best of our knowledge, diagnostic AI models classifying current patient status in CHD and cCHD have yet to be published. A fair comparison of predictive versus diagnostic models in CHD is limited due to their different aims and setup; however, their methodological comparison is possible to some degree.

In 2013, Clifton et al [14] proposed an algorithm for adults in the emergency department through the use of an integrated monitoring system that combines high-frequency physiological data to predict upcoming escalation of care. Here, they have developed and tested several ML methods against an existing evidence-based early warning score. The different approaches to predicting escalation of care had mixed results, where the SVM had a high detection rate (>85%, time frame–dependent), yet, also, a high false positive rate (27%). If their algorithm were applied to, for example, the population of patients with cCHD, their inherent dynamic circulation would not be taken into account, most likely decreasing the detection rate.

In this study, it is argued that the 2-fold analysis of stability (ie, parameters in relation to each other and with different time points) is of significant value to the monitoring or predicting of outcomes in heterogeneous populations, such as pediatrics, using high-frequency physiological data. As such, future studies aiming to monitor, classify, or predict outcomes in the pediatric population are encouraged to evaluate the need for adjustment to their patients’ dynamic physiology and consider their model’s resilience to these dynamic conditions. However, it must also be acknowledged that robust statistical methods for transparent advanced data science models, such as those proposed in this study, remain scarce to this date, especially in complex clinical time-series data. 

Additionally, a multitude of “black box models” (eg, deep neural networks) have shown spectacular results in various fields, including the prediction of clinical deterioration [10,16,19]. In 2022, Ruiz et al [19] demonstrated their retrospective data-driven extreme gradient boosted model aimed at predicting clinical deterioration (defined as adverse events, such as intubation, cardiopulmonary resuscitation or initiation of extracorporeal membrane oxygenation) in cCHD over a time frame up to 8 hours. Through the model’s assessment of 1028 variables (eg, medication, vital parameters, laboratory values, etc), they have achieved accurate predictions and good calibrations with at least 4 hours prior to intubation (area under the receiver operating characteristic curve 0.927, 95% CI 0.825-0.994) or cardiopulmonary resuscitation and extracorporeal membrane oxygenation (area under the receiver operating characteristic curve 0.914, 95% CI 0.796-0.991).

However, the methodological foundations of such complex models remain beyond the grasp of most clinicians. It is likely that models with explainable methods are more likely to be implemented in daily practice and, therefore, explainable modeling techniques were used in this study. The clinical usefulness of our proof of concept, however, has yet to be proven as it is currently limited by its underpowered sample size and the retrospective analysis of model performance. In the near future, the model will be trained and evaluated on a more heterogeneous population to increase performance and versatility, boosting the chances of successful (external) validation while maintaining a sharp clinical perspective: how can the algorithm be most valuable to both patients (eg, early intervention and reduced risk of injury) as well as the medical team (eg, reduced cognitive load)? 

Several other limitations to this proof-of-concept study must be addressed. First (and foremost), selection bias was introduced through the inclusion of patients with cerebral rSO₂ measurements, as well as IBP. Cerebral rSO₂ monitoring is currently not available as a standard of care in global (cardiac) PICUs, and as such, the clinical value of our model will decrease outside the research institution. Additionally, a relatively high sample rate of 1 Hz was used to extract data. As not all parameters are transmitted at the same frequency, internal sampling or resampling is inevitable, possibly affecting data quality.

Second, we have chosen a retrospective approach to analyze model performance. Analyzing patient stability solely based on retrospective parameters remains particularly challenging, even for medical experts. Increases or decreases in parameter values may, for example, originate for a number of reasons, such as feeding or movement, and may have little to no clinical significance. The classification of episode stability or sensor dysfunction was evaluated by expert consensus based on the same data available to the model. However, no hard judgments can be made on the clinical relevance of that episode, as the data were not labeled prospectively (ie, containing labeled events). Arguably, prospective validation with members of the medical team performing a simultaneous bedside evaluation on agreement with the model will be one of the future goals.

Third, in the stability analysis of parameter combinations, an SVM was trained to recognize stability across 5 dimensions. In selecting presumably stable parameter combinations, an 80th percentile split of the vector’s corresponding Mahalanobis distance was made, partially based on earlier work by Clifton et al [14]. However, since no explicit labeling was possible in the data set, the chosen cutoff percentile remains arbitrary. Additionally, through the use of normalized data, an assumption is made that any deviation from subset-specific mean values reflects an adverse development. However, for some parameters, an increase or decrease does not necessarily reflect an adverse event, which may result in an overestimation of clinical status and aid in the induction of alarm fatigue. Future research may point out different methods to be more effective.

Several steps must be taken to progress this model—and others alike—toward implementation in daily clinical practice [28,29]. Primarily, a data infrastructure is required to enable real-time or near–real-time data availability to AI models, allowing their prospective validation. In the near future, such a platform will be constructed, speeding up the qualitative performance analysis of data science models while promoting guideline adherence, such as the TRIPOD guidelines [22]. Eventually, AI models will be implemented into the daily workflow, aiding the medical team and likely decreasing their cognitive load, which is beneficial for, in this instance, the continuous interpretation of clinical data streams in hemodynamically challenging patients.

In this study, a proof-of-concept algorithm aimed at detecting clinical deterioration in patients with cCHD at the PICU was developed and retrospectively evaluated, achieving reasonable performance considering the heterogeneous population of neonates with cCHD. Combined analysis of baseline (ie, patient-specific) deviations and simultaneous parameter-shifting (ie, population-specific) proofs to be promising with respect to enhancing applicability to heterogeneous critically ill pediatric populations.

Although performance should be improved and prospectively validated, advanced data science models such as the one presented here may, in the future, be used in automated detection of clinical deterioration, providing real-time data-driven monitoring support in the case of hemodynamically challenging patients and allowing for timely intervention.