This was a prospective, observational study conducted in the Neonatal Intensive Care Unit (NICU) of the Department of Pediatrics, Semmelweis University Budapest, Hungary. The study was approved by the Ethics Committee of the National Medical Research Council (ETT-TUKEB 11790-2/2016/EKU and 7469-12/2021/EUIG). We enrolled patients who were born between July 2016 and February 2019, were diagnosed with moderate-to-severe HIE as described by Azzopardi et al. (4), and were cared for in our level III NICU, a regional cooling center. All patients were outborn. Inclusion criteria were (1) whole body TH for moderate-to-severe HIE, (2) non-invasive hemodynamic monitoring during TH. Exclusion criteria were (1) gestational age of <37 week, (2) postpartum asphyxia, (3) cooling initiated >6 h of life or other cooling protocol violations, (4) technical problems with data registration (equipment interference or sensor problems), (5) congenital malformations, (6) death during TH, (7) missing neurodevelopmental follow-up examination. Patient recruitment and exclusions are shown on Figure 1. Eight patients with persistent pulmonary hypertension (PPHN) receiving high-frequency ventillation (HFO) were excluded for technical problems with data registration, because of the interference between the non-invasive hemodynamic monitor and the HFO ventilation mode. None of the patients had sepsis in our cohort.

Non-invasive hemodynamic monitoring was applied using the ICON electrical velocimetry device (Ospyka Medical GmbH, Berlin, Germany). The ICON continuously recorded HR, SV and CO and measurements were averaged in every 60 s leading to a time-series dataset with minute frequency. For data analysis, CO and SV values were divided by the infant’s birth weight. Measurements are shown in the following format: ml/kg/min for CO, ml/kg for SV, and beats/minutes for HR.

The quality of the measured signal was verified using the signal quality index (SQI) values of the monitor. Low-quality data points of the measurement signal (SQI <80) were discarded from the statistical analyses, as recommended previously (15).

Whole-body hypothermia treatment was induced as described in the TOBY trial; initiated within the first 6 h of life, maintained for 72-hours, followed by gradual rewarming (4, 16). All patients were mechanically ventilated during the hypothermia treatment according to our institutional clinical protocol. Respiratory management of individual cases is decided by the attending neonatologist, but our institutional guideline is to use SIMV-VG (synchronised intermittent mandatory ventilation with volume guarantee) mode without pressure support, with a PEEP (positive end-expiratory pressure) of 5 cmH₂O. The target expired tidal volume is 4–5 ml/kg initially, which is adjusted later based on blood gases.

All patients were sedated with morphine sulfate infusion (loading dose of 100 µg/kg, followed by continuous infusion of 10 µg/kg/hour). Cerebral activity was recorded using continuous single-channel (biparietal, P3-P4) amplitude integrated electroencephalography (aEEG) monitoring (Olympic CFM 6,000 monitor; Natus Medical or EEG-1200 K; Nihon Kohden). Recording was started as soon as possible but always before 6 h of age and continued for the duration of therapeutic hypothermia and rewarming. The aEEG background activity was scored according to the Hellstrom–Westas (17) criteria, and aEEG normalization was defined as attaining continuous normal voltage activity.

Cardiovascular management was based on the mean arterial pressure targeted to reach the value of the gestational age in weeks (18). If hypotension occurred, vasopressor agents were initiated based on the decision of the clinician. Dopamine infusion was started usually as first-line therapy at a rate of 6 µg/kg/min, and if necessary, the dose was increased in 2 µg/kg/min increments up to 20 µg/kg/min. As per local protocol, 0.5 mg/kg hydrocortisone every 6 h was used frequently as a second-line therapy, in cases of suspected relative adrenal insufficiency (based on clinical and laboratory findings) or vasopressor-resistant hypotension (4, 16, 19). If needed, further vasopressor-inotropes were added including dobutamine, epinephrine, norepinephrine or milrinone. Vasotropic-inotropic score [VIS] was calculated as: dopamine dose (µg/kg/min) + dobutamine dose (µg/kg/min) + 100 × epinephrine dose (µg/kg/min) + 10 × milrinone dose (µg/kg/min) + 10,000 × vasopressin dose (unit/kg/min) + 100 × norepinephrine dose (µg/kg/min), it was calculated in every hour during hypothermia and rewarming, which allowed for direct comparison of inotropic needs in each patient over time (20). Clinicians were blinded to the ICON measurements.

MRI was performed at the Medical Imaging Center, Semmelweis University, Budapest, using a 3 Tesla Philips Achieva scanner (Philips Medical System, Best, The Netherlands). T1-, T2-, and diffusion-weighted MRI with apparent diffusion coefficient measurements were performed. MRI scans were classified using the scoring system developed by Weeke et al. (21).

Bayley Scales of Infant Development (II. Edition) (22) was used at two years of age to evaluate the neurodevelopmental outcome undertaken by a trained psychologist blinded to the clinical history. Favorable outcome was defined by scoring ≥70 points on both the mental development index (MDI) and psychomotor development index (PDI) scales, while adverse outcome was defined by scoring <70 points on either scale or death. Patients who were unable to complete the Bayley exam due to severe neurological impairment were automatically assigned to the adverse outcome group.

Descriptive statistics were expressed as mean ± standard deviation (SD) or median with interquartile ranges [IQR] for continuous variables and as number with percentage (%) for categorical variables. Fisher’s exact test and Wilcoxon rank sum test were used to compare groups on categorical and continuous variables, respectively.

Time-series features were derived from ICON signals as follows: (1) A common time window was selected with relation to the start of rewarming to ensure that the same period is analyzed for every patient; (2) individual HR, SV, and CO signals in the selected time window were plotted on the y axis against time on the x axis by applying LOWESS (locally weighted scatterplot smoothing) both for each patient separately and at the group-level; (3) signals with visible breakpoints at the start of rewarming were modeled by segmented robust linear regression (as implemented by the lmrob function of the robustbase R package (23), that is, two separate slope coefficients were estimated per signal (one for hypothermia and one for rewarming period) besides the intercept (representing the estimated value at the onset of rewarming); (4) signals with no sign of rewarming-triggered change were modeled by robust linear regression as a function of time, resulting in two features describing the signal: intercept at the start of rewarming period and slope.

To control for the vasoactive-inotrope need during the study period, a simple linear regression model was built to model VIS change over time in the selected time window, and the estimated VIS score at the start of rewarming was used afterwards. Due to skewed data, VIS scores were logarithmically transformed before calculations.

A multiple logistic regression model was developed to ascertain the effects of the extracted ICON time-series features on neurodevelopmental outcome. Since the CO signal is a function of HR and SV, only SV, HR, and VIS features were included in the model.

Finally, receiver operating curve statistics were calculated with the area under the curve analysis to evaluate the predictive power of the logistic regression model.

We accepted p < 0.05 as a level of significance. R Statistical Software, Version 4.2.1 (R Foundation for Statistical Computing, Vienna, Austria) was used for data analysis and plotting.

Born between July 2016 and February 2019, a total of 116 HIE patients received TH in our institute, of whom 50 patients were monitored with ICON. For the present analysis, we excluded patients who had low gestational age (4 patients) or had postnatal collapse (2 patients), or were not adhering to cooling protocol (3 patients). Another 15 patients were excluded due to late screen failure: technical problems with data registration (11 patients); rewarming was not registered (2 patients); death during hypothermia (1 patient); and lost to follow-up (1 patient). After these exclusions, 26 patients were included in the study, of whom 14 (54%) patients had favorable and 12 (46%) had adverse neurodevelopmental outcome at two years of age (performed at a median 19 [18–20] months). Flow diagram of study enrollment is shown in Figure 1. Demographics and clinical characteristics of the patients are shown in Table 1. Patients in the adverse outcome group had more severe brain injury on the MRI (performed at a median 4 [4; 5] days of life), and aEEG normalization occurred less frequently, and later than in patients with favorable outcome.

To treat hypotension, vasotropic-inotropic agents were administered in 23 (88%) patients. Dopamine was used in almost all patients, only 2 patients received a combination of vasoactive therapy and 1 patient received norepinephrine along with dopamine and dobutamine. There was no difference between the need for vasoactive drugs between the two groups, as shown in Table 2. We used a continuous VIS-score (19) to assess possible differences in vasoactive treatment between the groups that may influence cardiac function. The hourly VIS-score during TH was median 5.4 [3.5; 11.0] in the favorable outcome group and 5.3 [0.0; 7.3] in the adverse outcome group (p = 0.049), as shown in Figure 2. In addition, 50% of patients received hydrocortisone therapy as well.

Descriptive statistics of the raw ICON signals are shown in Table 3. Recordings started at a median 11.8 [7.3; 24.0] hours of life and ended at 84.0 [82.0; 86.8] hours. The median length of ICON measurements and number of valid data points, with signal quality index higher than 80, were similar in the adverse and favorable outcome groups.

All individual recordings were aligned to the onset of rewarming. The lengths of pre-rewarming and rewarming phases did not differ between groups (Table 3), but there was substantial variation at the individual level. Therefore, a common cutoff was applied to ensure that time-series features were extracted from the same time windows in relation to rewarming. Cutoff points were determined by visual inspection of the signals smoothed individually by locally weighted scatterplot smoothing (see Figure 3). Because the signals were relatively stable in the −24 h to +6 h interval, and all but two patients had full coverage in that interval (one having shorter pre-rewarming, the other shorter rewarming phases), individual recordings were cut at −24 and +6 h with relation to the onset of rewarming.

In line with our expectations, based on the results of previous studies (10, 11, 13), the slopes of HR and CO signals clearly changed after the start of rewarming (see Figure 3), whereas no apparent breakpoint could be detected for SV signals. Thus, segmented robust linear regression was performed for each patient to extract HR and CO features, and non-segmented robust linear regression models were fitted to SV signals (resulting in 3 features [2 slopes and 1 intercept] per patient for HR and CO, and 2 features [1 slope, 1 intercept] for SV signals, respectively). Thus, feature extraction yielded a total of 8 features per patient for describing the time-series HR, CO, and SV data, their descriptive statistics are presented in Table 4. We also present the comparison of estimated values for the cardiac function parameters at −24 h (hypothermia), 0 h (start of rewarming), and at 6 h (end of rewarming) (Table 5).

The estimated individual time-series features (that is, the individual regression lines) and their group-level averages are depicted in Figures 4–6. In general, the estimated slopes in the two groups were highly similar: HR and CO values were relatively stable in hypothermia, and started to increase during rewarming, with an estimated 25% increase in CO. In contrast, SV values showed only a slight increase in the analyzed time window. Regarding the estimated intercepts, that is, the estimated values of the signals at the onset of rewarming, the HR was significantly higher in the adverse group compared to the favorable outcome group (104 vs. 83 beat/min), and SV was lower (1.29 vs. 1.55 ml/kg, respectively).

Next, a logistic regression model was developed to ascertain the effects of the cardiac function features on neurodevelopmental outcome. As predictors, we included the three HR features (slope in hypothermia, value at the onset of rewarming, and slope during rewarming), two SV features (slope in the −24 to +6 h period and value at the onset of rewarming), and the VIS score (logarithm at the onset of rewarming). The CO was not included, as it is a function of HR an SV. The multiple regression yielded that only HR value at the onset of rewarming had a significant effect on the outcome, with higher heart rate increasing the odds of adverse outcome (odds ratio: 3.2, CI: 1.38–10.81, p = 0.020). The result suggests that a 10 beat/minute difference in HR at the onset of rewarming is associated with an increase of 3.2 in the odds of developing adverse neurological outcome (see Table 6). By testing all models which consist of all possible combinations of the predictor variables, we confirmed it is only the HR intercept which remains significant (data not shown). The AUC of the model is 0.893 (Figure 7).