Information Based Similarity Analysis of Oxygen Saturation Recordings to Detect Pulmonary Hypertension in Preterm Infants

Abstract

Pulmonary hypertension (PH) is a complex cardiovascular condition associated with multiple morbidities and mortality risk in preterm infants. PH often complicates the clinical course of infants who have bronchopulmonary dysplasia (BPD), a more common lung disease in these neonates, causing respiratory deterioration and an even higher risk of mortality. While risk factors and prevalence of PH are not yet well defined, early screening and management of PH in infants with BPD are recommended by consensus guidelines from the American Heart Association. In this study, we propose a screening method for PH by applying a signal analysis technique to oxygen saturation in infants. Oxygen saturation data from infant groups with BPD (41 with and 60 without PH), recorded prior to their clinical PH diagnosis were analyzed in this study. An information-based similarity approach was applied to quantify the regularity of SpO₂ fluctuations represented as binary words between adjacent five-minute segments. Similarity indices (SI) were observed to be lower in subjects with PH compared to those with BPD alone (p<0.001). These measures were also assessed for performance in screening for PH. SI of 7-bit words, exhibited 80% detection accuracy, 76% sensitivity and specificity of 83%. This index also exhibited a cross-validated mean (SD) F1-score of 0.80 (0.08) ensuring that sensitivity and recall of the screening were balanced. Similarity analysis of oxygen saturation patterns is a novel technique that can be potentially developed into a signal based early PH detection method to support clinical decision and care in this vulnerable population.

Graphical Abstract

Introduction

Survival rates of extremely preterm infants, born between 24 and 28 weeks’ gestation have risen over the past decades due to advances in neonatal medicine [1]. The risk of respiratory morbidity and mortality is nevertheless high in these infants, primarily due to neonatal chronic lung disease referred to as bronchopulmonary dysplasia (BPD) [2]. Pulmonary hypertension (PH), increased pressure in the pulmonary arterial system, complicates the clinical course of extremely preterm infants and is associated with cardiovascular and neurocognitive morbidities and increased risk of late-stage mortality, in this vulnerable group [3-5] . Among infants with BPD, PH is a complication in ~ 40% of infants and increases their risk of mortality [3,6] . While risk factors for development of PH in extremely preterm infants are still ill-defined, the need for evaluation and management of this condition is recommended by consensus guidelines from the American Heart Association [3,7-9] . In addition, early detection of PH using screening techniques may lead to early and better treatment [3].

However, infants with BPD regardless of PH presence, frequently exhibit deterioration of respiratory status with persistent or increasing oxygen requirement, making clinical features of PH such as severe oxygen desaturations and poor growth indistinguishable from that of the underlying BPD [10] . Results from recent studies suggest that the frequency and duration of these desaturations, known as intermittent hypoxemia (IH), may provide clinical utility for BPD prediction[11,12] .

We compared the frequency, duration and severity of such events between infants with BPD and PH (BPD-PH) and those with BPD alone (BPD group). This was carried out using oxygen saturation data collected in the week before PH was diagnosed clinically by echocardiogram. We reported that the average duration of IH events were longer in duration in infants with BPD-PH as compared to infants with BPD-alone and, in infants with established BPD-PH, IH event duration differed by survival status [12] . Such differences in desaturation events led us to consider the signal analysis of oxygen saturation data in this population to develop a screening method for PH. From our previous study, we realized that identification of IH events that might differentiate infants with BPD alone from those with BPD-PH would require longitudinal data analysis over the entirety of infants’ hospitalization. However, fluctuations in oxygen saturation may be more severe and persistent in the PH group than in the group with BPD alone. Our objective was hence to derive a time limited signal characteristic to effectively quantify similarities in the fluctuations of short-term oxygen saturation levels to infants with both BPD and PH.

Various linear and nonlinear methods applied to physiological heart rate variability and oxygen saturation are being actively investigated in risk stratification of infant and pediatric subjects for conditions such as BPD, obstructive sleep apnea and sepsis [13-16] . These measures mostly describe the overall static properties over a time interval related to the variability, irregularity or signal complexity, but are inadequate to quantify the dynamic changes to the regularity of signal fluctuations. Hence in this study, SpO₂ data was analyzed using an information-based similarity technique, that quantifies similarities in the occurrence of binary sequences representing changes in signal levels. This measure was introduced by Yang et al. to discriminate between heart rate patterns generated from healthy subjects and those with pathological congestive heart failure [17] . In addition to physiological signals, this measure has been applied extensively to analyze data in diverse fields such as genomic sequences and linguistic texts [18,19] . Recently, adjacent segments of inter heart beat data were studied to derive similarity indices that distinguished normal subjects from groups with moderate and severe obstructive sleep apnea (OSA) [20] . Akin to this scenario, BPD with PH may be considered as manifesting a more severe contributor to IH than BPD alone. Persistent patterns of SpO₂ fluctuations over neighboring time segments were hence quantified to derive differences in short-term oxygenation patterns between infants with BPD alone and those with BPD and PH. These were finally leveraged to screen for PH presence prior to clinical diagnosis. Such a screening tool based on continuously monitored signals in neonatal intensive care units (NICU) can potentially aid early clinical decision and provide a continuous bedside marker for development of PH in neonates.

1. Methods

1.1. Data and Subjects

Oxygen saturation data was prospectively collected in extremely preterm infants admitted at the University of Alabama at Birmingham (UAB) between 2018 and 2020 as part of the Prematurity-Related Ventilatory Control (Pre-Vent) study. This study approved by the Institutional Review Board (IRB-UAB) at the University of Alabama at Birmingham, had oversight provided by both IRB-UAB and an observational and safety monitoring board, appointed by the NHLBI. Approval from the institutional review board was obtained for physiological and clinical data collection with waiver of consent. At UAB, infants remaining on respiratory support after 28 postnatal days are systematically screened every month for pulmonary hypertension by echocardiogram. Infants were included if born at <29 weeks’ gestation, remained on respiratory support on postnatal day 28, had an echocardiogram performed to screen for PH, and had oxygen saturation data available in the week preceding echocardiographic diagnosis. Infants with major congenital anomalies or genetic syndromes were excluded. Table 1 gives a description of demographic and clinical characteristics of the subjects. Among these, data on some known risk factors for BPD-PH such as, small for gestational age, (congenital) patent ductus arteriosus (persistent opening between the two major blood vessels leading from the heart), sepsis and respiratory support in the study population have been presented.

Table 1:

Demographic and clinical characteristics of the infant groups with BPD and PH (BPD-PH) and BPD alone (BPD). Values are given as median (IQR) and number (%) where applicable.

	BPD group N=60	BPD-PH group N=41
Gestational age (weeks)	25.9 (24.1,28.1)	24.4 (23.3,25.9)
Birth weight (grams)	700 (604, 939)	587 (522,650)
Sex	Male: 32 (53)	Male: 21 (51)
Race	White: 28/59(47)	White: 11 (27)
Small for gestational age	9 (15)	13(32)
Maximum fraction of inspired oxygen at 28 days, FiO2	40 (30,51)	51 (45,75)
Ventilatory support at 28 days	49 (82)	38 (93)
Outcomes
Patent ductus arteriosus	15 (25)	23(56)
Necrotizing enterocolitis	7(12)	4(10)
Sepsis	9 (15)	15 (37)
Death	4 (7)	9 (22)

Bedside cardiorespiratory monitoring was carried out by Philips IntelliVue MP70 or MP50 monitors using Nellcor pulse oximetry sensors, an averaging time of 8 seconds, and recorded using the BedMaster system (Excel Medical Electronics, Jupiter FL, USA) at 125Hz with data analyzed using numeric data at 1 Hz. Forty- one infants with BPD and diagnosed PH (BPD-PH group) were included in the study along with sixty infants without PH (BPD group).

Six hours of SpO₂ data acquired seven days prior to PH diagnosis were included in this study. In our previous study [12] , we studied distribution of hypoxemic event statistics on a subset of these subjects and used continuous data from all 7 days leading up to the clinical diagnosis. This analysis is therefore focused on using a short-term recording to seek differences in oxygenation between the groups and to provide PH detection with good diagnostic accuracy metrics. Similarity indices were estimated over short non-overlapping data segments and averaged over the available observed time for each subject. Hence, subjects were not excluded, as enough data segments were available from all to average, even though some subjects had shorter than 6 hours of data. Data were preprocessed prior to analysis as follows. Recording errors caused missing SpO₂ values which were identified from the simultaneously recorded time vector. Signal gaps of less than 10 seconds were imputed by the average of previous three seconds. Segments with over 5% of missing values were excluded from the analysis. Missing data over observed time [median (min,max)] for all subjects was 0 (0, 3.9%).

1.2. Information based similarity analysis

An information based approach to quantify the similarity of two symbolic sequences has been developed to classify varied physiological signals including inter heart beat (RR) intervals and DNA sequences [17,18] . This method relies on finding temporal dynamical structures hidden in the complex fluctuations of output signals from biological systems and examining their rank order statistics to differentiate between states, such as healthy and pathological [17] . Noninvasive detection of conditions such as congestive heart failure and obstructive sleep apnea has been achieved previously by applying this measure to heart rate variability signal [20,21] .

This technique quantifies the similarity between two data series by first defining a binary symbolic sequence for each. Here, we converted the oximetric data series, x = {x₁,x₂,…, x_N} to a binary symbolic sequence as, 0 if x_i+1 = x_i; i=1,2,…N and 1 otherwise. Hence, in the case of SpO₂ data, the sequence represents variations or changes from the previous value. Following this, the sequence was divided into m-bit words by sliding the binary number one bit at a time. Hence a sequence of length (N-1) could give 2^m, unique m-bit words. Figure 1 illustrates the binary sequence creation and construction of 5-bit words for an SpO₂ data snippet. The first four SpO₂ values in the snippet are the same and are hence represented by three leading 0 bits in the binary sequence indicating no difference from their respective previous values. The fifth value is different from the fourth and this is represented by 1 in the next position and so on. From the binary sequence, the first m-bits are included to form the first word, the next m-bits starting from the second bit, form the second word etc. as shown for the case of m=5 in Figure 1.

Figure 1:

Schematic illustration for creation of binary sequence and construction of 5-bit words for SpO₂ data. The words generated by sliding one bit at a time, and the decimal equivalent of each word, 3,7, 14 etc. are also shown here.

Word frequencies are enumerated by their occurrence in the sequence and these rank numbers are sorted for words with highest frequency to lowest. The similarity of two data series is quantified by such word ranks generated from the symbolic sequences of each series. For this, the rank of each word in one series is plotted against the corresponding word in the other one. If the two series are similar in their rank order of words, the scattered points will lie close to the diagonal line. The average deviations of these points from the diagonal are considered a measure of ‘distance’ between them. The more scattered the points are, the less similar the two segments and vice-versa [17] . For two data series, $S_1$ and $S_2$, this average deviation for m-bit words is defined as,

$$D(S_1,S_2)=\frac{1}{L}{\sum }_{k=1}^L\mid R_1(w_k)-R_2(w_k)\mid F(w_k)$$ (3)

where L denotes the total number of unique m-bit words, $R_i(w_k)$, i = 1,2 is the rank of word $w_k$ for the binary sequences from $S_1$ and $S_2$. The rank difference is weighted by the factor $F(w_k)$ that accounts for contributions from different words to the overall measure and is calculated as

$$F(w_k)=\frac{[-p_1(w_k)log(p_1(w_k))-p_2(w_k)log(p_2(w_k))]}{Z}$$ (4)

and the normalization factor, $Z$ is,

$$Z={\sum }_{k=1}^L[-p_1(w_k)log(p_1(w_k))-p_2(w_k)log(p_2(w_k))]$$ (5)

It can be noted that $-p_i(w_k)log(p_i(w_k))$, i = 1,2 represents the Shannon entropy of the word $w_k$ for each sequence. The weighing factor, $F(w_k)$ hence incorporates the information content carried by the m-bit words in the sequences being compared [22]. In this manner, the higher the frequency of occurrence of words, greater will be their contribution to quantifying similarity. $D(S_1,S_2)$, the information-based similarity index represents the average deviation of the ranks from the diagonal line and in this sense is a measure of the ‘dissimilarity’ between $S_1$ and $S_2$. The greater the scatter from the diagonal, higher the $D(S_1,S_2)$ or SI value and lower the similarity between the data series being compared. In addition, it has been specified that this similarity index is an empirical measure which does not always obey the triangular inequality criterion for a distance measure [22]. The normalization by the total number of words keeps the index in the range [0,1].

In this analysis, the data series compared were adjacent non-overlapping segments of SpO₂ signal series for the two subject groups. This ensured that dynamic changes to signal similarity that might be related to the underlying pathology were studied and quantified within each subject group. Information based similarity index for word lengths 3 to 10 for adjacent segments of 5 minutes, $(S_{i+1},S_i)$, 1 ≤ i ≤ (n − 1) where n is the number of segments in each recording were derived. The final similarity index for a subject, SI was obtained by averaging over all the values computed for each recording. Figure 2 displays a schematic of estimating the SI from the data record of a single subject.

Figure2:

Schematic representing the screening for PH by estimating information-based similarity index for a subject.

1.3. Statistical analysis and validation

Signal preprocessing, estimation of indices as well as statistical analyses were carried out using MATLAB (R2021b, MathWorks, Natick, MA, USA). Similarity indices were calculated using the software in [23] . Kolmogorov–Smirnov test was used to check normality of the variables. The indices for BPD and BPD-PH groups were tested for equality of means by the independent samples t-test in case of normal variables and the Wilcoxon rank test for equality of medians otherwise. The results were expressed as the mean ± standard deviation (SD). Significant difference was indicated by p < 0.05. Univariate logistic regression was used to classify infants with PH using each SpO₂ similarity feature as predictor and the performance was evaluated using accuracy, sensitivity, and specificity. In addition, the Mathews correlation coefficient which accounts for agreement between observed and predicted labels by including all elements of the confusion matrix was also estimated [24,25]. This metric ranges from −1 to 1, with 0 corresponding to a random classifier. The receiver operating characteristic (ROC) curve and area under the ROC curve (AUROC) of the screening models were also determined for each classifier. In addition, the bootstrap method with 1000 replications was applied to derive the 95% confidence interval for the ROC.

A hold-out strategy was adopted for cross validating the classifiers. 30 infants from the BPD-PH group (40%) and 45 from the BPD group (60%) were randomly sampled to form a stratified training set with the remaining (25%) as test data. Stratification ensured that the same distribution of positive (BPD-PH) and negative (BPD) classes was retained across the original and test-train sets. Performance was evaluated by calculating the balanced accuracy and F1 score, metrics recommended for evaluating classification on imbalanced data sets [26] . Balanced accuracy denotes mean of the sensitivity and specificity while F1 is the harmonic mean of the sensitivity and positive predictive value of the classifier. This was repeated 100 times and the mean and SD of these metrics were calculated for each feature.

2. Results

2.1. Similarity Index estimation and analysis between BPD-PH and BPD groups

The information-based similarity index (SI) was estimated for adjacent non-overlapping SpO₂ data segments. Representative examples from a PH subject are shown in Figure 3. adjacent segments with more constant levels of SpO₂ as in subplot (a) showed more ranks for 5-bit words closer to the line of identity, resulting in a lower SI and another pair of segments as in (b) that were more dissimilar had higher SI estimates. Hence the similarity index can track the variations in the regularity of SpO₂ fluctuations between adjacent non-overlapping segments.

Figure 3:

Adjacent SpO₂ segments from a BPD-PH subject (a) and (b) representing different regularity in their fluctuations. The rank comparison of 5-bit words for the segments in (a) is plotted in (c). The corresponding plot for the segments in (b) is shown in (d). The line of equality in (c) and (d) helps visualize agreement between ranks for different words between the two segments.

2.2. Comparison of SpO₂ similarity indices between BPD-PH and BPD groups

Similarity indices were estimated over all adjacent segments available from the recording for each subject. These were averaged over the observation period to provide a mean similarity index for each subject and compared between groups. 98 subjects had all 6 hours, one BPD-PH infant had 4.63 hours and two BPD infants had 5.31 and 5.64 hours of data respectively. The choice of 5-minute long segments was made to ensure adequate frequency for the binary words of differing bit-lengths from m = 3 to 10 [27,28] .

The dynamic changes of SpO₂ measures for the two groups are shown in Figure 4. SI for 7-bit words was well separated between subject groups, with the BPD group’s values topping the BPD-PH group’s indices over the observed periods.

Figure 4:

Temporal variation of the estimated SI for 7-bit words for the two groups. The lines represent the mean value of the index and the shaded portion, standard error for the BPD and BPD-PH groups at each analyzed segment over the whole recording.

The mean (SD) of these measures are summarized by groups in Table 2. All measures passed the Kolmogorov test for normality. The SI indices were all significantly higher in the BPD group compared to the BPD-PH group (p<0.001). We also conducted this analysis for the shorter recording times of 2 and 4 hours with similar results (Supplementary Table 1). However, the measures averaged over the entire 6 hour period had more 5-minute segments included and consequently gave lower standard deviation estimates and the greatest differences between the BPD-PH and BPD groups.

Table 2:

SpO₂ similarity indices (mean (SD)) for the BPD-PH and BPD groups.

Measure	BPD N=60	BPD-PH N=41	P
SI_m3	0.10(0.03)	0.08(0.01)	<0.001
SI_m4	0.12(0.02)	0.10(0.01)	<0.001
SI_m5	0.14(0.02)	0.12(0.01)	<0.001
SI_m6	0.16(0.03)	0.14(0.02)	<0.001
SI_m7	0.20(0.03)	0.17(0.02)	<0.001
SI_m8	0.24(0.05)	0.20(0.03)	<0.001
SI_m9	0.29(0.06)	0.24(0.03)	<0.001
SI_m10	0.33(0.08)	0.27(0.04)	<0.001

SI_mi, i = 3,4, …,10: Similarity index of , i-bit word

2.3. Screening Performance of Similarity indices

Univariate logistic regression was used to screen for PH in the subject sample. The estimated coefficients, model equations and p-values are given in Supplementary Table 2. The ROC curve was considered and the area under it (AUROC) computed to provide a measure of the discriminative ability of each SI feature. Screening analysis was also conducted using estimates averaged over a shorter period of 4 hours and the comparison between the computed AUROC metrics is presented in Figure 5. This analysis ascertained whether the discriminative ability differed when the observation time was longer, and secondly, decided which measures exhibited superior classification capability.

Figure 5:

Area under ROC of screening models for each of the SpO₂ features averaged 4 hours and 6 hours.

SI of binary words m ≥4, averaged over 6 hours observation time exhibited AUROC > 0.75. When averaged over 4 hours, only four of these indices had AUROC higher than 0.75. SI_m4 had the highest discriminative ability among all measures considered and close values (~ 0.81) at both averaging periods. This was also observed for SI_m5. SI for all other word lengths had higher AUROC when averaged for 6-hours. This analysis clearly demonstrated that similarity indices discriminated between BPD-PH and BPD groups, and estimates averaged over (at least) 6 hours would likely perform better classification than those over shorter observation periods. However, SI at shorter word lengths of 4 and 5 may still be suitable for PH screening when only short SpO₂ data sets are available.

The performance metrics of PH screening by similarity indices of words with 3 ≤ m ≤ 10 bits are presented in Table 3. Sensitivity of the screening is of utmost importance in a clinical setting, so that ideally all PH subjects are correctly classified. With more correctly classified PH subjects, SI for m = 4 to 7 showed incremental sensitivity, specificity, and accuracy. SI_m7 had the best sensitivity, specificity pairing as (0.76, 0.83). Figure 6 displays the ROC curve and confusion matrix of the screening based on this similarity index. The Mathews correlation coefficient (MCC) evaluated for models with these features was found to vary from 0.21 at m=3 to 0.3 at m=10, with a peak value of 0.59 for word length of 7, all of which were higher 0, the value for a random classifier. In addition to the area under ROC values themselves, the lower bound of 95% confidence interval was also more than 0.5 corresponding to a random classifier. The similarity features thus differentiated between the two subject groups considered. Based on all these metrics, the best screening performance for PH was achieved by the information-based similarity indices estimated for adjacent 5-minute segments and averaged over 6 hours and for word lengths of 4 ≤ m ≤ 7.

Table 3:

Screening performance of similarity indices

Classifier Feature	TP	FP	TN	FN	Acc.	Sen.	Spe.	AUROC (95% CI)
SI_m3	20	17	43	21	0.624	0.488	0.717	0.73(0.62,0.82)
SI_m4	28	15	45	13	0.723	0.683	0.750	0.81(0.71,0.88)
SI_m5	28	15	45	13	0.723	0.683	0.750	0.78(0.69,0.86)
SI_m6	29	13	47	12	0.752	0.707	0.783	0.78(0.68,0.86)
SI_m7	31	10	50	10	0.802	0.756	0.833	0.79(0.69,0.87)
SI_m8	27	13	47	14	0.733	0.659	0.783	0.78(0.66,0.87)
SI_m9	25	13	47	16	0.713	0.610	0.783	0.77(0.64,0.85)
SI_m10	23	14	46	18	0.683	0.561	0.767	0.76(0.66,0.85)

SI_mi, i = 3,4, …,10: Similarity index of i-bit word, TP: true positive, FP: false positive, TN: true negative, FN: false negative, Acc.: accuracy, Sen.: sensitivity, Spe.: specificity, AUROC: area under ROC, CI: confidence interval

Figure 6:

PH screening performance of the SI_m7 feature. (a) Receiver operating curve (b) confusion matrix

Finally, screening performance of features was validated on 100 trials with held out test sets of ~ 25 subjects. Figure 7 displays the mean (SD) of balanced accuracy and F1 metrics for the classification by SI of m=4 to 9 words. For unbalanced data, with more controls (BPD) than cases (BPD-PH), these metrics are recommended for performance evaluation of prediction models. SI_m7 exhibited the highest F1 and balanced accuracy of 0.80(0.08) and 0.77(0.08) respectively, while other SI had test metrics ~ 70%. With high sensitivity, specificity, and positive predictive values, SpO₂ similarity indices can thus be effective in PH screening.

Figure 7:

Cross-validation metrics (Mean +/− SD) of screening with SpO₂ features over 100 trials. (a) Balanced Accuracy, (b) F1 score.

3. Discussion

Information-based similarity analysis of oxygen saturation data was demonstrated to detect pulmonary hypertension (PH), a debilitating cardiorespiratory complication, in a population of preterm infants with BPD. This non-invasive technique could detect PH in infants with good diagnostic accuracy by analyzing adjacent data segments, one week before clinical diagnosis. The similarity index corresponding to 7-bit binary words achieved maximum screening accuracy, sensitivity, and specificity of 80%, 76% and 83% respectively. To the best of our knowledge, this is the first application of a signal analysis technique leveraging routinely monitored oximetric data in the NICU for PH screening.

3.1. Proposed method and screening performance

Information-based similarity analysis quantifies the regularity of SpO₂ fluctuations between adjacent segments of data by enumerating patterns of m-bit binary words. The method identified greater similarity between adjacent SpO₂ segments for the BPD-PH group when averaged over time (Table 1) and over subjects (Figure 3), supporting our initial hypothesis. Similarity indices of different word lengths were consistently lower for BPD-PH in comparison to the BPD group (p<0.001). Based on area under ROC and diagnostic metrics, these indices also discriminated BPD subjects with PH from those with BPD alone. For segments of 5-minute duration, best screening performance was by 7-bit binary words beyond which performance deteriorated. We can therefore infer that if segment length increased, longer words may likely occur, plausibly with differing frequencies, thus providing different estimates for the similarity index in each group and hence better screening performance. In this study, we used 6 hours of data, but have also shown discrimination between groups with SI estimates averaged over 4 hours, albeit with lower performance metrics (Figure 4). The optimal choice of segment length for similarity analysis, the length m of the binary word and the averaging time will depend on this available data length.

The cross-validation of the univariate screening showed variability (SD: 8%-10%) in the balanced accuracy and F1 metrics. The training set was appropriately stratified based on the subject sample to reflect outcomes that may occur in a realistic setting where more subjects without PH are likelier than those with PH. The observed variability in the cross-validation metrics may therefore stem from two sources. Firstly, the test sample size was small and due to this, even a few misclassifications could affect the performance metrics detrimentally. Secondly and perhaps more importantly, the screening is affected by how well differentiated the groups are, based on the similarity index. Because these infants all have underlying BPD, the cardiorespiratory patterns they exhibit may be similar despite a subset of them also having PH. During the observed 6 hour period, if some BPD-PH infants had similar patterns of oxygenation to that of a BPD infant (or vice-versa), their similarity index may overlap the BPD group and cause them to be misclassified. This is a challenge in developing efficient screening models for this condition that can only be overcome by including large sets of subject data accounting for all sources of variability in cardiorespiratory behavior and identifying signal measures that can quantify subtle differences between groups reliably. Despite this, the lowest validation metrics corresponding to SI_m7 and SI_m8 were over 70%, displaying good sensitivity and PPV balanced with specificity, all important metrics from a clinical perspective.

3.2. Current methods for PH diagnosis and strengths of SpO₂ based PH screening

A serious cardiovascular complication in preterm infants, PH is associated with higher risk of mortality, non-optimal growth and developmental outcomes, especially in those with BPD [3-5]. Diagnostic echocardiogram in infants with moderate to severe BPD at a postnatal age of 36 weeks’ postmenstrual age is the current recommendation for PH screening [29]. However, serial electrocardiography even when evaluated by trained neonatologists has limited sensitivity to detect the disease and its severity [30] . Risk factors for this condition suggest early detection of PH recommended by AHA consensus [3,7] . Screening models based on clinical characteristics and physiological signals which provide early and additional predictions of PH presence, may therefore support development of preventive protocols.

A recent model with demographic and known clinical risk factors such as small for gestational age and BPD severity grade was found to predict PH with AUROC of 0.93 [31] . Degree of respiratory support over the first five days of life was used to predict development of future PH with AUROC of 0.76 [32]. We studied distribution characteristics of severe desaturations in a case-control sample and found significantly longer durations per event in the PH group. The odds ratio failed to reach significance in detecting PH after adjustment for baseline and clinical characteristics, though mortality was well predicted within the PH group (AUROC = 0.77 ) [12]. These findings prompted us to explore indices of oxygenation dynamics that can identify characteristics of PH-compromised cardio-vascular functioning in the affected group. Similarity analysis of short term SpO₂ fluctuations as presented here is a novel methodology which can potentially be developed into a bedside screening protocol for PH. In addition, as it is based on portable, cost-effective pulse oximetry, this method can be easily adopted in low-resource settings with limited access to NICU hospitals [33].

3.3. Limitations and Future directions of the work

In this study, we focused on the similarity index as a marker of the differential oxygenation dynamics between the BPD and BPD-PH groups of infants with BPD. While this showed good performance, measures characterizing other aspects such as signal irregularity and spectral content may carry complementary information that can differentiate between groups. These could be investigated as potential signal based features along with the current measure, to identify infants with PH. Detecting cardio-respiratory diseases using physiological signals is challenging because disease markers may be spread over time or have signal patterns similar to the underlying condition, in this case BPD[10] . Hence, more effective signal analysis techniques and indices that extract information related to the pathology need to be investigated. An analogous situation arises in applying SpO₂ or heart rate variability analysis to distinguish severe from mild cases of OSA in different subject populations [14,34] . The methodology developed here should be validated on more subjects from different centers and its efficacy for early warning tested on data from earlier periods, preferably a few weeks before diagnostic testing. Our future work will pursue these directions as well as incorporate clinical characteristics to risk stratify pre-term infants likely to develop PH.

4. Conclusion

This work proposed information-based similarity analysis of routinely monitored oxygen saturation to screen preterm infants with BPD for PH. Quantifying the similarity in SpO₂ fluctuation patterns between adjacent segments from data recorded one week prior to the clinical diagnosis, this method outperformed other measures describing varied signal characteristics, in identifying PH in infants with BPD. The validation F1-score of screening with similarity indices was high, pointing to balanced sensitivity and positive predictive value, both important metrics in clinical screening. In summary, similarity analysis of oxygen saturation is a promising physiological signal-based early screening tool for PH, a complicated cardiovascular condition in preterm infants.

Highlights.

• Pulmonary hypertension (PH) complicates the clinical course and outcomes in preterm infants.

• Early screening for the condition can inform medical decision and neonatal care.

• Information based similarity analysis of oxygen saturation patterns identified infants with PH

• Our method using short data sets detected PH prior to diagnostic echocardiography

• Noninvasive signal-based screening can be a continuous marker of disease development

List of Abbreviations

PH

Pulmonary hypertension

BPD

Bronchopulmonary dysplasia

AHA

American Heart Association

SI_mi

Similarity index for a word of i bits

AUROC

Area under receiver operating curve

Acc.

Accuracy

Sen.

Sensitivity

Spe.

Specificity