Intermittent hypoxemia during hemodialysis: AI-based identification of arterial oxygen saturation saw-tooth pattern

Abstract

Background

Maintenance hemodialysis patients experience high morbidity and mortality, primarily from cardiovascular and infectious diseases. It was discovered recently that low arterial oxygen saturation (SaO₂) is associated with a pro-inflammatory phenotype and poor patient outcomes. Sleep apnea is highly prevalent in maintenance hemodialysis patients and may contribute to intradialytic hypoxemia. In sleep apnea, normal respiration patterns are disrupted by episodes of apnea because of either disturbed respiratory control (i.e., central sleep apnea) or upper airway obstruction (i.e., obstructive sleep apnea). Intermittent SaO₂ saw-tooth patterns are a hallmark of sleep apnea. Continuous intradialytic measurements of SaO₂ provide an opportunity to follow the temporal evolution of SaO₂ during hemodialysis. Using artificial intelligence, we aimed to automatically identify patients with repetitive episodes of intermittent SaO₂ saw-tooth patterns.

Methods

The analysis utilized intradialytic SaO₂ measurements by the Crit-Line device (Fresenius Medical Care, Waltham, MA). In patients with an arterio-venous fistula as vascular access, this FDA approved device records 150 SaO₂ measurements per second in the extracorporeal blood circuit of the hemodialysis system. The average SaO₂ of a 10-second segment is computed and streamed to the cloud. Periods comprising thirty 10-second segments (i.e., 300 s or five minutes) were independently adjudicated by two researchers for the presence or absence of SaO₂ saw-tooth pattern. We built one-dimensional convolutional neural networks (1D-CNN), a state-of-the-art deep learning method, for SaO₂ pattern classification and randomly assigned SaO₂ time series segments to either a training (80%) or a test (20%) set.

Results

We analyzed 4,075 consecutive 5-minute segments from 89 hemodialysis treatments in 22 hemodialysis patients. While 891 (21.9%) segments showed saw-tooth pattern, 3,184 (78.1%) did not. In the test data set, the rate of correct SaO₂ pattern classification was 96% with an area under the receiver operating curve of 0.995 (95% CI: 0.993 to 0.998).

Conclusion

Our 1D-CNN algorithm accurately classifies SaO₂ saw-tooth pattern. The SaO₂ pattern classification can be performed in real time during an ongoing hemodialysis treatment, provide timely alert in the event of respiratory instability or sleep apnea, and trigger further diagnostic and therapeutic interventions.

Background

Maintenance hemodialysis (HD) patients suffer from unacceptable high morbidity and mortality, with cardiovascular disease (CVD) being the main cause of hospitalization and death. The association between low arterial oxygen saturation (SaO₂) during dialysis and poor patient outcomes is well established [1, 2]. Physiologically, the transport of oxygen to peripheral tissues and organs depends on cardiac output, hemoglobin concentration, and saturation of hemoglobin molecules with oxygen. In HD patients, heart failure is a well-recognized cause of tissue hypoxia. Less appreciated is peripheral hypoxia due to structural and functional pathologies of the respiratory system, resulting in reduced SaO₂. Next to cardiac output and hemoglobin concentration, SaO₂ is a crucial determinant of the amount of oxygen delivered to tissues and organs. Clinically, SaO₂ can be assessed using both invasive and non-invasive methods. Recent advancements have introduced technologies that enable continuous monitoring of SaO₂ in the HD extracorporeal blood circuit during a treatment session. These innovations have proven valuable for identifying intradialytic hypoxemia [2], monitoring COVID-19 patients [3], and tracking other important clinical outcomes [1] such as respiratory instability [4]. Nocturnal hypoxemia is associated with cardiovascular complications in incident hemodialysis patients [5]. Prolonged intradialytic hypoxemia (PIH), defined as SaO₂ below 90% for more than one third of the HD treatment time, is associated with increased hospitalization and mortality rates. In addition, maintenance HD patients with PIH show a pro-inflammatory phenotype, as indicated by higher inflammatory markers, lower serum albumin levels, and resistance to erythropoietin. Furthermore, PIH is more prevalent in HD patients with comorbidities such as congestive heart failure (CHF) and chronic obstructive pulmonary disease (COPD) [1]. Intradialytic hypoxemia is also associated with peridialytic hypertension, possibly related to sympathetic activation and endothelin-1 secretion [6].

Recently, we observed distinct intermittent oscillatory saw-tooth pattern (STP) of SaO₂ in patients on HD [4]. In some instances, through detailed clinical observation, we were able to link STP to sleep-related breathing disturbances. This finding is significant, as sleep-disordered breathing is common in dialysis patients [7]. A study found that 37% of males and 34% of females with end-stage kidney disease (ESKD) treated with dialysis had sleep apnea syndrome (SAS). SAS was significantly associated with age, obesity, diabetes, hypertension, and other sleep disorders [8, 9]. These findings highlight the importance of monitoring SaO₂ during HD, as early identification of and intervention for intradialytic hypoxemia could mitigate its impact on long-term morbidity and mortality in this vulnerable population.

A recent study using recurrence analysis quantifies an optimal recurrence threshold (ε_opt) to identify STP [4]. Constructing ε_opt time series with a rolling window allows real-time detection of intermittent high-frequency high-amplitude STP. ε_opt correlates highly with SaO₂ desaturation density, and, using oxygen desaturation index as a surrogate for the apnea–hypopnea index, ε_opt accurately distinguishes SAS events.

One-dimensional convolutional neural networks (1D-CNN) have gained significant attention in the medical field, particularly for analyzing time series of biomedical signals such as electrocardiograms (ECG) and electroencephalograms (EEG). These networks can efficiently process time-series data and other one-dimensional formats without extensive pre-processing. Applications include patient-specific ECG classification and early diagnosis of diseases [10].

Quasi-continuous measurement of SaO₂ in the extra-corporeal circuit by the Crit-Line monitor (CLM) affords the opportunity to follow the temporal evolution of SaO₂ during HD. Against this backdrop, we aimed to automatically identify repetitive episodes of intermittent STP using 1D-CNNs.

Methods

In HD patients with an arterio-venous fistula as vascular access, the oxygen saturation in the extracorporeal HD circuit resembles SaO₂. SaO₂ during HD was measured using the CLM (Fresenius Medic Care North America, Waltham, MA), a device certified by the US Food and Drug Administration (FDA) for the measurement of hematocrit and oxygen saturation in the extracorporeal blood circuit of the HD system. CLM measures oxygen saturation and hematocrit 150 times per second. These measurements are then averaged over a 10-second period and the respective means transmitted to Amazon Web Services (AWS) via Apache KAFKA, a real time streaming software [11].

Patient selection

To build our AI model, we collected intradialytic SaO₂ recordings from HD patients. The inclusion criteria were (a) being an in-center maintenance hemodialysis patient, and (b) availability of CLM SaO₂ recordings. The patients were dialyzed in clinics of the Fresenius Kidney Care network that comprises about 2,500 clinics in the US. In about 600 of these clinics, CLM is used as a standard of care. There were no exclusion criteria. These rather wide inclusion criteria and the absence of exclusion criteria ensure that the base population is typical for a US HD population and that the results are generalizable. In these about 600 clinics, we have collected data from about 500,000 HD treatments. Out of these treatments we selected 89 treatments from 22 patients.

Visual inspection

We segmented the treatment time series into 5-minute consecutive sections, each segment had 30 SaO₂ measurements. The 5-minute segments were visually inspected and independently adjudicated by two researchers who categorized them as either no saw-tooth pattern (NSTP, stable and consistent SaO₂ with minimal fluctuations; Fig. 1A) or STP. NSTP indicates respiratory stability. In contrast, STP shows much greater variability and fluctuations with noticeable peaks and troughs over time; it indicates respiratory instability (Fig. 1B). The visual classification of SaO₂ patterns by the two independent observers was based on two characteristics: amplitude and duration. NSTP were considered SaO₂ variations with an amplitude of less than 2%, with only occasional isolated spikes exceeding 5% in amplitude and lasting less than 30 s. In contrast, STP were considered by oscillations in SaO₂ with an amplitude greater than 5% and a duration exceeding 30 s. Severe STP had an SaO₂ < 90%. The reason for the < 90% SaO₂ criterion was that earlier work showed that intradialytic SaO₂ < 90% is associated with increased morbidity and mortality [1]. Discrepancies between the two observers were resolved through a joint review.

Fig. 1

Two examples of SaO₂ time series. Panel (A) shows an SaO₂ time series without saw-tooth pattern. Panel (B) shows a SaO₂ saw-tooth pattern with about one cycle per minute

Training

We randomly allocated the 5-minute SaO₂ time series segments to either training (80%) or test (20%) data sets. Initially, most segments in the training data were NSTP. To address this class imbalance in the training data, we up-sampled the STP group to match the number of NSTP samples. This was done by randomly selecting STP segments with replacement from the original pool and replacing the original STP samples in the training dataset with the newly selected ones. The test data set was not changed. The updated training dataset was then shuffled and randomly partitioned into two subsets. To improve model generalization, random Gaussian noise (mean = 0, standard deviation = 0.1) was added to every SaO₂ measurement in one subset, which was then used to train the 1D-CNN. The other subset served as the validation set for computing the validation error throughout the optimization process.

For the learning process we used the Keras open-source deep learning library [12] to build a 1D-CNN. The input layer of our 1D-CNN receives the 5-minute time-series segments of SaO₂. The first convolutional layer applies 32 filters, each with a kernel size of 6, to extract low-level temporal patterns from the input sequences. A Batch Normalization layer follows the convolution to normalize activations and accelerate training. A Max Pooling layer with a pool size of 3 and a stride of 2 is applied to reduce spatial dimensions while retaining the most important features. A second convolutional layer is applied to the pooled output to refine feature extraction. This layer has 32 filters with a kernel size of 3. Batch Normalization is again used to improve stability during training. Another Max Pooling layer with a pool size of 2 and a stride of 2 is applied. Following the convolutional and pooling layers, the output is flattened to a 1D vector to be fed into the fully connected layers. The fully connected layer with 32 neurons is applied to capture deeper representations of the extracted features. Finally, the model outputs a probability distribution over two classes using a softmax activation function, providing classification probabilities for NSTP and STP, respectively (Fig. 2).

Fig. 2

Architecture of a 1-dimensional convolutional neural network (1D-CNN). In this example, a STP is presented to the input layer. The trained 1D-CNN computes the probabilities for NSTP and STP, respectively

The model was trained using the Adam optimizer from the Keras 3.5 library with a categorical cross-entropy loss function [13]. Training was conducted for 1,500 epochs with a batch size of 32 and a learning rate of 0.000003.

Test/validation

The model with the smallest validation error was chosen as the final model. This final model was then evaluated on the test set using the following metrics: accuracy, sensitivity, specificity, F₁-score and area under the receiver operating curve (AUC).

Results

We sampled 89 HD treatments with SaO₂ measurements from 22 in-center patients across four different clinics (Table 1). Continuous variables are reported as mean (± standard deviation) or median [Q1, Q3] and categorical variables are presented as proportions. In total, 4,075 consecutive 5-minute segments were adjudicated, 21.9% with STP and 78.1% with NSTP. The training dataset comprised 3,260 segments, while the test set contained 815 segments. Within the training set, NSTP and STP segments were distributed as 2,542 (78.0%) and 718 (22.0%), respectively. In the test set, these distributions were 642 (78.8%) for NSTP and 173 (21.2%) for STP (Table 2).

Table 1 Patient and treatment characteristics

Table 2 Rate of saw-tooth pattern and no saw-tooth pattern among 5-minute segments in the training and test data, respectively

Our 1D-CNN model reached a 96.4% accuracy and an AUC of 0.995 (95% CI: 0.993 to 0.998) in the test set (Fig. 3). The respective confusion matrix is shown in Table 3. In the test data, the model reached 96.5% sensitivity, 96.4% specificity and 0.92 F₁-score, indicating that it can work as a reliable method to detect STP.

Fig. 3

Receiver operating curve and AUC on the test data set

Table 3 Confusion matrix of the classification performance in the test data set

For external validation, we randomly selected 20 recent HD treatments from 18 clinics. For this sampling, we did not perform the up-sampling of STPs to ensure a more accurate representation of their actual prevalence. Within this external validation dataset, the model achieved a sensitivity of 95.1%, a specificity of 96.7%, and an F1-score of 0.79 (Table 4).

Table 4 Confusion matrix of the classification performance in the external validation data set

Figure 4 shows the SaO₂ measurements of two different HD treatments. Both sessions were about four hours long. In Fig. 4A, NSTP was classified, while for the session shown in Fig. 4B the model classified STP for many of the 5-minute segments (shown in orange).

Fig. 4

Intradialytic SaO₂ measurements during two hemodialysis sessions. (A) Session without saw-tooth pattern classified. (B) Session with saw-tooth patterns (orange)

To distinguish presumably benign from more concerning STP, we decided to further categorize STP in mild and severe STP. If a 5-minute segment was classified as STP and the minimum SaO₂ in that segment was below 90% then we classified it as severe STP. If a 5-minute segment was classified as STP and all SaO₂ measurements were equal or above 90% then we classified it as mild STP. Figure 5 shows the same HD treatment as Fig. 4B after separating mild and severe STP.

Fig. 5

Intradialytic SaO₂ measurements during a hemodialysis session with intermittent mild (orange) and severe (red) saw-tooth patterns

Discussion

Our primary finding demonstrates that a 1D-CNN can classify quasi-continuous intradialytic SaO₂ recordings with high accuracy into STP and NSTP, respectively. The method can be scaled easily, enabling efficient classification of a large volume of SaO₂ measurements. Given the prevalence and clinical significance of hypoxemia in HD patients, this result may hold important clinical implications.

While adequate SaO₂ is fundamentally important for health and quality of life, there is a notable knowledge gap about SaO₂ dynamics during HD. To address this issue, dialysis providers have implemented CLM devices in about 600 dialysis clinics across the US, many of which are connected to the cloud for seamless, quasi-continuous data gathering. Recently, we collected over six billion intradialytic SaO₂ measurements from over 70,000 HD patients [14]. The analysis of such a vast data repository necessitates automated methods, which led us to develop an AI-based tool, a 1D-CNN. This model accurately classifies STP in SaO₂ time series recorded during HD. We plan to deploy the model in the cloud, focusing on quantifying the SaO₂ pattern at both the treatment and patient levels. This approach aims to enhance our understanding of intradialytic SaO₂ dynamics and their clinical correlates.

SaO₂ is a critical parameter in the management of patients undergoing HD, as it provides insights into the patient’s respiratory and circulatory status. In HD patients, maintaining adequate SaO₂ is essential due to the unique physiological challenges posed by HD. Studies have shown that a significant proportion of hemodialysis patients experience episodes of hypoxia, with some reporting SaO₂ levels dropping below 90% during treatment [1]. SaO₂ serves as a prognostic indicator, with studies demonstrating that lower arterial oxygen saturation levels during HD are associated with poorer clinical outcomes, including higher rates of hospitalization and mortality [1]. In addition, low SaO₂ during HD has been associated with intradialytic hypertension, possibly due to activation of the sympathetic nervous system activation, stimulation of the renin-angiotensin-aldosterone system (RAAS), and endothelin-1 release [6]. During the Covid pandemic, low SaO₂ before the diagnosis of Covid-19 identified patients at high risk for hospitalization and death [3].

However, to date, the relationship between SaO₂ STP and outcomes in HD patients is unknown. The reasons for this knowledge gap are multifaceted. First, it was only a few years ago that cloud-connected CLM devices were deployed, enabling the necessary data storage. Second, these newer CLM devices report SaO₂ every 10 s, a significant improvement over the previous devices, which reported only once per minute. Nevertheless, it is conceivable that the ability to classify SaO₂ STP could inform clinical decision-making or diagnostic and therapeutic interventions. For example, SaO₂ STP are associated with SAS. This pattern typically reflects the intermittent nature of oxygen desaturation events that occur during apneic episodes. Galuzio et al. [4] highlight that oximetry data in HD patients can reveal intermittent oxygen desaturations that correlate with the apnea–hypopnea Index (AHI), a key metric used to assess the severity of sleep apnea. Furthermore, the relationship between SaO₂ and SAS is underscored by findings from Jhamb et al. [15], who reported that lower mean oxygen saturation is associated with increased mortality risk in patients with advanced kidney disease, many of whom may also suffer from undiagnosed SAS. In addition, SAS in HD patients is associated with inflammation and oxidative stress [16]. In HD patients with SaO₂ STP, specific clinical exploration and testing for SAS could be considered. If confirmed, e.g., through polysomnography, appropriate treatment could be initiated. This is relevant, as sleep-disordered breathing is frequent in dialysis patients [7]. The reported prevalence of sleep apnea is > 50% among dialysis patients [17,18,19]. SAS is a condition that is associated with lower health-related quality of life [20] and increased cardiovascular and mortality risk in kidney patients [15, 21].

Other methods have been developed to identify SaO₂ STP. Galuzio et al. [4] devised a method that builds on dynamical systems theory and not on AI. The authors have developed a metric to quantify the occurrence of abnormal behaviors in the time series. This approach achieves an area under the ROC curve between 0.93 and 0.94 to identify SaO₂ pattern indicative of SAS. It will be interesting to see the relative performance of the two approaches in a well-defined patient cohort with a clinical diagnosis of SAS.

When implemented at scale, the 1D-CNN will allow us to identify SaO₂ STP and to investigate clinical correlates of the SaO₂ STP, including demographic factors, laboratory data as well as treatment-related parameters. Additionally, one can examine the relationship between quantifiable measures of the SaO₂ STP and key clinical outcomes, including hospitalization and mortality. This correlation seeks to provide deeper insights into the clinical significance of these patterns.

The quantification of STP at treatment level and patient level could also be used as feature inputs for other predictive models, e.g. hospitalization predictive model. With current cloud setting, this saw-tooth pattern recognition could be performed in real-time during the dialysis treatment and trigger a more specific diagnostic evaluation (e.g., polysomnography for definitive diagnosis of SAS) and therapeutic interventions.

Conclusions

In summary, we have developed an AI-based tool capable of detecting irregularities in SaO₂ time series collected during hemodialysis. This tool will deepen our understanding of intradialytic SaO₂ dynamics and support the implementation of advanced diagnostic and therapeutic interventions.