Predicting Survival in Extracorporeal Membrane Oxygenation Patients With Optical Microcirculation Sensing

In this study, we propose a novel method to use functional near-infrared spectroscopy (NIRS) to monitor patients’ lower limb microcirculation with extracorporeal membrane oxygenation (ECMO). We controlled the ECMO system's speed and measured hemodynamics using NIRS devices which attached to both calves at approximately 60% of the tibia length. Features from the collected blood oxygen data were extracted and utilized as machine learning inputs for classification. The patients were divided into two groups based on discharge and mortality. In venovenous (VV) ECMO, we found that the construction of the classification model based on the characteristics of this type with better discriminating ability can effectively distinguish the two groups.

terminal part of the circulatory system, and it is also the place where the material is exchanged between blood and tissue cells. Clinically, patients with acute and severe diseases are often unable to provide sufficient oxygen to maintain the metabolic needs of tissues in the body due to impaired cardiovascular function and thus causing shock. During shock, the peripheral vascular system will contract to ensure blood supply to the heart and brain, causing various pathological changes and even perfusion failure between body organs and tissue cells due to ischemia and hypoxia. Therefore, monitoring peripheral microcirculation changes in acute and critical patients helps clinicians evaluate and judge patients' physiological conditions [1], [2], [3], [4].
Extracorporeal membrane oxygenation (ECMO) is a lifesaving rescue treatment in refractory respiratory and cardiac failure [5], [6], [7], [8], [9], [10]. It can be divided into the following two basic types: venovenous (VV) and venoarterial (VA)-ECMO [11]. VV-ECMO removes blood from the inferior vena cava, passes it through the ECMO device, and returns it to the body via the femoral or jugular vein. It can replace the respiratory function of the lungs. Suppose the patient's heart function is normal, but there is a problem with lung function. In that case, this device is suitable for preventing the complications caused by arterial intubation [12], [13]; VA-ECMO removes blood from the inferior vena cava, passes it through the ECMO device, and returns it to the body via the femoral artery. It can provide heart pumping and respiratory function support. When patients experience sepsis or hypotension because of other factors, they will undergo treatment with this type of device. According to the data from the annual international Extracorporeal Life Support Organization (ELSO) Registry Reports through July 2020, a total of 57958 adult patients have received the ECMO device since 1990, of which approximately 60% of the patients were removed from the aid of the ECMO device, and 48.8% of patients survived and were discharged successfully [14]. Although ECMO has been effective in improving the survival rate of patients [15], there are still considerable risks and complications that are difficult to resolve after its use [16], [17], [18]. Currently, there is no reliable and immediate mechanism for detecting hemodynamic changes in the peripheral tissues. We can merely use basic physiological parameters to adjust the ECMO setting. Therefore, monitoring peripheral micro-circulation changes in acute and critical patients helps clinicians evaluate and judge patients' physiological status.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0 /  TABLE I  INFORMATION ON PATIENTS TREATED WITH ECMO IN THIS STUDY NIRS is a noninvasive optical method to evaluate regional oxygenation of the tissue circualation. Our preliminary study discussed patients' physiological status while undergoing ECMO and optimized ECMO therapy based on real-time peripheral NIRS probing [11]. This study continued the previous research methods to monitor hemoperfusion of patients' lower limb microcirculation. A feasibility study was conducted to evaluate the disease condition of patients receiving ECMO by combining NIRS data with a machine learning algorithm. In the future, we aim to develop an optical analysis system that assists in ECMO treatment, makes an objective assessment of prognosis and provides clinicians with more treatment information.

II. MATERIALS AND METHODS
The Institutional Review Board of Taipei Veterans General Hospital approved this study (TVGHIRB-2019-02-007AC). This approval is valid for two years, from Feb 12, 2019, until Feb 11, 2021. The study was conducted with adult patients (age >20 years) undergoing ECMO at Taipei Veterans General Hospital.

A. Patients
Patients undergoing VA-ECMO and VV-ECMO were included in this study. Patients in the following conditions were excluded from the study: (1) patients receiving ECMO with central cannulation; (2) patients for whom limbs were unsuitable for distal NIRS oximetry monitoring. Forty-four patients participated in this study, of which twenty-two patients were receiving VA-ECMO, and twenty-two patients were receiving VV-ECMO (Table I).

B. NIRS Measurement and Analysis
Hemodynamic responses were measured using two PortaLite systems (Artinis, Netherlands) at a distance from the ankle of approximately 60% of the tibia length ( Fig. 1(a)). The instrument has three pairs of light sources and one photodetector. The distances between the sources and the detector are 30 mm, 35 mm, and 40 mm. The three light sources emit near-infrared rays at both wavelengths (760 nm and 850 nm). The light intensity data measured by the probe at a sampling rate of 25 Hz is transmitted to a computer via Bluetooth, and converted into the relative concentration changes of oxyhemoglobin (HbO 2 , µ· − M minute −1 ) and deoxyhemoglobin (HHb, µ· − M · − minute −1 ), using PortaLite's dedicated software Oxysoft (Artinis, Netherlands). To determine absolute hemoglobin concentrations to estimate the tissue saturation index (TSI %), PortaLite employs spatially resolved spectroscopy (SRS). Subsequently, the blood oxygen information is filtered and normalized, and then brought into our machine learning algorithm.

C. Measurement Protocol
At initiation, the ECMO flow rate was kept for 15 minutes (min) as the reference value to ensure the stability of the signal. In the VA-ECMO group, the flow rate was reduced by 500 revolutions per minute (rpm) for 10 min and returned to the initial flow rate for the next 10 min. Subsequently, the flow rate was increased by 500 rpm for 10 min and then by another 500 rpm for the next 10 min. Finally, the flow rate was returned to the original rate for 15 min ( Fig. 1(b)). VA-ECMO needs to pump low-pressure venous blood to high-pressure arteries, so the adjusting pump speed unit in VA-ECMO is higher. In the VV-ECMO group, the adjustment protocol was the same as that for the VA-ECMO group, but each step made adjustments of 300 rpm. The adjustment criteria for determining the rotational speed of ECMO in the protocol are determined by the clinician's experience. Makes changes in blood flow large enough to be observed without compromising patient safety. The PortaLite system monitored the patient for 70 min throughout the whole experiment.

D. Microcirculation Monitoring Results
Regardless of VA-ECMO or VV-ECMO, there is usually a cannula through the femoral vein or femoral artery on one side of the patient's leg in the ECMO device, so the lower limb of the patient could be divided into the cannulation side and the noncannulation side. The cannulation side greatly influences peripheral tissue hemoperfusion. Therefore, this study mainly discussed the changes in blood oxygen on the noncannulation side of the subjects.

E. Statistical Analysis
There are many periodic physiological signals in the raw data measured by NIRS, and these fixed-frequency physiological signals are the leading cause of the noise, including respiration (approximately 0.15-0.4 Hz) and heartbeat (approximately 0.4-1.6 Hz) [19]. We used an exponential moving average (EMA) to filter the signal to reduce the noise impact on the data analysis.
From the hemodynamic signal, the average values of HbO 2 , HHb and TSI in each stage, stage slope, and stage beginning slope, and also calculate the activation degree and the stage difference for HbO 2 and HHb at each stage ( Fig. 1(c)). These values were used as the classification model's characteristics, and the classification effectiveness was evaluated to select classification features with better discriminative ability.  high-dimensional or infinite-dimensional space for classification and other tasks [21], [22], [23]. We use the radial basis function (RBF) as the model's nonlinear decision boundary's transformation function. The function can transform the input features into a high-dimensional feature space and solve some problems that linear functions cannot distinguish. The confusion matrix and k-fold cross-validation were used to test the model's classification performance and generalization ability.

A. Statistics Between VA and VV Groups
We classified as a discharge group for patients who successfully left the hospital after weaning from ECMO. Patients who could not wean from ECMO or passed away after weaning from ECMO were classified as the mortality group. In the VA group, their mean age was 56 years, and eight of them were female. In the VV group, ten were female, and their mean age was 60 years. The mortality rates were 72.73% (6 discharge/16 mortality) and 43.48% (12 discharge/10 mortality) in the VA and VV groups, respectively. Fig. 2 shows the data distribution in every stage. The adjusting pump speed unit in VA-ECMO is higher than that in VV-ECMO in the experimental setting because VA-ECMO needs to pump low-pressure venous blood to high-pressure arteries, requiring a higher pump speed. Therefore, HbO 2 levels were higher in the  VA-ECMO group than in the VV-ECMO group. From the box plot of the changes in HbO 2 and HHb concentrations of the two groups, the distribution of data in the VA group was relatively dispersed.

B. Feature Comparison and Extraction
After further analysis, we extracted 78 characteristic parameters from the NIRS signal on the noncannulation side of the VV-ECMO and VA-ECMO groups to classify the discharge and mortality groups. However, excessive input of useless feature parameters into the classification model will lead to overfitting the classification results. This research used an independent sample t test as a performance measure for machine learning algorithms. Table II lists the characteristic parameters selected from the VV-ECMO and VA-ECMO groups with better p values after comparison and ranking. Fig. 3 illustrates the two sets of chosen parameters from the VV-ECMO and VA-ECMO groups and shows the distribution of these two sets of characteristic parameters in the discharge and mortality groups. We observed that the two features  III  AUC VALUES OF THE CHARACTERISTICS IN THE VV-ECMO AND  VA-ECMO GROUPS with better p values in the VV-ECMO group were the stage difference of ΔHbO 2 in stage 1 and the beginning slope of ΔHbO 2 in stage 5, both of which had statistically significant differences. In contrast, in the VA-ECMO group, the mean ΔHbO 2 in stage 1 and the mean ΔHbO 2 in stage 2 were extracted. However, the p values of the above characteristics were 0.0788 and 0.062, respectively, showing no statistically significant differences. Therefore, we could not prove that these two groups of characteristic parameters could be used as a reasonable basis for classification. Thus, we referred to Gottemukkula et al.'s research [24]. The team used receiver operating characteristic (ROC) curves and area under curve (AUC) values to rank NIRS signal features and selected the feature parameters with better performance to incorporate into the classification model. Table III shows the AUC values for the characteristics of the VV-ECMO and VA-ECMO groups. ROC is a probability curve, and AUC represents the degree or measure of separability. It tells how much the model is capable of distinguishing between classes. Therefore, the higher the AUC value, the more influential the evaluation features for the classification model [25], [26]. Table III showed that the two features, which were stage differences in ΔHbO 2 in stage 1 and stage beginning slope of ΔHbO 2 in stage 5, had better p values previously obtained in the VV-ECMO group. It still had a good performance in the ROC curve and the AUC value evaluation, with AUC values of 0.8148 and 0.7778, respectively ( Fig. 4(a) and (b)), which had excellent discrimination. We extracted the stage mean of ΔHbO 2 in stage 1 and the stage difference for TSI % in stage 5 in the VA-ECMO group (Table III). Although both AUC features were 0.74, which indicated acceptable discrimination, the ROC curve's performance was abysmal (Fig. 4(c) and (d)).

C. SVM Classification Results
In this study, the VV-ECMO and VA-ECMO groups were included 22 subjects, respectively, and 70% of the data were used as a training set, and 30% as a testing set. After selecting the characteristic parameters with better discriminating ability, we built the SVM model to predict the outcome. Fig. 5(a) and (b) shows the data distribution and model classification results of the VV-ECMO group classified by SVM under  two-dimensional features. The blue dots represent the discharge group's data, and the red dots represent the data points of the mortality group. The accuracy was 93.3% for the training set and 71.4% for the testing set. Fig. 5(c) and (d) shows the data distribution and model classification results of the VA-ECMO group classified by SVM under two-dimensional features. In the same way as the VV-ECMO classification model mentioned above, the blue dots represent the discharge group's data points, and the red dots represent the data points of the mortality group. The accuracy was 86.7% for the training set and 57.1% for the testing set.

D. Result Validation
To prevent the deviation caused by the classification model's dependence on a specific training set or testing set, we need to test the classification performance and generalization ability of the two ECMO classification models using a confusion matrix and 5-fold cross-validation.
In the confusion matrix, the positive samples of the VV-ECMO testing set data label were three mortality patients, and the negative samples were four discharge patients. The matrix values were normalized to facilitate the observation of sensitivity and specificity results ( Fig. 6(a)). The sensitivity and specificity of the proposed VV-ECMO classification model were 67% and 75%, respectively. The higher the two values are, the better the model's recognition for positive and negative samples. The five samplings' average accuracy and standard deviation in the 5-fold cross-validation were 0.78 ± 0.13 (Fig. 6(c)), indicating that the model had relatively good generalization ability and stability.
In the VA-ECMO classification model's confusion matrix, the positive samples of the testing set label were five mortality patients, and the negative samples were two discharge patients. The values of the matrix were also normalized. We observed that the VA-ECMO classification model's sensitivity and specificity were only 60% and 50%, respectively ( Fig. 6(b)), while the average accuracy and standard deviation of 5-fold cross-validation were 0.63 ± 0.19 (Fig. 6(d)). The above model evaluation results indicate that the VA-ECMO classification model is relatively low in generalization ability and stability.

IV. DISCUSSION
In the VV-ECMO classification model, we evaluated the feature's performance by the above method. We selected two features with better discriminative ability, namely, the stage difference of ΔHbO 2 in stage 1 and the beginning slope of ΔHbO 2 in stage 5. The two features showed a great difference between the two groups in VV-EMCO because the time point in which these two features occur is when the microcirculation in the patient's body is more severely affected by the adjustment of the pump speed. When the patient has reached the final stage of physiological condition, the body's maintenance mechanism will concentrate blood to the body's core to ensure the functioning of vital organs such as the brain and liver. For the discharge group, the blood oxygen changes larger when affected by the pump speed adjustment; for the death group, the blood oxygen changes were small, and the measured characteristics of the two were significantly different. Thus, blood vessels such as peripheral arterioles, the peripheral circulatory system, and the intestinal system will constrict or close. Therefore, no matter how adjustments were made to the pump speed in ECMO, the hemoperfusion changes we measured were not significant. On the other hand, there are significant changes in blood oxygenation in the adjustment of the ECMO pump speed, which may indicate that the patient's peripheral regions can obtain sufficient oxygen for gas exchange between the cells and capillaries, help the damaged tissue to repair, and reduce the risk of complications. From the model's decision boundary ( Fig. 5(a) and (b)), the two ethnic data points of VV-ECMO could be roughly separated, and there was no overfitting. Although it was not a very good classification result according to the accuracy of approximately 70% of the testing set classification, we believed that the classification error of the model at any data point would significantly affect the final result due to the insufficient amount of data collected at present. A small number of incorrectly classified data points appeared if we discussed the data distribution in the two-dimensional classification and the model's circle selection. Nevertheless, these data points are distributed near the model's decision boundary and have no extreme outliers. The normalized confusion matrix and 5-fold cross-validation show that the VV-ECMO classification model we built has relatively good classification generalization ability and model stability.
In the VA-ECMO classification model, the independent sample t test evaluation from the two groups in VA-ECMO failed to obtain statistically significant differences in characteristic values. Therefore, we referred to the relevant literature's feature selection methods and tried to use ROC curves and AUC values for feature evaluation. We selected two characteristic parameters with better AUC value performance, which were the stage mean of ΔHbO 2 in stage 1 and the stage mean of TSI % in stage 5. The model's decision boundary circled most of the discharge group's data points without overfitting (Fig. 5(c) and (d)). Nevertheless, several misclassified data points existed in the upper left corner of the training model. Even though the training set's classification model was approximately 90% accurate, we only obtained less than 60% accuracy when importing the testing data into the trained model. The results indicated that the classification model did not have good generalization ability in the VA-ECMO group classification. We believe that the reason for this result was that, on the one hand, the number of discharge populations of VA-ECMO collected in this study was insufficient. There was no statistically representative data distribution of the maternal population and no noticeable difference in distinguishing data, which led to the failure of proper fitting of the model. On the other hand, the poor performance of VA-ECMO classification results may be due to the complexity of the disease types of patients receiving this type of ECMO (Fig. 6(e)). For example, a patient suffering from shock due to severe bacterial infection and another patient suffering from myocarditis without bacterial infection received the same ECMO type. Patients' pathological changes and physiological responses were very different, so it was difficult to explain whether the blood oxygen changes measured in the same population were consistent. The normalized confusion matrix and 5-fold cross-validation results indicated that this model's classification performance and generalization ability did not meet the requirements. Therefore, it was impossible to distinguish the two groups of VA-ECMO from the present data.
This research work is a preliminary achievement of precision medicine of ECMO treatment. The ultimate goal of precision medicine is to minimize iatrogenic damage and minimize medical resource consumption to maximize the benefits of treatment. This research hopes to collect more data in the future, further refine the model and improve its performance, and predict whether the patient can be discharged from the hospital, which will provide a reference for clinicians. For patients who are predicted to be discharged from the hospital, better care should be given according to the actual situation; for patients who are predicted to death, it can also be considered whether to give up ineffective medical treatment.

V. CONCLUSION
This study is a preliminary feasibility study of the mechanism for evaluating patients' clinical conditions with ECMO. In VV-ECMO, we found that the construction of the classification model based on the characteristics of this type with better discriminating ability can effectively distinguish the two groups. The inspection of the classification model also shows that it has relatively good classification generalization ability and model stability. Regarding VA-ECMO, the complexity of the disease types and the insufficient amount of data did not lead to better results. Additionally, due to the classification accuracy, the classification model could only be used to make a preliminary assessment. More than 66.7% of patients are currently clinically equipped with VV-ECMO, indicating that the application of NIRS combined with machine learning has considerable potential. In the future, we hope to continue collecting more subjects to expand the ECMO patient database and improve the accuracy and generalization ability of the classification prediction model. After the model performance improves, more appropriate medical resources can be given to patients predicted to be discharged from the hospital, and unnecessary treatment can be reduced for patients predicted to death. We hope that this assessment tool will provide clinicians with a more accurate prediction of critically ill patients' conditions and effectively reduce the problems of medical resource abuse and ineffective care.