Oximeter-Based Autonomic State Indicator Algorithm for Cardiovascular Risk Assessment FREE TO VIEW

Cardiovascular (CV) disease is the leading cause of death with contributing risk factors including age, gender, smoking, dyslipidemia, diabetes, and hypertension.¹ Diagnostic and therapeutic decisions are often based on risk assumptions such as the Framingham Risk Score² and the European Systematic Coronary Risk Evaluation,³ which are derived from large, long-term outcome studies and have been validated prospectively.^4,5

Although undisturbed sleep is associated with a marked unloading of the CV system, physiologic parameters derived during sleep have not been used traditionally for CV risk assessment. However, some sleep-related respiratory and CV parameters have been shown to be strongly related to CV morbidity and mortality. For instance, repetitive hypoxemia and sympathetic activation associated with obstructive sleep apnea (OSA) events promote cardiac and vascular disease.^6,7 The nondipping of both BP⁸ and heart rate⁹ during sleep are independently associated with all-cause mortality. The high-frequency component of heart-rate variability during sleep is blunted in patients with coronary artery disease.¹⁰ Nocturnal arterial vascular tone determined by finger photoplethysmography is elevated in patients with essential hypertension.¹¹ Therefore, it seems plausible that systematic combined analysis of cardiac and vascular reactivity recorded during sleep may provide novel and improved tools for CV risk assessment.

Finger pulse oximetry, a noninvasive photometric technique that measures the pulsatile nature of the arterial blood flow, is widely used for the diagnosis of sleep-disordered breathing. Multiple physiologic parameters, including oxyhemoglobin saturation, heart rate, and pulse waveform, can be derived from the oximetry signal. We hypothesized that these physiologic parameters reveal information about CV characteristics that reflect CV risk. Overnight pulse wave and oximetry parameters were collected from patients referred to a sleep center for diagnostic evaluation. An autonomic state indicator (ASI) algorithm was developed to categorize CV risk and was validated using the well-established European Society of Hypertension/European Society of Cardiology (ESH/ESC) risk-prediction matrix.¹²

A total of 213 subjects were randomly selected from clinical patients referred to the sleep laboratory at Sahlgrenska University Hospital, Gothenburg, Sweden, during the period of 2004 to 2006. No specific selection criteria were applied, and recruitment was determined only by current access to personnel and study-specific equipment. The study was approved by the internal review boards, and the patients were informed about the study.

A modified version of the Basic Nordic Sleep Questionnaire¹³ covering the occurrence of all major CV diseases was completed by the subjects. In addition, BMI and office BP were assessed according to World Health Organization standards. The presence of CV risk factors and/or concomitant CV disease was further determined by a sleep physician during a clinical interview and physical examination. Finally, each patient’s comprehensive hospital medical record was reviewed for diagnoses of CV disease. Drug treatments were assessed and coded according to the Anatomical Therapeutic Chemical classification system. All this information was compiled and used to determine CV risk scores in accordance with the ESH/ESC risk factor matrix, ranging from “average risk” (CV risk category 1), “low added risk” (CV risk category 2), “moderate added risk” (CV risk category 3), “high added risk” (CV risk category 4), and “very high added risk” (CV risk category 5).¹² This matrix depicts an overall estimation of the total CV risk in European populations and is calibrated to Framingham criteria and the Systematic Coronary Risk Evaluation chart.

Patients underwent either attended or ambulatory overnight polysomnography (PSG) (Embla A10; Reykjavik, Iceland) (n = 133) or polygraphy (Somnocheck II; Weinmann; Hamburg, Germany) (n = 80). All recordings included thoracic and abdominal respiratory effort signals, nasal-oral airflow pressure, and body position. The PSG recordings included EEG channels C4/A1, Cz /A1, and C3/A2; left and right electrooculograms; chin and anterior tibialis muscle electromyograms; and ECG.

All the recordings were scored manually according to the American Academy of Sleep Medicine criteria.¹⁴ Obstructive apnea was defined as a ≥ 90% decrease in airflow, with evident respiratory effort, lasting at least 10 s. Hypopnea was defined as a clear reduction of at least 50% airflow amplitude associated with an oxygen desaturation (Spo₂) of ≥ 3% or an arousal (in PSG recordings). The apnea hypopnea index was calculated as the total number of apnea and hypopnea events per hour of total sleep time (or total analysis time in polygraphy recordings). The Spo₂ index was calculated as the number of Spo₂ events of at least 4% per hour of total sleep time (or total analysis time).

A novel finger photoplethysmographic pulse oximeter sensor (ChipOx; MCC; Karlsruhe, Germany) was developed to quantify the pulse wave signal. These pulse signals were recorded on a separate analog data logger (Physiologg data logger; Institute for Aviation and Space Research, University of Cologne; Cologne, Germany) in parallel with PSG recordings. During polygraphy recordings, the signals were directly recorded by the inbuilt oximetry module.

Recordings with a minimum 3-h artifact-free oximeter signal were included in the ASI analysis. Sixty-five patients were discarded because of technical failure and/or insufficient signal quality, which occurred mainly during the start-up phase of the study. Final analysis was performed in 148 patients (Fig 1).

The photoplethysmographic pulse wave signal was first collected with a sampling frequency of 50 Hz and was filtered using a second-order Butterworth low-pass filter with cutoff frequency of 10 Hz to remove noise. Five physiologic parameters, including pulse wave attenuation (PWA), pulse propagation time (PPT), respiration-related pulse oscillation (RRPO), pulse rate (PR), and Spo₂, were derived from the beat-to-beat signal without averaging and were saved at a sampling rate of 5 Hz. Time and frequency information from PWA, PR, and Spo₂ were extracted using a Matching Pursuit algorithm,¹⁵ which is an established, wavelet-related signal analysis method. Indices of the prevalence of PWA, PR, and Spo₂ were calculated from the total recording time. These three indices and the whole-night means of PPT and RRPO (five ASI components) were applied to the development of the CV risk classification algorithm (see following for detailed information).

For the identification of high-risk patients, a neuro-fuzzy system¹⁶ (see e-Appendix 1 for detailed information) was constructed and trained using a randomly selected subset of patients (n = 99). A three-layer fuzzy perceptron with one hidden layer was used. The perceptron contained five input nodes from the five ASI components, a four-node hidden layer, and one output layer, which determined the probability that the input vector belonged to a high-risk patient (Fig 2). For the fuzzification process, trapezoid membership functions were used. The fulfillment grade of each data vector to belong to a high-risk patient was computed by using min/max inference and was scaled to the output range of [0;1]. The final classification was done by testing the output value of the neuro-fuzzy system with a threshold of 0.5 to classify high-risk patients.

Descriptive data are presented as mean ± SD. Group differences were tested by Student t test and χ² test. The level of statistical significance was set at < 5%, two tailed, for all statistical tests. Analyses were performed using SPSS, version 15.0 (SPSS Inc; Chicago, Illinois) and R software language¹⁷ (R Foundation for Statistical Computing; Vienna, Austria) with the Design library.¹⁸

We examined the relationships between CV risk and PWA-I, PR-I, Spo₂-I, mean PPT, and mean RRPO using a multivariate proportional odds logistic regression model. The predictive strength of the model (c index) is the probability of concordant rankings between CV risk scores and the model’s predictions. This analysis is completely independent of the neuro-fuzzy classification and was performed on the whole group of 148 patients.

Agreement between ASI and the ESH/ESC risk strata (high risk vs low risk) was indexed using Cohen κ. Sensitivity and specificity of the ASI algorithm were evaluated by applying a diagnostic cutoff level between the average to moderate (ESH/ESC CV risk 1, 2, and 3) vs high and very high (ESH/ESC CV risk 4 and 5) groups.

A total of 148 subjects (98 men, aged 50 ± 13 years, BMI 27.6 ± 5.4 kg/m²) were included in the final analysis. Previously known systemic hypertension and ischemic heart disease were found in 47 and nine patients, respectively.

Repetitive and/or spontaneous attenuations of PWA were found during the night in all subjects. The most frequent attenuations were spontaneous and between 10% and 30% from baseline. More pronounced attenuations of the pulse wave signal (> 40% from baseline) and PR accelerations were typically associated with respiratory events, periodic limb movements, or body movements (data not shown). An example of three PWA episodes without detectable EEG arousal is shown in Figure 3.

According to the ESH/ESC CV risk matrix, 47, 43, 32, 10, and 16 patients were classified as CV risk category 1, 2, 3, 4, and 5, respectively. We examined the relationships between the five ASI parameters and the ESH/ESC CV risk score using multivariate logistic regression models. Individual models of each of the predictors (PWA-I, mean PPT, mean RRPO, PR-I, and Spo₂-I) accounted for between 60% and 70% of the models’ discriminatory power (c index). The performance of these variables in predicting CV risk was variable, but data suggested that each variable provided a strong association with a shift in risk class that was statistically significant for all variables (mean PPT, P < .0001; mean RRPO, P = .02; PR-I, P = .01; Spo₂-I, P = .0001, respectively), except PWA-I (P = .08). Correlation analysis showed that the five ASI variables interrelated with Spearman correlation varying from r² = 0.04 to 0.18. A multivariate model using all five ASI variables was validated using 200 bootstrap replications with backward stepwise term elimination. Indeed, each of the five variables contributed significantly to the model fit and none could be eliminated effectively. This full model predicted risk class with a discrimination index of 79% (χ²₍₅₎ = 59.5, P < .0001). The individual risk odds are illustrated in Figure 4. Finally, the diagnostic sensitivity and specificity of the computed composite ASI CV risk score for separation of the ESH/ESC matrix-derived CV risk class 1 to 3 (average to moderate risk) from class 4 and 5 (high and very high risk) were determined. The overall agreements between the two risk assessment tools in an analysis including all 148 patients and 49 validation group subjects were 81% and 78%, with a Cohen κ of 0.51 and 0.48, respectively.

There was no systematic difference in the anthropometric data of the training and validation cohorts (Table 1). Comorbid CV and pulmonary diseases were equally prevalent in both groups. Concomitant medications with potential effects on cardiac and vascular reactivity were numerically more common in the validation group than in the training sample (Table 2).

Diagnostic sensitivity and specificity of the neuro-fuzzy-classifier-derived ASI CV risk score were compared with the ESH/ESC matrix-derived risk classification using separate training (n = 99) and validation (n = 49) patient samples. In the training set, 13 out of 16 (sensitivity 81%) patients with high or very high (class 4 and 5) CV risk were identified correctly, whereas 70 out of 83 (specificity 84%) patients with average to moderate (class 1 to 3) CV risk were identified. In the validation set, eight out of 10 (sensitivity 80%) patients with high or very high CV risk and 30 out of 39 (specificity 77%) patients with average to moderate CV risk were classified correctly. Six out of the nine patients who were classified falsely as high risk by the algorithm belonged to the ESH/ESC moderate added risk class. Positive and negative predictive values were 0.47 and 0.94, respectively. Concomitant use of medication (Anatomical Therapeutic Chemical codes C01, C03, and C07-10) was recorded in 55 patients (Table 2), and β-blocker treatment was identified as an important factor for classification that was not in line with the ESH/ESC matrix. A receiver operating characteristic curve for the ASI CV risk score was constructed using a defuzzification threshold > 0.5 to identify the patients at high CV risk in the validation group. The area under the curve was 0.84 (Fig 05).

This study constitutes a systematic attempt to identify markers suggestive of an elevated overall CV risk using oximetry-derived cardiac/autonomic measures from an overnight sleep recording. Our results show that the ASI algorithm can discriminate the low/high CV risk defined by the ESH/ESC risk matrix with acceptable sensitivity and specificity. This novel algorithm may provide new insights into sleep-related autonomic control of CV variability, which could add diagnostic value to the identification of individuals at increased overall CV risk.

CV risk assessment is applied widely in clinical medicine and is of considerable importance for decisions on preventive measures, allocation of health-care resources, and individual diagnostic and treatment strategies. Integrated matrices have been constructed to assist doctors and health-care providers to assess patients’ CV risk.^2,3,12 These matrices are based on patients’ history, anthropometrics, and biochemical variables, as well as point assessments of CV structure and function. Continuous physiologic measures of heart rate and BP variability across the 24-h cycle have also been embedded in risk prediction strategies, but to a lesser extent.^19,20 State of sleep is characterized by physical and autonomic stability and is less influenced by other stimuli. To our knowledge, there are no risk prediction tools that selectively refer to the sleep period. The hypothesis addressed by the ASI algorithm assumes that the state of sleep, because of its physiologic integrity,²¹ can provide a unique window for investigation of CV function and disease.

Physiologic signals during sleep are typically recorded to provide diagnosis of sleep-related disorders. Classic signals reflecting CV function, such as BP variability,²² pulse transit time,²³ and peripheral arterial tone,²⁴ have also been applied in this context. Recently, skin microcirculation has been identified as an area of interest in studies of vascular health and disease.²⁵ The nocturnal PWA has been used to reflect changes in autonomic activity²⁶ and has been linked to daytime BP.²⁷ In the current study, we obtained the physiologic signals from the oximetry during an extended time period, selected the parameters based on their demonstrated relevance in reflecting CV pathophysiology and outcomes,^28‐30 combined these variables, and aimed to provide an integrated functional measure of the CV risk.

The current study was performed in a sleep clinic with a high prevalence of sleep-disordered breathing in the study population. Although this population had no, or a variable degree of, OSA and CV risk, future studies with this algorithm need to address its applicability in the general population. Compelling evidence from well-performed studies has established a robust association between OSA and CV morbidity as well as mortality.^31‐34 Recent prospective studies suggest that OSA is a cause of incident type 2 diabetes^35,36 and is associated with increased mortality, particularly related to coronary artery disease.³⁷ Consequently, there is intense interest in identifying methods, particularly those based on easily performed noninvasive techniques, for evaluation of CV risk in patients investigated for OSA. Such tests would serve both as an investigative tool for additional research and as a means of helping identify patients with OSA who are at particular risk for adverse sequels.

Several study limitations need to be addressed. Whether night-to-night variability of sleep could potentially affect the ASI classification outcome was not systematically evaluated in this cross-sectional feasibility study. Blood samples for biochemical assessment of risk factors were not obtained systematically from patients, which could have led to an underestimation of CV risk. Finally, the effect of interventions like CV drug treatment, sleep apnea treatment using continuous positive airway pressure, and medication affecting sleep propensity and sleep regulation on the ASI algorithm need to be investigated further. We did a subanalysis of the misclassified high-risk patients and found that the concomitant use of β-blockers appeared to affect our measurements of CV reactivity patterns. It is well known that β-blocking agents could decrease the CV mortality risk. Hence, it is possible that the risk reduction provided by β-receptor blocking drugs was indeed correctly interpreted by the ASI algorithm. Future studies in patients with and without β-blocker treatment will further clarify this issue.

In conclusion, a novel ASI algorithm was developed to analyze overnight cardiac and vascular variability dynamics. The measurement technique uses a straightforward continuous recording from an oximeter probe. ASI-derived CV risk markers were strongly associated with risk prediction based on the ESH/ESC CV risk matrix and showed high sensitivity and specificity for the detection of high CV risk. Prospective studies are needed for evaluation of the accuracy of the ASI algorithm to predict CV disease outcomes such as myocardial infarction, stroke, or death.

Author contributions:Dr Grote: contributed to study design, data acquisition and analysis, and preparation of the manuscript.

Mr Sommermeyer: contributed to study design, data acquisition and analysis, and preparation of the manuscript.

Dr Zou: contributed to study design, data acquisition and analysis, and preparation of the manuscript.

Dr Eder: contributed to study design, data acquisition and analysis, and preparation of the manuscript.

Dr Hedner: contributed to study design, data acquisition and analysis, and preparation of the manuscript.

Other contributions: The authors express their gratitude for support throughout the study to Helena Axelsson (Sahlgrenska), Lena Engelmark (Sahlgrenska), Martina Bögel (Weinmann GMBH), Matthias Schwaibold (MCC), and Bernd Schöller (MCC) and for their technical and intellectual input during different parts of the study.

Financial/nonfinancial disclosures: The authors have reported to CHEST the following conflicts of interest: Dr Grote has served as a medical advisor and has received travel grants and honoraria for lectures from Weinmann. Mr Sommermeyer is an employee of MCC. Dr Hedner has received travel grants and honoraria for lectures from Weinmann. Drs Zou and Eder have reported that no potential conflicts of interest exist with any companies/organizations whose products or services may be discussed in this article.

Additional information: The e-Appendix can be found in the Online Supplement at http://chestjournal.chestpubs.org/content/139/2/253/suppl/DC1.