Extracting Cardiac Information From the Pressure Sensors of a Dialysis Machine

University dissertation from Department of Biomedical Engineering, Lund university

Abstract: This doctoral thesis in biomedical engineering deals with cardiac monitoring using the built-in extracorporeal blood pressure sensors of a hemodialysis machine. Patients treated with hemodialysis often suffer from cardiovascular disease. Despite this, cardiac monitoring is not routinely performed during dialysis treatment, since external devises required for such monitoring causes patient discomfort and increased workload for the clinical staff. Extraction of cardiac information from the pressure sensor signals is complicated by the fact that the cardiac pressure pulses are obscured by pressure pulses caused by the peristaltic blood pump of the dialysis machine. The cardiac signal component is of much lower magnitude than the pump pressure component, and the pump and heart rates may coincide.The thesis comprises an introduction and four papers describing methods for extraction and characterization of cardiac information from the built-in pressure sensors of a dialysis machine. In the first paper, a method is proposed for estimating the cardiac signal by subtracting an iteratively refined blood pump model signal from the signal measured at the extracorporeal venous pressure sensor. The method was developed based on simulated pressure signals, and evaluated on clinical pressure signals acquired during hemodialysis treatment. Heart rate estimated from the clinical pressure signal was compared to heart rate derived from a reference photoplethysmographic (PPG) signal. The results suggest that the accuracy is sufficient for analysis of heart rate and certain arrhythmias. In the second paper, a method is proposed for improved cardiac signal extraction by combining the arterial and the venous pressure sensor signals of the hemodialysis machine. Using different techniques for combining the arterial and venous pressure signals, the performance is evaluated and compared to that of the method in the first paper. Heart rate and heartbeat occurrence times, estimated from the extracted cardiac signal, are compared to the corresponding quantities estimated from the PPG reference signal in nine hemodialysis treatments. The results show that the proposed method offers superior estimation at low cardiac signal amplitudes, enabling cardiac monitoring during treatment without the need of extra sensors for more patients. Ventricular premature beats (VPBs), being abundant in hemodialysis patients, can provide information on cardiovascular instability and electrolyte imbalance. The third paper describes a method for VPB detection in cardiac signals extracted using the method described in the second paper. A set of features characterizing the cardiac pressure pulses is extracted, and linear discriminant analysis is performed to classify beats as normal or VPB. Performance is evaluated on signals from nine hemodialysis treatments, using leave-one-out cross validation. The simultaneously recorded and annotated PPG signal serves as reference. The results show that VPBs can be reliably detected for average cardiac pulse pressures exceeding 1 mmHg. In the fourth paper, we propose a novel method for detection of venous needle dislodgement. Four features, extracted from the arterial and venous pressure sensors signals, are used as input to a support vector machine which determine whether the venous needle is dislodged. The support vector machine is trained on a set of laboratory data, and tested on an-other set of laboratory data as well as on four clinical recordings. The results show that dislodgement is detected long before the blood loss has caused serious injury to the patient. In summary, the work in this thesis contributes with novel methods for extraction of cardiac information using the built-in pressure sensors of the dialysis machine. The main benefit of such approach for cardiac monitoring, compared to existing techniques, is that no extra sensors are required. All results are preliminary, and the methods needs to be validated on a larger set of clinical recordings.