Signal artifacts in wearable-based biosignal acquisition in patients with ventricular assist devices

Background

Consumer wearables are increasingly used for self-monitoring, but their performance in patients with continuous-flow LVADs is uncertain. We evaluated feasibility and accuracy of smartwatch-based biosignal acquisition in advanced heart failure (HF) with and without LVAD support.

Methods

In a prospective, single-center pilot, 30 patients were enrolled (15 AHA/ACC stage D HF, 15 continuous-flow LVAD). Each underwent parallel measurements with a consumer smartwatch and clinical references: cuff blood pressure, fingertip SpO2, ECG-derived heart rate (HR), and a 12-lead ECG. Following skin antisepsis, the device was secured to the wrist; HR recording excluded an initial 20-s calibration. Simultaneous SpO2 and three 30-s single-lead ECGs were acquired. Up to three attempts per modality were allowed; failures after three attempts were deemed unsuccessful. Agreement was assessed by Bland–Altman analysis.

Results

Thirty patients were included: heart failure (n = 15; age 65 ± 5 years; LVEF 25 ± 2%) and LVAD (n = 15; age 62 ± 2 years; LVEF 17 ± 2%; flow 4.5 ± 0.2 L/min). In the HF cohort, smartwatch data acquisition demonstrated high success for HR, SpO2, and single-lead ECG (overall success 93%). QRS width on the smartwatch lead deviated from standard lead I by 16±4% with 33% of patients within a 10% margin (95% CI, 15–58%).

In LVAD patients, HR detection was feasible in 100% and accurate in 80% (≤5% deviation) and 93% (≤10% deviation), independent of aortic valve opening. Bland–Altman bias was −0.1 bpm (SD 5.7 bpm; limits of agreement [LoA] −11.3 to 11.1 bpm). Single-lead ECGs were interpretable in 73%. All non-interpretable traces exhibited QRS amplitudes <0.3 mV relative to baseline noise, suggesting a practical amplitude cut-off. QRS width deviation was 12±4%, with 53% (95% CI, 30–75%) within a 10% margin; bias was −5.3 ms (SD 16 ms; LoA −36.7 to 26.1 ms). SpO2 monitoring largely failed (20% success) without identifiable predictors of success (hair density, blood pressure, LVAD flow, and pulse index not associated). Bland–Altman bias was −1.7% (SD 2.1%; LoA −5.8% to 2.4%).

Conclusions

In advanced HF without LVADs, smartwatch acquisition of HR, SpO2, and ECG signals was largely feasible. In continuous-flow LVAD patients, smartwatches provided reliable HR and frequently interpretable single-lead ECGs, whereas SpO2 acquisition remained unreliable. These findings delineate the potential and current limitations of wearable-based telemonitoring in LVAD populations and underscore the need for algorithmic and hardware optimization to improve oximetry and low-amplitude ECG detection.Bland–Altman analysisECG Curves