Session Item

Recent studies into radiation dose to particular cardiac sub-regions during lung cancer radiotherapy show a strong association with adverse cardiac events. Wearable health monitors such as smartwatches are becoming popular and allow comfortable long-term heart-rate monitoring. We present the development of a methodology to reliably extract smartwatch ECG features to gain insight into the development of cardiac toxicity and preliminary results from an initial 10 recruited patients.

Withings scanwatches were used to take 30s single-lead ECGs by 4 volunteers and 10 patients under approved study protocol. ECG features (R, T and P peaks, Figure 1) were used to identify abnormalities in cardiac functions: premature ventricular contractions (PVC), atrial fibrillation (AF) and premature atrial contractions (PAC). The quality of ECG varied considerably due to poor connection during acquisition. ECGs containing too much noise were identified and removed. For each ECG, repeating peaks were first identified, constraints were placed on the average, maximum and relative peak separation, followed by template matching to the average beat shape. Next, smoothing was applied, consisting of a bandpass filter (1.3-45Hz), 5-point first-order differentiation, back-cumulation and squaring to emphasise the peak shape and suppress high-frequency noise. R-peaks were identified as the tallest regularly repeating peaks and used to calculate heart-rate. R-peaks were used to determine irregular heartbeats, which were used to remove the QRS complex and identify AF, PVC and PAC. From the remaining data, two moving averages were calculated with differing smoothing widths to identify T-waves.

For validation, identified peaks were compared to peaks manually identified by clinicians, and accuracies graded by sensitivity (Se), positive predictivity (+P) and average time error in ms (TE). Volunteer data was used to validate R-peak identification. T and P wave verification was conducted using publicly available annotated databases. The average heart-rate, PR and RT intervals for each ECG were plotted against each patient’s 6 weeks of monitoring.

Volunteer and public data showed good accuracy (TE=7.10ms, Figure 2), comparable to the literature values, despite the increased noise level removing 13% of ECGs. Initial analysis of 10 patients (taking a mean of 2 ECGs per day) shows, on average, a decrease in resting heart-rate over the 6 weeks during which the watch was worn. Signs of PVC and PAC were identified in 6 patients.

We have developed a framework to identify key features in single-lead smartwatch ECGs. We have validated on multiple ECG databases and found accuracies comparable to the literature. Preliminary analysis on 10 patients with lung cancer showed feasibility for long-term monitoring, with the framework identifying potential abnormalities in cardiac function.