In the study, 100 consecutive patients between age 5 and 17 at the pediatric heart center of Skåne University Hospital Lund, Sweden, were enrolled. Sixteen patients were considered for the study but did not participate due to comorbidity such as severe autism or Down syndrome, or because they chose not to participate. The median age was 12.9 (IQR 8.7–15.6) years, and 54 (54%) participants were male (Table 1).

In some cases, it was not possible to perform certain measurements (e.g. PR interval in complete AV block or isoelectric T wave; Fig. 2a-b). Manual smartwatch measurements could be performed on more ECGs than automated measurements (Table 2).

Examples of smartwatch ECG tracings with; a good quality. b low quality due to baseline artefacts. c mismatch between the automated (73 ms) and manual (200 ms) measurement of the QRS duration. d a flat T-wave where an automated measurement was performed but a manual measurement was deemed not feasible. ECG = Electrocardiogram. Ms = Milliseconds. QRS = QRS duration. QT = QT interval. SW = Smartwatch

Overall, manual and automated measurements of the 12-lead and smartwatch ECGs differed significantly for all variables (Table 3).

Looking at the agreement between the different ECGs, we first focused on the agreement between the automated and manual lead I and II measurements of the 12-lead ECG. Agreement between automated and manual measurements were excellent for all parameters on the 12-lead ECG (ICC ≥ 0.9, p < 0.001; Table 4). Similarly, agreement between automated and manual measurements derived by smartwatch was excellent for HR, RR interval, PR interval and QT interval (ICC ≥ 0.9, p < 0.001; Table 4). By contrast, agreement for the QRS duration and QTc interval was worse – though still good (ICC 0.75–0.79, p < 0.001; Table 4).

Next, we compared the automated and manual smartwatch measurements to the 12-lead ECG measurements (automated and manual lead I and II; Table 5). The ICCs for the standard 12-lead ECG compared to the smartwatch ECG were all highly significant (p < 0.001). The ICC for the automated smartwatch and 12-lead measurements were excellent for HR (ICC ≥ 0.9, p < 0.001), good for the PR interval and the QT interval (ICC 0.75–0.89, p < 0.001), and moderate for the QRS duration and the QTc interval (ICC 0.5–0.74, p < 0.001). In comparison to the 12-lead ECGs, the agreement of the manual smartwatch measurements was superior to the automated smartwatch measurements for the PR interval, QRS duration, QT interval and QTc interval. The agreement for manual smartwatch measurements compared to 12-lead ECGs was good or better for all variables (ICC > 0.84).

In order to visualize the agreement between the automated 12-lead and smartwatch ECG, we performed Bland Altman plots for each parameter (Fig. 3a-e). There was no significant difference in HR between the two modalities (bias − 0.6 beats per minute (bpm) (95% limits of agreement (LoA) − 11.3 to + 10.2 bpm), p = 0.499; Fig. 3a). Only one patient (1%) was considered an outlier, with a difference of 46 bpm between the two modalities. The difference was attributed to the 12-lead ECG overestimating the HR in a patient with complete AV-block, while the smartwatch derived HR was correct.

Bland Altman plot picturing the level of agreement measured automatically with a smartwatch and a standard 12-lead ECG (12-lead ECG – Smartwatch ECG). The solid red line represents the bias and the dotted lines the upper and lower limits of agreement. a HR (bpm). b PR interval (ms). c QRS duration (ms). d QT interval (ms). e QTc interval (ms). Bpm = Beats per minute. ECG = Electrocardiogram. HR = Heart rate. LoA = Limits of agreement. Ms = Milliseconds

For the PR interval, we also found no significant bias (bias 2.2 ms (95% LoA -35.8 to + 40.2), p = 0.403; Fig. 3b). Seven patients (8%) had a difference in PR interval above 30 ms between the two modalities. In four of those cases, the P-wave was flat and difficult to discern.

Regarding the QRS duration, the smartwatch significantly underestimated the values compared to the 12-lead ECG (bias 18.4 ms (95% LoA -37.0 to + 73.8), p = 0.002; Fig. 3c). 28 participants (29%) had a difference > 30 ms between the two ECGs, mostly due to the watch underestimating the QRS interval through inaccurate determination of the S-wave’s ending (Fig. 2c).

The QT interval measured by the smartwatch had a tendency towards being underestimated (bias 28.8 ms (95% LoA -43.0 to + 100.5), p = 0.102; Fig. 3d). Fifty-one patients (60%) had a difference above 30 ms between the two readings. Outliers regarding the QT interval were usually due to difficulty deciding where the QRS complex began, and the T-wave ended (Fig. 2d). Furthermore, the smartwatch underestimated the QTc interval significantly (bias 30.6 ms (95% LoA − 54.8 to + 116.1), p < 0.001; Fig. 3e). 51 patients (60%) had a difference above 30 ms compared to the 12-lead ECG. The 51 outliers were not the exact same patients as for the QT interval despite the QTc interval being calculated from the QT.

Interobserver agreement for the manual smartwatch measurements was excellent for all times and intervals (RR: ICC 1.00 (95% CI 0.99 to 1.00), p < 0.001; PR interval: ICC 0.97 (95% CI 0.90 to 0.99), p < 0.001); QRS interval: ICC 0.90 (95% CI 0.70 to 0.97), p < 0.001; QT interval: ICC 0.96 (95% CI 0.88 to 0.99), p < 0.001).