Evaluation

The algorithm described above was implemented in a short C-language program (see the appendix below). The program was tested using a set of 35 ECG recordings (originally compiled for another study) for which a simultaneously recorded gait (walking cadence) signal was available as an independent activity indicator[1]. These recordings were made using conventional two-channel ambulatory ECG recorders, but up to four foot switch signals were multiplexed and recorded in lieu of one of the ECG signals. Each recording was between 18 and 79 minutes long; most were approximately 35 minutes. Three activities were recorded: in the first phase (R), subjects rested in a supine position for five to twenty minutes; in the second phase (S), subjects were asked to stand for six minutes; and in the third phase (W), subjects walked at a comfortable pace for at least eight minutes. The population included twelve young (aged 27 $\pm$ 5) and eleven elderly (aged 76 $\pm$ 5) subjects; the ages of the remaining subjects were not noted.

The ARISTOTLE arrhythmia detector[2] analyzed the ECG signal from each recording, generating an annotation file containing encoded beat labels. The tach postprocessor[3] extracted the interbeat intervals from ARISTOTLE's annotation file and generated a uniformly sampled (2 Hz) instantaneous heart rate signal for each recording using the method of equation (1). This signal was used as the input to the activity estimator.

Software developed for the purpose analyzed the foot switch signals, producing step annotation files similar to ARISTOTLE's beat annotation files for each record. In order to obtain an accurate independent index of activity, the step annotations were corrected manually as necessary (note that no corrections were applied to the beat annotations, however). The corrected step annotations were then processed using tach as for the beat annotations, to obtain a uniformly sampled instantaneous gait signal. This signal was taken as the ``gold standard'' measurement of physical activity. The three phases of the experimental protocol (R, S, and W) are easily identified from the gait signal.

The ECG-based activity index was derived from the instantaneous heart rate signal at 2.5-minute intervals throughout each recording. A typical example is presented in Figure 2.

Figure 2: An example of the recordings used to evaluate the activity index. G: instantaneous gait (steps per minute); H: instantaneous heart rate, $H(t)$ (beats per minute); A: activity index, $A(t)$; S: stationarity index, $S(t)$; P: power, $P(t)$. $A(t)$, $S(t)$, and $P(t)$ are updated at intervals of 2.5 minutes. The recording begins during the resting phase. The standing phase begins during minute 14 and ends during minute 23, when the walking phase begins. The minimum activity index occurs in the first five minutes in this case.

For each recording, the five-minute period for which the activity index was minimum was identified. In all 35 cases, this period fell entirely within the resting phase, as had been determined independently from observation of the gait signal. For comparison, alternative indices containing only one component of $A(t)$ were tested. The resting phase was identified successfully in 30 of 35 cases using $\overline{H}_{N}(t)$, in 21 of 35 cases using $S(t)$, and in 27 of 35 cases using $P(t)$.