2.1.1. Lab Study

In the lab study, data from 27 amateur runners (21 male, 6 female) were recorded. The dataset included runners with different strike types. Six of the subjects were forefoot/midfoot runners, and 21 subjects were rearfoot runners. The classification of the strike type was based on the definitions of Altman et al. [17]. Further anthropometric data can be found in Table 1. Before data acquisition, all subjects were informed about the related risks and gave written consent to participate in the study and for the collected data to be published.


**Table 1.** Anthropometric data of subjects participating in the lab study.

The subjects were equipped with running shoes in matching sizes (Response Cushion 21, Adidas AG, Herzogenaurach, Germany), as depicted in Figure 1a. This model had a cavity in the right shoe midsole for the placement of a sensor. We cut another cavity of the same size at the same location into the left shoe midsole to be able to acquire data from both the left and the right shoe in order to record more data for the training and evaluation of the algorithms. The specific IMU we used was the miPod sensor [18]. The accelerometer of the sensor was configured with a range of ±16 g and the gyroscope with a range of ±2000 ◦ s , and data were sampled with a frequency of *fs* = 200 Hz and a resolution of 16 bit. Before each data acquisition, the IMUs were calibrated using the calibration procedure introduced by Ferraris et al. [19]. Figure 1a depicts the orientation of the sensor in the sole of the running shoe: *x* points in the lateral direction, *y* in the dorsoventral direction, and *z* in the craniocaudal direction.

As the gold standard for velocity and stride length, we used a motion capture system (Vicon Motion Systems Inc., Oxford, UK) with 16 infrared cameras and recorded data with a sampling rate of *fs* = 200 Hz. A submodel of the marker setup introduced by Michel et al. [20] containing six markers on each shoe (see Figure 1a) was used. The marker on the heel (for rearfoot runners) and the lateral sided toe marker (for forefoot/midfoot runners) were used to extract strides. Depending on the strike type, minima in the trajectory of the corresponding markers were used to label initial ground contacts [21]. The IMUs and the motion tracking system were synchronized using a wireless trigger [22], which was connected to light barriers (S40 Series, Datalogic, Bologna, Italy). The light barriers triggered the start and the end of the recording for each trial in both systems. Using the described synchronization technique, we were able to match strides in the motion capture gold standard data to strides in the IMU signal.

The subjects were asked to run various trials with different velocities in the range of 2–6 m/s. We defined these velocity ranges to cover a wide range of relevant running velocities. As the capture volume was restricted to 6 m, and the stride length varied depending on the running velocity, different numbers of strides were recorded for the different running velocities. We recorded five additional trials for the two high velocity ranges to increase the number of captured strides. The velocity ranges and number of trials recorded can be found in Table 2. The subjects were asked to accelerate before and keep the pace within the capture volume. We measured the velocity at the beginning of the motion capture system volume using the above-mentioned light barriers used for synchronization. The velocity measured by the light barriers was used to ensure that a sufficient number of trials were recorded within each velocity range, for each subject. If necessary, the subjects were instructed to run faster or slower in order to ensure the defined number of trials in each velocity range. The ground truth value for each stride's velocity *vref* was computed from the motion capture reference as:

$$v\_{nf} = \frac{d\_{ref}}{t\_{ref}} = \frac{d\_{nf} \cdot f\_{\text{s}}}{N\_{stride}} \tag{1}$$

where *dref* is the stride length obtained by the difference of the positional data obtained by the motion capture system between two consecutive initial ground contacts, *tref* the corresponding reference stride time, *Nstride* the number of samples in between two consecutive initial ground contacts, and *fs* the sampling rate. Figure 1b illustrates the setup and running path of the subjects during the lab data recording. Overall, 2377 strides were recorded during the lab study for the evaluation of the algorithms.


**Table 2.** Number of trials and recorded strides per velocity range in the lab study.

**Figure 1.** (**a**) Shoe equipped with a miPod sensor and the marker setup. The IMU is located within the sole of the running shoe. The marker setup allowed for a computation of velocity and stride length. (**b**) Illustration of reference system setup. The subjects ran through the capture volume of the motion tracking system, created by 16 infrared cameras, and looped back around.
