Individuals with stroke were recruited from the University Hospital Zurich research database. Inclusion criteria were age ≥ 18 years, chronic stage of stroke (≥ six months since the incident), and ability to walk without assistance on a level surface for at least six minutes (Functional Ambulation Category, FAC ≥ 3). The use of assistive devices was allowed. Potential participants were excluded if they had gait deficits unrelated to stroke or were unable to understand verbal instructions. Seventeen individuals with chronic stroke were recruited for this project. The study was approved by the cantonal ethics committee (BASEC-Nr. Req-2019-00565, Ethics Committee Zurich, Zurich, Switzerland). All participants provided written informed consent after being informed about the study procedure.

The participants were invited to attend a single measurement event consisting of two sessions of approximately one hour each. After obtaining informed consent, demographic information and outcome measures were collected to characterize the sample. This included age, gender, time since stroke, type of stroke, use of walking aids, 10mWT, gait speed, six-minute walk test, and Motricity Index. All participants were equipped with insoles placed within their personal footwear while performing two measurement blocks (timepoints T1 and T2), each with three repetitions of the 10mWT. A break of at least 30 min separated the two measurement blocks; during this break, the measurement system was removed from the shoe and charged. Post-break, system mounting was performed from scratch, facilitating the evaluation of test–retest reliability while minimizing the impact of day-to-day variation. All gait tests were recorded with both measurement systems, which were started and synchronized before each test.

A Stappone insole system (version 1.0, stapp one, Vienna, Austria) collected plantar pressure with 12 resistive textile pressure sensors in specific zones under the foot. Additionally, it comprised a triaxial accelerometer, a battery, onboard buffer storage, and a Bluetooth Low Energy 4.2 emitter for each insole. Once the data was collected, parameters were processed onboard the system using a proprietary adaptive thresholding algorithm. The sampling frequency was set to 100 Hz. Raw data for acceleration, plantar pressure, and gait event timings (FS and FO) were recorded and transferred via Bluetooth to the Stappone software (version 0.9.5.6, stapp one, Vienna, Austria) after each trial. The pressure calibration of the insoles was automatically adjusted by the software during postprocessing after each measurement was taken. The insole data were analyzed using an internal threshold-based algorithm for gait event detection. The sensor system was lightweight and could be worn for 36 h before needing a recharge.

A GoPro Hero 8 camera (GoPro, Inc., San Mateo CA, USA) served as the reference system because, unlike optoelectronic systems, it does not suffer from the uncertainty associated with human gait labeling (15). The use of manual labeling can lead to inaccuracies due to human error. To minimize these errors, the four-eyes principle was introduced. The camera was securely mounted on a rolling trolley at the height of the feet to capture all trials from a low vantage point with minimal occlusions (Figure 1). The sampling frequency was set to 120 Hz with a resolution of 1,080 pixels and no lens correction. During postprocessing, the gait events (FS and FO) were manually labeled frame by frame by one rater, and the timing was calculated as the time elapsed since the start signal.

Measurement signals were synchronized for each trial using an audio output that was generated by the insoles when the measurement was started. The GoPro was connected to the audio output using a microphone adapter, which recorded the start signal as a sound burst. During postprocessing, this burst was identified with a threshold algorithm and verified manually.

The insole system and the reference system were used to assess the stance duration for both the affected and the unaffected side. The accuracy, concurrent validity, and test–retest reliability of stance duration were analyzed from the steps taken in the six 10mWTs per participant. Stance duration, defined as the time between FS and the subsequent FO of the same foot, was calculated for every step. The stance duration was calculated once based on the insole timings and once based on the video reference timings. The timings of FS and FO were calculated as the time elapsed since the starting signal of the video. To ensure correspondence between FS and FO, e.g., that both belonged to the same step on the same side, a plausibility check was performed to remove all out-of-distribution values.

The accuracy of the event detection was assessed with the F1 score to balance precision and recall. Precision was defined as the total number of true positive (TP) identifications by the insole system and the reference system divided by the sum of the TPs and the false positives (FPs; events detected by the insole system but not by the reference system).Precision=TP(TP+FP)Recall was defined as the number of TPs divided by the sum of the TPs and the false negatives (FNs; events identified by the reference system but not by the insole system).Recall=TP(TP+FN)The F1 score was used to balance precision and recall and was defined as:F1score=2×Precision×RecallPrecison+RecallTo evaluate the concurrent validity of the insoles, the differences between the stance durations measured by the insole system and by the reference system were analyzed. Normality of the distributions of stance duration was checked by visual inspection of QQ-plots by two investigators. For conflicting ratings, a consensus was reached through discourse. Subsequently, a paired two-sided t-test (α < 0.05) was employed to detect statistically significant differences between the systems.

Furthermore, the agreement of the stance durations between the two measurement systems was characterized using Bland–Altman plots. Such plots have a strong history in the comparison of reference systems to experimental systems in the medical field (32, 33). Their utility lies in the fact that they not only compare means over the given range of the test population but also enable the identification of fixed and variable bias individually. Both features are important in judging whether a system is sufficiently precise for a given application and population.

Finally, the ICC_2,1 was used to examine the agreement between the two measurement systems. The ICC_2,1 was calculated using the following equation:ICC(2,1)=MSR−MSEMSR+(k−1)MSEwhere MS_R = mean square for rows, MS_E = mean square for error, and k = number of raters/measurements (34). According to Fleiss, an ICC ≥ 0.75 is considered “excellent,” 0.40 ≤ ICC < 0.75 is considered “moderate to good,” and ICC < 0.40 is considered “poor” (35). The ICC was applied, as it provides a measure-independent metric of agreement that allows for comparison between different metrics and measurement systems.

In a last comparison, we evaluated test–retest reliability for each system by comparing stance duration data from T1 with data obtained at T2. Differences were identified using a paired two-sided t-test after a visual check for normality of the distributions using QQ-plots. The ICC_2,1 (34) was again applied to permit comparability of reliability to other systems and measures.

Of all 3,820 gait events, 90.86% were correctly identified by the insole system. A cumulative confusion matrix illustrates gait event detection for continuous walking in all participants by summation (Figure 2). The data analysis showed that 91 right FSs, 86 right FOs, 87 left FSs, and 85 left FOs were not identified as gait events by the insole system. Further investigation showed that all missing events corresponded to either the first FO or the last FS of the trial. To simulate continuous walking, these gait events were excluded from further analysis. Recall values ranged from 0.994 to 1, with a precision of 1 for all measurements. The F1 score ranged from 0.997 to 1, as shown in Figure 2. Figure 2

Cumulative confusion matrix for continuous walking. The rows of the table correspond to the total number of gait events detected by the reference system. The columns correspond to the total number of gait events detected during continuous walking by the insole system.

Concurrent validity of stance duration was determined through comparison of the insole-derived timings and the video reference measurements. FP events were excluded from the analysis. Sixteen left (1.96% of a total of 815) and 17 right (1.94% of a total of 878) stance durations were excluded as outliers. No systematic reason was identified for the causation of these outliers.

The normality of the distributions of stance duration data was confirmed by both raters using the QQ-plots. The t-test performed on the stance durations calculated by the insoles and the reference system revealed no significant difference (p = 0.06). The stance durations recorded by the insole system tended to be slightly longer than those recorded by the reference system (difference of −0.01 s). The limits of agreement showed a deviation of insole measurements of up to 0.33 s (Figure 3). The correlation analysis between the insole and the video reference system was robust, with an excellent ICC_2,1 value of 0.874.

To enhance the visual clarity of our analysis, we have provided two alternative visualizations of the data reported by the Bland–Altman plot. Specifically, we show the difference in stance duration between both systems categorized by affected vs. unaffected side (Figure 4). Furthermore, we depict differences in stance duration of three categories of ambulators (Figure 5): limited community ambulators (<0.8 m/s), community ambulators (0.8–1.4 m/s), and those able to cross the street safely (>1.4 m/s) (36).

The normality of the distributions of stance duration data between T1 and T2 was confirmed by both raters using the QQ-plots. The paired, two-sided t-test indicated no significant difference (p = 0.44) between the two measurement blocks, T1 and T2. The test–retest reliability between measurement timepoints T1 and T2 was excellent, with an ICC_2,1 of 0.928. The mean difference in stance duration between T1 and T2 was 0.03 s.

The insole system correctly identified 90.86% of the 3,820 recorded gait events. This result aligns with a study that correctly identified 94.5% of all steps in a population with neurological conditions by employing insoles equipped with a piezoresistive device (23). Both recall and precision were >0.98. These rates are comparable to or exceed those reported in studies using accelerometers in individuals with stroke (37, 38). The insole system failed to detect 9.14% of the gait events. Analysis revealed that the initial FOs at the start of the trial and the final FSs at the end of the trial were consistently missed because participants initiated and concluded trials in a standing position. Simulating continuous walking by removing the first FOs and the last FSs from the analysis resulted in a detection error rate of only 0.2%. Among these erroneously detected gait events, eight were FSs detected before the participant started walking, suggesting potential algorithm sensitivity to weight shifts without actual foot movement (39).

The validity of the stance duration measurement was excellent. In comparison to the FeetMe® Monitor system (ICC > 0.77), which was validated in individuals with chronic stroke over three 8-min trials (18), the utilized insole system demonstrated superior performance. Our results align with the accuracy of the ZeroWire® footswitch system, which has four piezoresistive pressure sensors on each foot and was used to record stance time in healthy participants; its accuracy ranged from an ICC of 0.86 to 0.97, depending on walking speed (24). Due to the increased variability and decreased walking speed that are often found in pathological gait, Inertial Measurement Units (IMU) often report reduced accuracy at low walking speeds of 0.4 m/s or severe deviations from normal human gait (40, 41). This is reflected in the lower accuracy of IMU sensors used to detect stance duration in individuals with stroke (42). While we did not record any individuals walking at less than 0.4 m/s, the insole system performed remarkably well for 0.5 m/s–0.6 m/s speeds, which are associated with long stance durations (43). This consistency indicates that the system should function well, even for slow-walking individuals post-stroke.

The high level of agreement between the two measurement timepoints, T1 and T2, as indicated by the ICC_2,1 of 0.92, is consistent with findings for the FeetMe® Monitor system (ICC > 0.9) (18). The difference between T1 and T2 can be attributed to typical variation in stance duration (44). These findings indicate that the measures are repeatable over time, even if the insole is removed and reinstalled in the shoe. It is important to note that a short pressure calibration was performed after each introduction of the system into the shoe to compensate for changes in lacing pressure.

Alterations in walking patterns or the ability to walk can serve as a valuable source of information regarding patient trajectories throughout the rehabilitation pathway and beyond. Gait strategies are influenced by various clinical factors, including neurological, cardiovascular, and musculoskeletal conditions; cardiovascular and metabolic diseases; age-related changes; and trauma (9). Wearables have found extensive application in research, and their utilization is expanding into clinical settings. They can provide information about spatiotemporal, kinematic, and kinetic parameters. Nevertheless, caution is advised in the application of these devices, as they are specifically designed to extract certain parameters in particular patient populations. Therefore, these applications need to undergo a validation process in patients with specific diseases or conditions in clinical settings, as well as in real-world conditions if they are employed in such environments (45). Most importantly, the information gathered using these sensors must align with clinical endpoints and have clinical relevance (46). This ensures that such information can be effectively used by clinicians to provide valuable additional insights into individual health conditions and to facilitate clinical decision-making.

Our findings suggest that the insole system is applicable for assessing stance duration in individuals post-stroke. However, this conclusion is limited to clinical settings and continuous, straight walking. Further research is necessary to determine whether these findings can be transferred to a real-world environment and various activities of daily life.