**5. Experiments and Results**

The target of the present work is to develop a low-cost, long-term module to estimate load on a cane while users walk. Hence, to validate the system, it is necessary to check that the readings of the sensor depend on the load on the cane during the gait cycle. It is important to test the device with its target population, as healthy volunteers reportedly use walking aids in different, non-consistent ways even when they try to simulate pathological gaits. Our test population includes only people who rely on a traditional cane for mobility in their ADLs (Figure 7).

We had already tested that the cane reliably measures static weights during the calibration process (Figure 4). The error was bigger when the load was below 2.5 kg (mean squared error of 0.1389). However, it is necessary to check how the module behaves during a gait cycle, with dynamically changing loads. As commented, the load on the cane depends not only on user weight but also on their condition. Hence, it is interesting to test the system with people with different conditions to check that the module works well for different gait abnormalities. Unfortunately, although weight can be normalized or accounted for, major load variations due to the condition cannot be quantified, i.e., we cannot directly predict how much weight a given person is going to support on the cane at each time instant to check whether our module is accurate or not. To correlate our readings with the condition, we assess users via their gait speed, which has been consistently reported as a meaningful parameter. We have measured manually the gait speed using two markers in the floor and a highly precise chronometer.

Figure 8 shows the cane load over time for each user presented in Table 1. We can observe that some users load significantly more weight on the cane with respect to others. For example, users 1 and 4 present peak values of 8.55 kg and 11.78 kg on average when compared to users like 2 or 5 (0.18 kg and 0.32 kg, respectively). The main reason for this variability is that, as commented, load depends largely on the users' condition, even more than on the users' weight, i.e., users with poor condition need more assistance. We can also notice that some users increase the load on the cane the longer they walk, like user 3 (meniscus surgery in both knees).

**Figure 8.** Load on cane *y*-axis (kg) over time *x*-axis (seconds). Users were suggested to walk for one minute, some of them (users 1 and 7) walked less than the minute and others more than the minute.

As weight bearing depends largely on condition, we had no benchmark function to determine how much weight each volunteer loaded on the cane at each time instant. Hence, we used gait speed as an indirect measure of disability to check that users with poorer conditions returned larger loads on the cane. People with severe dependencies typically present gait speeds below 0.6 m/s [30] and they are expected to bear more weight on the cane. Figure 9 shows the relationship between gait speed during our 10 m tests and the load peaks for each of our volunteers. As expected, gait speeds below 0.615 m/s are related with higher loads on the cane: 7.47 kg on average (ranging from 1.15 kg to 17.13 kg); whereas volunteers with gait speeds above its have an average load on the cane of 0.89 kg (ranging from 0.18 kg to 1.45 kg). Additionally, load variances for gait speeds under 0.615 m/s (variation range from 0.84 kg to 2.69 kg) are higher when compared to gait speeds above this limit (variation range from 0.17 kg to 0.68 kg). This gait speed relation with load on cane confirms that users with poor condition need more assistance than others with lower condition and that the need for aid depends largely on their condition. Specifically, we obtained the Pearson correlation coefficient between gait speed and median load for our volunteers (*H*<sup>0</sup> is a correlation equal or greater than 0). The resulting coefficient to −0.7473 (*p*-value 0.0065), meaning that gait speed and load are inversely related, as reported by clinicians [31]. Volunteer 6 is an outlier in this analysis because he presents a vestibular disorder, i.e., he uses the cane for balance rather than for weight bearing. If we remove him from the correlation, the coefficient is equal to −0.7971 (*p*-value 0.0050).

**Figure 9.** Load on cane vs users' gait speed.

Finally, we can observe in Figure 8 that the upper bound for our test volunteers is equal to 29 kg. This limit fits well the target 3-days of use without charge for 8 h per day of loading. In conclusion, the proposed module meets the required constrains: (i) it is cheap and easy to add to a commercial cane; (ii) it reliably estimates load on the cane in static and dynamic conditions; and (iii) it can be used for long-term monitoring.
