This study investigated the feasibility of using plantar pressure data to identify frail people and predict fall events in the elderly. Over 700 senior people performed a balance standing test and a 20 m walking trial while wearing a 7-sensor plantar pressure measurement insole. One-hundred-eighty-two features were extracted from the collected plantar pressure data. Random forest models were built to identify subjects with a frail state or a recent history of falling. The overall balanced accuracy for the recognition of frail subjects was 0.75 ± 0.04 (F1-score: 0.77 ± 0.03). The overall balanced accuracy for classifying subjects with a recent history of falling was 0.57 ± 0.05 (F1-score: 0.77 ± 0.03). The classification of subjects relative to their frailty state primarily relied on features extracted from the plantar pressure series collected during the walking test. In particular, the classifiers frequently used features related to plantar pressure peaks, i.e., the “Wavelet analysis” and “Peak analysis and AUC” categories. In the future, plantar pressure data processed with random forest algorithms might be of interest to support the detection of gait-related frailty patterns. Further research works are necessary to understand how the tools used in the present study could complement the existing evaluation methods. In the present study, these tools were ineffective in classifying subjects according to their history of falling.
Plantar pressure measurement for classifying frail individuals and fallers
Studies proposing new assessment methods for frailty in senior people are regularly published [7]. The use of technology allows for more objective evaluations and is therefore attractive to clinicians. To date, several studies have successfully combined the use of inertial sensors with statistical classification techniques [11, 12, 15]. Only one study has tried to use plantar pressure to distinguish frail people from healthy individuals [12]. In a group of 186 senior people, Chkeir et al. extracted four parameters from the vertical ground reaction force analysis and COP position when stepping on a bathroom scale composed of a 4-sensor force platform. Unlike the present study, measurements were completed in static conditions only. The authors found statistical differences between healthy, pre-frail, and frail individuals but did not use machine learning techniques to develop classifying models.
The present study is the first to combine plantar pressure measurements with machine learning techniques to classify frail and healthy senior people. Among studies aiming to introduce new technology for assessing frailty, this is also the second study to test a large sample of over 700 senior people [13]. The accuracy score of 0.75 ± 0.04 may not be as high as some previous studies that used accelerometer sensors and functional tests [15, 36]. In one study aiming at classifying pre-frail and healthy subjects in a group of 124 elderly people, Greene et al. [15] reported accuracy scores of 0.84 (F1-score: 0.83) and 0.94 (F1-score: 0.94) in women and men, respectively. They collected kinematic data using a network of inertial sensors attached to different parts of the body during the completion of established clinical instruments, such as TUG, sit-to-stand, and standing balance tests. In another study consisting in classifying frail and robust subjects in a group of 309 elderly people (training sample:160, test sample: 149), Chang et al. (2013) reported an accuracy score of 0.83 (F1-scores: 0.81) [36]. They used a complex experimental set-up combining sensor units attached to several selected pieces of home furniture, again in conjunction with functional tests. They also input the data obtained from digital questionnaires surveying subjects abilities to perform activities of daily living. In contrast, the plantar pressure data used in the present study were obtained during a simple 45 s standing test and two 10 m walking trial segments; these data were obtained using a single easy-to-use instrument, i.e., the plantar pressure measurement insole, not complex multi-sensing systems used in conjunction with clinical instruments or functional tests, as the ones proposed in the above-mentioned studies [15, 36]. Perhaps, plantar pressure data obtained in the course of a TUG, sit-to-stand test, or any other challenging situation (e.g., dual tasks, etc.) would also result in higher accuracy scores. Future studies are necessary to verify this hypothesis and to understand whether the combination of features extracted from inertial sensors and in-shoe plantar pressure measurements would yield better results for identifying frail people.
Interestingly, higher performances have been noted for men than women (0.78 ± 0.07 vs. 0.72 ± 0.04). These observations are similar to those of studies that used inertial sensors and may be explained by some women-specific gait characteristics [15, 37]. Walking speed, step length, and step width were found to be lower in aging women than in their male counterparts, which points to the necessity of developing specific models for each population.
In the present study, models developed for identifying people with a recent history of falling did not show satisfactory results. The best performance was as low as 0.60 ± 0.10, only for the 65–69 age group. Further studies are needed to clarify whether models using plantar pressure data obtained in functional tests, rather than simple standing balance tests and 20 m walking trials, could yield better predictions. Plantar pressure data could also be collected in free-living conditions to try detecting near-fall events (i.e., slips, trips, missteps), the frequency of which has been shown to be associated with the risk of future actual falls [38]. The question of the adequacy of the extracted features may also be considered. While COP-related features have already shown statistical relationships with falling events in at least on previous study [22], features describing one-dimensional ground reaction forces had never been suggested in the literature and may not have the same prediction capabilities as for the frailty state prediction models. Finally, the fall history recall questionnaire used in the present study did not allow distinguishing events caused by intrinsic physical factors from the ones caused by extrinsic/environmental factors. Factors falling into the second category may not involve any physical change that could be captured by the 7-sensor plantar pressure measurement insole device used in the present study.
Plantar pressure measurements and feature extraction
Investigating the features that contribute the most in random forest classifier models may provide early insight into the physical changes that could be important for the early detection of frailty patterns. Considering previous observations on the age-related changes in walking COP trajectories and the call for using walking COP measurements for the evaluation of gait stability and postural control abilities, features extracted from COP excursion and trajectories during standing and walking trials were expected to rank among the most important features for the detection of frail individuals in the present study [23]. Instead, features providing the most valuable information to the random forest models were those related to the ground reaction force (Fig. 3C and D). Ratios of the height of peaks to the height of the valley, alongside several other parameters from the wavelet analysis, were among the most contributive features. Such parameters are associated with moving the center of gravity efficiently during the gait stance phase [39]. While the sharpness of the ground reaction force wave is closely related to walking speed in healthy subjects, alterations of this wave during walking trials have also been linked with pathologies of the lower limbs. For instance, Kotti et al. successfully used similar parameters to identify knee osteoarthritis patients [31]. Moreover, other parameters from the sensor-specific peak and AUC analysis have also been identified among the most contributive ones. Features related to the heel and center midfoot sensors are especially well represented, indicating that features reflecting the ability to sustain landing load at the beginning of the stance phase may also be considered early frailty indicators.
Interestingly, all the 27 important features identified in the present study emanate from plantar pressure data collected during the walking trial, pointing to the limit of the force plate for the evaluation of frailty and the necessity to develop systems capable of performing measurements during ambulatory trials. Moreover, random forest classifiers built with data features extracted from the standing balance test only showed a lower accuracy (balanced accuracy: 0.57 ± 0.05, weighted F1-score: 0.56 ± 0.04, detailed data not shown).
Smart insole for the early detection of frailty patterns
The objective evaluation of frailty state and falling risk in senior people remains a critical contemporary challenge in the health science field. The assessment of plantar pressures could provide crucial pieces of information, more specifically for the evaluation of the physical dimension of frailty. Indeed, aging-related gait alteration is associated with some loss of strength or with the development of sarcopenia [40]. The early evaluation of parameters that inform on the physical dimension of frailty would enable tailoring appropriate interventions early in the aging process. Plantar pressure measurements have been linked with promising preliminary observations in the past. Cheap and wireless smart-insoles similar to the one used in the present study could overcome some of the practical issues related to the use of force plates, especially when measurements are carried out during walking trials [20,21,22,23].
In addition, at the dawn of the IoT era, it is certainly possible to design smart shoe devices that can systematically collect plantar pressure data during daily life walking segments and monitor changes in COP trajectories and ground force reaction waves over several years. Considering the relatively good frailty classification accuracies produced in the present study with data extracted from a minimal number of steps, it is possible to expect higher scores with longitudinal approaches. Moreover, longitudinal monitoring of plantar pressure data in free-living conditions through smart shoe devices should not be restricted to walking segments only. Physical behavior recognition using 7-sensor plantar pressure measurement insole devices is feasible [29]. Therefore, it would be possible to isolate sit-to-stand events that naturally occur during the daily life of older adults and analyze the plantar pressure data to detect deviations in frailty patterns. Piau et al. tested the feasibility of using smart shoes to collect behavioral information in free-living conditions and for long periods of time. They observed a high level of acceptance in senior people [41]. Their smart insole device could track the number of steps, walking distance, gait speed, and active walking duration, but no functional evaluation of the participants was performed. Therefore, longitudinal and prospective studies are needed before stating on the relevance of the smart shoe approach for the individualized surveillance of gait and balance function alteration and the early detection of frailty patterns. These studies should consider how to use this new approach concomitantly to the existing methods in order to properly evaluate how they can complement them by bringing new or earlier information to the clinicians.
Limitations and strengths
One limitation of the current study is related to the imbalanced nature of the dataset. Twenty-nine percent of the subjects were defined as frail using the Kihon checklist, and only 20% of the subjects declared having experienced at least one fall event in the year preceding the measurements, resulting in a limited amount of data to train the algorithm with regard to the characteristics of these two groups. Consequently, one cannot rule out that the lower performance observed with the faller classification algorithms could be a consequence of the limited available data rather than irrefutable evidence that the 7-sensor plantar pressure measurement system proposed in the present study is unsuitable for the identification of fallers. In some subgroups, the ratio of frail subjects to non-frail subjects was extremely low. For example, no more than 11% of people aged between 65 and 69 years old were categorized as frail by the Kihon checklist, resulting in classifiers having lower performance in this age group (accuracy: 0.68 ± 0.11, F1-score 0.78 ± 0.05). Steps were taken to address this limitation. First, a large number of people (774) were recruited. The minority class could include enough samples and a variety of postural and gait patterns representative of the senior Japanese population. The study included 712 participants, which means that data from over 110 and 160 participants, for the faller and frailty analyses, respectively, were available for training the whole population models. To date, only one other study has tested the effect of wearable technology for assessing frailty or the risk of falling on such a high number of subjects [7, 12]. The 30–70 ratio between frail and non-frail people found in the present study is 7.4%. This is higher than the reported estimated prevalence of frailty in senior Japanese people [42]. This higher figure may be explained by the fact that healthier individuals are less present in the spaces through which the subjects were recruited (i.e., health and welfare administration and healthcare provider company) or less interested in having this type of postural and gait assessment. Second, the majority class has been under-sampled according to the method described elsewhere [33], in order to avoid 1) classifiers performing poorly on minority classes due to overtraining in the majority class and 2) overfitted outcomes that come from the paucity of information in the minority class.
Another major limitation relates to the standards adopted in this study to identify frailty. In the absence of the gold standard method, the Kihon Checklist was used to determine frailty status. The method is widely used in Japan and has been described as a valid frailty prediction tool in several reports [43, 44]. However, many tests are available to predict frailty. Some authors have suggested that results could vary widely, especially between self-administrated methods such as the Kihon checklist and tests administered by nurses or physicians [45]. Therefore, it is possible that the accuracy score of the present study could have been different, either increased or decreased, if another frailty assessment tool had been used as a reference instead of the Kihon checklist. In the future, new objective assessment methods, such as the one proposed in the present study, should be tested against a broader panel of frailty assessment methods to strengthen the interpretability of the results.
Another limitation of the present study is the non-inclusion of variables related to the medical history of participants in the predicting models. Indeed, a combination of plantar pressure data and medical information could strengthen accuracy scores for either falling history or frailty state predictions. However, building such kind of models would not only have required a systematic collection of medical history, but also an even larger group of participants to have enough individuals per medical condition so that the learning algorithms can identify patterns.
Finally, the present protocol does not allow excluding the presence of inaccuracies for the falling history parameter, which may have negatively impacted the accuracy scores. Future studies should include a more robust protocol for the collection of information related to falling events, be they recalled data or prospective protocols.