Biometrics Identification and Gait
June 2018 Issue
Any parent with a child in organized sports has spent many hours watching from the sidelines, bleachers, and sometimes the parking lot as players practice and compete. After an extended period watching my daughter and her soccer teammates practice and play, I was able to identify most of the players based on their distinctive running style, even from a significant distance and as the light faded at the end of an evening practice. At times, it was even possible to identify that a player was altering her usual gait due to injury. Observing, characterizing, and documenting specific features of gait is an important aspect of many O&P encounters. Accurate characterizations can improve the credibility of treatment recommendations and outcome assessments. Most clinicians perform gait assessment daily by simple observation, without technological assistance. With time, these observation skills and habits become second nature, and many O&P clinicians can relate to the experience of subconsciously observing the unique features and patterns in individuals' gaits outside of a clinical environment. I'm guessing that I'm not the only clinician who regularly critiques the gait patterns of the individuals I'm casually observing in a public setting. We may even be able to make reasonable assumptions based on our clinical experience about the underlying causes of the deviations we observe clandestinely.
The characteristics of many human physical features are unique, including fingerprints, handprints, voices, irises, and facial features. Biometrics can refer to the measurement of these and other human features and the use of that data for personal identification. Biometry is often implemented for security purposes, to deny or grant access to information or physical facilities. Behavioral biometrics involves the measurement of unique patterns of behavior. One example that is often cited is the way in which telegraph operators in the late 1800s could identify each other based on their unique patterns of sending the dots and dashes of Morse code. Since gait is a repetitive behavior that is influenced by individual physical characteristics, it has been investigated as a possible data source to confirm identity. One reason for this interest in gait is that the acquisition of many other forms of biometric data requires close physical proximity, while gait observation can be performed at a physical and temporal distance. Live video footage can be used to surveil individuals in a public space at a distance and without their knowledge. Recorded footage can be reviewed later, such as when reviewing CCTV footage after the commission of a crime. This article summarizes information related to personal identification based on gait patterns and discuss possibilities for the application of that technology in clinical practice.
Early in 2018, Connor and Ross published a survey titled "Biometric recognition by gait: A survey of modalities and features," which reviewed research on systems designed to identify individuals by their gait.1 The article is a thorough overview of the collection methods and analysis of gait data for personal identification, and a summary of the research related to the accuracy of those methods. The authors describe the history of gait assessment and distinguish between gait analysis (the area of interest for clinicians), gait forensics (the use of gait data in criminal investigation), and gait biometrics (personal identification of individuals based on gait-related data). They explain the use of 3D motion capture, accelerometers, underfoot pressure mats, and acoustics to identify unique gait patterns.
Video-based biometric systems can incorporate model-based or model-free approaches. As Connor and Ross describe in their article, model-based approaches involve matching a walking model, such as a skeletal structure, to video images and making computations based on the parameters of that model (Image 1). Model-free approaches evaluate features of gait directly from video images or from silhouette images that have been derived from those images. Model-based approaches have similarities with the types of video capture and analysis common in gait assessment systems used in many O&P practices (Image 2). For example, PnO Data Solutions allows practitioners to document kinematic data during key gait events by plotting lines on a still image captured from the video footage. Joint angles are calculated by the software and can be used to identify gait deviations and determine how orthotic or prosthetic intervention impacts those deviations. When recording gait in a clinical setting, the practitioner must take care to obtain footage that will result in accurate measurement and assessment. The recording is typically performed in a controlled environment with the camera properly positioned in relation to the subject. Video-based biometric identification systems, on the other hand, must be capable of using footage captured in real-world environments where motion may occur in poor lighting and at various angles from the recording device. The data analyzed by these systems includes not only joint angles, but also segment lengths and position trajectories over a series of steps. Complex calculations are required to match this information obtained from subjects to a database of known individuals.
As Connor and Ross explain, model-free approaches rely on the analysis of silhouettes of individuals while they walk (Image 3). After the silhouette of the individual has been separated from the background of the images, a variety of features of that silhouette are analyzed, including body segment lengths and widths, and physical dimensions such as height. Alternatively, features of the silhouette outline can be analyzed. Systems may use a sequence of silhouettes or an average silhouette for the analysis. It seems unlikely that practitioners are making detailed kinematic assessments while performing unassisted observation of gait in a clinical setting. It is more likely that we assess features of gait in a more holistic and intuitive manner. This type of assessment may involve identification of patterns of movement that are similar to the analysis of silhouettes used in model-free biometric systems. Just as individual players can be identified on a soccer field by observing their silhouettes framed against the horizon, experienced clinicians can identify that something is not right about a patient's gait without evaluating specific joint angles. We are often able to identify general features of pathological gait without describing those deviations in terms of specific goniometric estimates. We are more likely to say that a patient's knee flexion angle is too excessive or their ankle too plantarflexed than to attempt to assign a specific goniometric magnitude to those deviations.
Both model-based and model-free video-based biometric approaches rely on complex calculations to compare data points collected for a specific subject to a database of known subjects. According to Connor and Ross, "the most commonly reported metric in the gait biometric literature is the identification rate or the classification rate (CR)" that "indicates how likely a random test or ‘probe' sample is associated with the correct person in the labeled ‘gallery' based on the match scores generated."1 CRs for many model-based and model-free systems range from 80 to 100 percent. Errors occur by either failing to identify a subject or by incorrectly identifying them as someone else. Since controlling access is a key motivation for biometric identification, research has been conducted to determine whether gait identification systems can be fooled. According to the review, "purposely and repeatedly circumventing a gait recognition system by imitation is either very difficult or impossible." While it may not be possible to mimic another person's gait sufficiently to circumvent a security system, "it is possible to willfully avoid detection by gait" because we can change our gait pattern significantly. In other words, it is possible to make our gait significantly different than our normal gait pattern, but it is more difficult to imitate another person's gait pattern accurately enough to fool a gait identification system.
Identification of Pathological Gait
A recent review by Figueiredo et al. of "studies that employed machine learning algorithms…to automatically recognize the clinical status of human locomotion" specifically excluded image-based systems of gait identification and focused only on systems that collect biomechanical features of gait use sensor technology.2 The diagnoses of the subjects in the studies reviewed by Figueiredo et al. included Parkinson's disease, cerebral palsy, multiple sclerosis, and knee osteoarthritis. The systems reported on in this review were able to distinguish pathological gait from normal gait with a similar level of accuracy as the video-based biometric systems described earlier (80-100 percent). This review includes highly technical descriptions of data analysis that are beyond the scope of clinical practice, but the authors' rationale for the usefulness of this type of technology is worth considering. Like 3D gait assessment, these newer forms of biometric analysis could be useful for diagnosing pathological gait, recommending specific gait training strategies, planning treatment tailored to individual patient needs, and identifying patients' progress by comparing their gait pattern in follow-up sessions to previous baselines. For this purpose, wearable sensor technology offers a more practical and lower-cost option than a 3D gait lab. A group of German researchers demonstrated in 2015 that data gathered using a GAITRite pressure sensing walkway could accurately "identify pathological gait patterns and establish clinical diagnosis with good accuracy."3 Subjects in this study exhibited gait deviations as a result of cerebellar ataxia, phobic postural vertigo, bilateral vestibulopathy, and progressive supranuclear palsy. Both sensor and video-based technologies have significant advantages over instrumented walkways because of their portability and ability to collect data in everyday environments.
It is unlikely that biometric recognition of pathological gait using video-based systems will be feasible in clinical practice in the near future. Identifying nuances of the gait patterns of athletes on a soccer field and our patients in our facilities is a complex skill and replicating that skill across the range of human experience is complicated. However, it is intriguing to imagine systems that would allow clinical gait assessment using a combination of video footage recorded on smartphones and software systems that perform complex analyses similar to those previously mentioned. This technology could add credibility to treatment recommendations and may be helpful in clinical education. Experienced clinicians often struggle to communicate exactly what they are observing and what their experienced intuition tells them is wrong with a patient's gait. Difficulty communicating this information to students or residents can be a barrier to their learning gait assessment skills, and biometric technology could assist learners in acquiring those skills. This technology could also allow practitioners to perform an in-depth gait analysis remotely using video recorded by a patient, caretaker, or other health professional. In the meantime, practitioners and educators will continue to rely on visually observing gait and implementing video- and sensor-based technology that is already available to supplement that observation.
Even if computerized gait recognition eventually becomes clinically viable, it will still be necessary for a clinician to identify which deviations can and should be addressed and in what manner. Figueiredo et al. caution that "the classification of data is not necessarily equivalent to diagnosis, as there should be sufficient clinical evidence supporting such an argument in specific cases/conditions. This fact agrees with a major drawback of the described techniques, which is that they do not consider the subject's clinical history." Whatever future technology supports our clinical decision-making, assessment, documentation, and communication will ultimately remain the unique contribution of human practitioners.
John T. Brinkmann, MA, CPO/L, FAAOP(D), is an assistant professor at Northwestern University Prosthetics-Orthotics Center. He has more than 20 years of experience treating a wide variety of patients.
1. Connor, P., and A. Ross. 2018. Biometric recognition by gait: A survey of modalities and features. Computer Vision and Image Understanding 167:1-27.
2. Figueiredo, J., C. P. Santos, and J. C. Moreno. 2018. Automatic recognition of gait patterns in human motor disorders using machine learning: A review. Medical Engineering & Physics53:1-13.
3. Pradhan, C., M. Wuehr, and F. Akrami et al. 2015. Automated classification of neurological disorders of gait using spatio-temporal gait parameters. Journal of Electromyography and Kinesiology 25(2):413-22.