Biomechanical analysis has historically been trapped within the "gold standard" laboratory. To accurately measure how a person moves, researchers have traditionally required hundreds of thousands of dollars in equipment: infrared multi-camera arrays for 3D motion capture, floor-embedded force plates to measure ground reaction forces, and synchronized electromyography (EMG) to track muscle activation.
While these tools provide high-fidelity data, they are immobile, expensive, and require specialized staff. This creates a "laboratory gap" where advanced diagnostics are unavailable in the places they are needed most—geriatric wards, local sports fields, and patients' homes. A growing body of research is now bridging this gap by using machine learning to extract laboratory-grade insights from ubiquitous, low-cost sensors like smartphones, wearable inertial measurement units (IMUs), and standard 2D video.
The Problem of the Laboratory Gap
The primary barrier to widespread movement analysis is the dichotomy between subjective observation and resource-intensive measurement. In clinical settings, practitioners often rely on visual scales or simple stopwatches, which David Ahmedt-Aristizabal and colleagues note are prone to human error and lack the resolution to detect subtle pathological changes Computer Vision for Clinical Gait Analysis: A Gait Abnormality Video Dataset.
Conversely, high-precision tools like pressure-sensitive walkways or 3D optical systems are restricted to specialized environments. As Johannes C. M. Schlachetzki points out, this lack of portability prevents the collection of objective data during routine medical appointments or in a patient's daily life Wearable sensors objectively measure gait parameters in Parkinson’s disease.
From Pixels to Kinematics
The first step in democratizing biomechanics is replacing expensive 3D cameras with ubiquitous 2D video. This transition relies on pose estimation algorithms—neural networks trained to identify human "keypoints" (like the ankle, knee, and hip) in standard RGB frames.
Researchers like Ziyu Gong have demonstrated that frameworks like OpenPose can be used to assess injury risks, such as Anterior Cruciate Ligament (ACL) tears, using nothing more than a smartphone Assessment of Anterior Cruciate Ligament Injury Risk Based on Human Key Points Detection Algorithm. Similarly, Fritz Webering showed that by analyzing the physics of a body in free-fall through video, a system can self-calibrate to measure vertical jump height without any physical markers Markerless Camera-Based Vertical Jump Height Measurement Using OpenPose.
While 2D video is excellent at timing events (like when a foot hits the ground), it historically struggled with 3D spatial accuracy. To solve this, researchers like Sina Mehdizadeh and Hoda Nabavi are validating deep-learning-based pose trackers against gold-standard 3D systems to identify exactly where these "sensor-light" methods succeed and where they require refinement Concurrent validity of human pose tracking in video for measuring gait parameters in older adults.
Inferring Kinetics from Kinematics
The most significant breakthrough in this field is the ability to estimate kinetics (forces and moments) without force plates. This is a "blind" estimation problem: can we guess how hard the foot is hitting the ground just by looking at how the leg moves?
Machine learning makes this possible by capturing the non-linear relationship between motion and force. For example:
- Joint Moments: Archit Kambhamettu and Samuel Audia developed VidSole, which combines instrumented insoles with video to estimate the Knee Adduction Moment (KAM), a critical biomarker for osteoarthritis, with clinical-grade accuracy VidSole: A Multimodal Dataset for Joint Kinetics Quantification and Disease Detection with Deep Learning.
- Ground Reaction Forces: Jesse Davis and colleagues used a single sacral-mounted sensor to predict running impact forces, allowing for "lab-grade" injury monitoring in the real world Predicting vertical ground reaction force characteristics during running with machine learning.
- Exoskeleton Control: Clément Lhoste and Emek Baris Küçüktabak used Long Short-Term Memory (LSTM) networks to estimate weight distribution in robotic exoskeletons using only joint angles, eliminating the need for fragile external force sensors Deep-Learning Estimation of Weight Distribution Using Joint Kinematics for Lower-Limb Exoskeleton Control.
Gait as a Digital Biomarker
By moving analysis out of the lab, researchers are transforming movement into a digital biomarker for neurological and systemic health. Because walking requires complex coordination between the brain, nerves, and muscles, subtle irregularities in gait can signal disease progression long before other symptoms appear.
Heiko Gaßner and Dennis Jensen have used wearable sensors to track gait variability as a marker of disease severity in Huntington’s Disease Gait variability as digital biomarker of disease severity in Huntington’s disease. Similarly, Yanmin Tang and colleagues are using vision-based AI to predict cognitive decline in patients with "silent" cerebrovascular disease Association between gait features assessed by artificial intelligent system and cognitive function decline.
The ultimate goal of this research direction, as seen in the work of Robert J. Ellis on the SmartMOVE app, is to provide personalized, real-time therapy—such as auditory gait cues—delivered through the same smartphones used to measure the impairment A Validated Smartphone-Based Assessment of Gait and Gait Variability in Parkinson’s Disease. By translating high-fidelity biomechanics to ubiquitous sensors, the field is moving toward a future where "the lab" is wherever the patient happens to be.