Unlocking health insights: Estimating advanced walking metrics with smartwatches

Gait metrics — measures like walking speed, step length, and double support time (i.e., the proportion of gait cycle when both feet are on the ground) — are known to be vital biomarkers for assessing a person’s overall health, risk of falling, and progression of neurological or musculoskeletal conditions. Analyzing how a person walks, known as gait analysis, offers valuable, non-invasive insights into general well-being, injuries, and health concerns.

Historically, measuring gait required expensive, specialized laboratory equipment, making continuous tracking impractical. While smartphones now offer a portable alternative using their embedded inertial measurement units (IMUs), they demand precise placement — such as a thigh pocket or belt — for the most accurate results. In contrast, smartwatches are worn on the wrist in a fixed location. This provides a much more practical and consistent platform for continuous tracking, even expanding the tracking window to phone-less scenarios like walking around the house.

Despite this crucial logistical advantage, smartwatches have historically lagged behind smartphones in comprehensive gait metric evaluation. In our work, "Smartwatch-Based Walking Metrics Estimation", we sought to bridge this gap. We demonstrated that consumer smartwatches are a highly viable, accurate, and reliable platform for estimating a comprehensive suite of spatio-temporal gait metrics, with performance comparable to smartphone-based methods.

To achieve this, we developed a multi-output (i.e., multi-head) deep learning model built on a temporal convolutional network (TCN) architecture identical for both smartwatch and smartphone data. This multi-head model is a key differentiator from prior TCN-based approaches, which often only provide temporal events (like contact points) that require drift-prone integration for spatial metrics like step length and gait speed. Our model, in contrast, directly estimates all spatio-temporal gait metrics.

Our model takes two key inputs, user height (a single scalar demographic input) and raw IMU signals, which include 3-axis accelerometer and 3-axis gyroscope data from a single on-wrist Pixel Watch at 50 Hz sampling frequency. The model architecture extracts embeddings from the IMU sensor input, which are then concatenated with the demographic data before the final prediction layers. The multi-head model output then directly estimates a comprehensive suite of measures, including bilateral metrics, one head each for gait speed and double support time, and unilateral metrics, two heads each (one for left and one for right foot) for step length, swing time, and stance time. Definitions for each of these metrics are defined as:

• Gait speed: Distance an individual travels divided by the time taken (in cm/s)
• Double support time: Proportion of gait cycle when both feet are on the ground (in %)
• Step length (unilateral): Distance from the initial contact of one foot to the initial contact of the other foot (in cm)
• Swing time (unilateral): Duration within the gait cycle when the foot is not in contact with the ground (in ms)
• Stance time (unilateral): Duration within the gait cycle when the foot is in contact with the ground (in ms)

For data segmentation, we used 5-second windows with a 1-second overlap. We utilized mean absolute percentage error (MAPE) for the loss function, which uniquely optimizes for the relative accuracy across all multi-unit outputs (e.g., step length in cm, double support time in ms).

To rigorously evaluate the model, we conducted a large-scale validation study featuring a large cohort of 246 participants and approximately 70,000 walking segments. Participants were screened to be over 18, not using assistive devices, and without balance- or gait-affecting conditions. Data was collected from two international cohorts: Google in Mountain View, California and Kyoto University in Japan.

For the reference (ground truth) measurements, we used a lab-grade Zeno Gait Walkway system. Participants were outfitted with a Pixel Watch 1 on each wrist and four Pixel 6 phones placed in the front pocket, back pocket, backpack, and a cross-body bag.

The study protocol included a diverse range of walking patterns to ensure comprehensive evaluation:

• Six minute walk test (6MWT): Complete loops along the track at a self-paced speed
• Fast pace walking: Walk at a comfortably fast but steady pace
• Mild and moderate asymmetry: Walk self-paced while wearing a hinged knee brace locked into specific flexion/extension angles

The smartwatch model was trained using data from both wrist-worn devices, while the smartphone model's testing phases exclusively utilized data from front and back pocket phone placements, given their expected prevalence and highest accuracy. We employed a five-fold cross-validation strategy to maximize the test cohort and prevent data leakage by assigning all data from a single participant to a single split.

The results collectively demonstrated the accuracy, correlation, and reliability of the smartwatch-based method, showing comparable performance to smartphone estimates, despite the smartphone model being trained with approximately two times more data segments.

1. Strong validity and excellent reliability: Across the 70,000 walking segments, smartwatch estimates demonstrated strong validity (Pearson r >0.80) and excellent reliability (intraclass correlation coefficient (ICC) >0.80) for most metrics, including gait speed, step length, swing time, and stance time. Double support time showed moderately lower but acceptable ICCs (0.56−0.60) with narrow 95% confidence intervals, underscoring reliability.
2. Comparable to Smartphone Estimates: A quantitative comparison of the MAPE (see below) and mean absolute error (MAE) showed non-significant differences (p >0.05) between the Pixel Watch and Pixel phone across all measured gait metrics. This establishes the smartwatch as a viable and highly comparable platform for accurate gait analysis. Both the smartwatch and smartphone models significantly outperformed a naïve estimator, which merely predicted the mean of the training samples.
3. The role of user height: An ablation study confirmed that providing user height significantly improved the smartwatch's accuracy for estimating gait speed and step length, which highlights the distinct importance of user height for wrist-based gait metric estimation.

These findings are a major step in establishing the ubiquitous on-wrist smartwatch as a foundational technology for accurate and reliable gait-based health tracking. By bringing comprehensive gait analysis out of the lab and onto the wrist, we can enable:

• Continuous and accessible tracking: Longitudinal monitoring of gait metrics outside traditional clinical and laboratory settings
• Early detection and prevention: Greater potential for the early detection of disease, fall risk, and personalized rehabilitation planning.

The smartwatch offers a practical and consistent platform for health tracking that overcomes the placement issues associated with smartphones. Our continued work will explore refining and expanding the suite of metrics to maximize the utility of smartwatches in proactive health tracking and recommendations.

The following researchers contributed to this work: Amir B. Farjadian, Shun Liao[e3e428], Alicia Y. Kokoszka, Kyle DeHolton, Jonathan Hsu, Jonathan Wang, Lawrence Cai, Mark Malhotra, Shwetak Patel, Anupam Pathak, Ming-Zher Poh.