Researchers found automated foot strike detection is viable for smartphone-based fall risk classification in patients with lower limb amputations.
Foot strikes automatically detected by a smartphone attached to the body during a 6-minute walk test (6MWT) can be used to calculate step-based features such as fall risk for patients with lower limb amputations, according to research published in PLOS Digital Health.
In this study, a smartphone was attached to the participant’s belt and specifically positioned at their lower back while the participant completed a 6MWT along a 20-meter hallway.
The session was video recorded and accelerometer, gyroscope, and smartphone orientation data were collected using The Ottawa Hospital Rehabilitation Centre (TOHRC) Walk Test app at 50 Hz. These data and timestamps for the video recordings were imported into MATLAB 2020b.
The authors then re-interpolated smartphone signals to 50Hz using linear interpolation, and afterwards applied a fourth-order zero-lag Butterworth low pass filter with a cut-off frequency of 4 Hz.
A random forest model has been shown to be effective at classifying fall risk among patients with lower limb amputations, although manual labelling of foot strikes was required.
The authors’ goal in the current study was to determine whether using the random forest model and an automated foot strike detection approach would result in similar fall risk classification results. Step-based features were determined using manually-labelled or automated foot strikes.
Inclusion criteria for the study were:
After exclusions, the study included 80 patients recruited from the University Rehabilitation Institute with transtibial, transfemoral, or bilateral lower limb amputations. Most (78.8%) participants were male and the mean (SD) age was 64.2 (12.2) years, with ages ranging between 19 and 90 years.
The most common type of amputation was transtibial (90%), with a low number of participants having a bilateral (6.2%) or transfemoral (3.8%) lower limb amputation.
The authors found manually-labelled foot strikes correctly classified fall risk for 64 of the 80 participants, reflecting an 80% accuracy, 55.6% sensitivity, and 92.5% specificity in classification.
Automated foot strikes performed at a slightly lower quality, correctly classifying fall risk for 58 participants. This reflects an accuracy of 72.5%, an equal sensitivity of 55.6%, and a specificity of 81.1%. The automated foot strike method also had 6 more false positives compared with the manual label method, classifying non-fall risk as fall risk.
Despite this, the authors determined the classifications were similar enough between approaches for the automated method to be viable for smartphone-based fall risk classification in patients with lower limb amputations.
“Automated foot strike detection is necessary in a clinical environment, where timely manual labelling is not feasible,” the authors said. “A smartphone-based fall risk classification model from a 6MWT can benefit the patient and clinician since one assessment with a single sensor placement can provide functional capacity, stride parameters, and fall risk information to aid clinical decision-making.”
The authors also noted limitations in this research, including small sample size and little variety in type of amputation. With this in mind, they said additional research is warranted and would benefit from a greater number of patients with transfemoral or bilateral amputations and sub-group analysis.
Reference
Juneau P, Lemaire ED, Bavec A, Burger H, Baddour N. Automated step detection with 6-minute walk test smartphone sensors signals for fall risk classification in lower limb amputees. PLOS Digit Health. Published online August 18, 2022. doi:10.1371/journal.pdig.0000088
Reducing Readmissions in the Safety Net Through AI and Automation
March 6th 2025Artificial intelligence (AI) and electronic health record–based automation tools helped a safety-net health system meet performance-based readmission metrics, thereby retaining critical funding while improving clinical and equity outcomes.
Read More
Health Equity in the Era of Large Language Models
March 5th 2025This article presents challenges and solutions regarding health care–focused large language models (LLMs) and summarizes key recommendations from major regulatory and governance bodies for LLM development, implementation, and maintenance.
Read More
More than 5 years after the American Society of Clinical Oncology warned of emerging disparities in precision medicine, efforts by community practices to embrace technology, form partnerships, and use data show how patients can gain access to personalized approaches. But challenges remain, especially for those covered by Medicaid.
Read More