Deep Learning for Early Scoliosis Detection Using mmWave Radar Gait Data
Trial statusNot yet recruiting
Trial phaseNot listed
Trial typeObservational
Biological sexAll
Age2-75
SponsorGebze Technical University
About this trial
Scoliosis is a sideways curvature of the spine that often develops during childhood and adolescence. When detected early, scoliosis can be managed effectively with non-invasive approaches such as bracing and physiotherapy, while late detection frequently leads to surgical intervention. Current screening methods rely on physical examination and X-ray imaging, which exposes children to ionizing radiation and may miss early-stage cases.
This observational study investigates whether millimeter-wave (mmWave) radar, combined with deep learning (a type of artificial intelligence), can detect early signs of scoliosis by analyzing how a child walks. The radar sensor records subtle movement patterns during walking without using cameras and without producing any identifiable images, fully preserving the participant's privacy. No ionizing radiation is involved.
Pediatric participants attending the orthopedic clinic for routine scoliosis evaluation are invited to walk a short distance in front of a mmWave radar sensor. The collected gait recordings are then analyzed using deep learning models, and the results are compared with the participant's standard clinical scoliosis assessment performed by a pediatric orthopedic specialist. The diagnostic performance of the deep learning model is evaluated using sensitivity, specificity, and overall accuracy.
If the approach proves accurate, it could offer a radiation-free, privacy-preserving, and low-cost alternative for early scoliosis screening in schools, primary healthcare centers, and pediatric orthopedic clinics, ultimately supporting earlier diagnosis and reducing the long-term clinical burden of untreated scoliosis.
Eligibility criteria
This trial accepts healthy volunteersQualifiers
Being between 2 and 75 years of age at the time of registration
Having applied to the pediatric orthopedics outpatient clinic for an assessment of suspected or known scoliosis
Being able to walk independently for at least 7 meters without assistive devices
Written informed consent from a parent or legal guardian
Disqualifiers
Severe scoliosis requiring urgent surgical intervention that prevents participation in walking tasks
Refusal to give informed consent or consent
Trial population
Pediatric and adolescent participants attending the pediatric orthopedic outpatient clinic at Başakşehir Çam and Sakura City Hospital for routine scoliosis evaluation. Consecutive eligible patients are invited to participate, including both those with confirmed scoliosis and those in whom scoliosis is ruled out following clinical and radiographic assessment.
Trial design
Design model
Cohort
Time perspective
Prospective
Treatments tested in this trial
mmWave Radar Gait Assessment
Diagnostic testEach participant performs a standardized walking task along a defined path in front of a millimeter-wave (mmWave) radar sensor. The radar continuously records the participant's gait micro-Doppler signatures during the walk. The mmWave radar device is contactless, non-ionizing, and does not capture identifiable visual images, fully preserving participant privacy. The recorded gait signals are subsequently processed and analyzed using deep learning models (including convolutional and transformer-based architectures) trained to classify scoliosis status. The full radar-based assessment takes approximately 5 to 10 minutes per participant. The standard clinical and radiographic scoliosis evaluation performed as part of routine care serves as the reference standard.
Treatment groups
200 Participants
are divided into 1 treatment groupGroup A: Pediatric Participants Undergoing Scoliosis Evaluation1 intervention
Trial outcomes
Primary outcomes
1
Diagnostic Accuracy of the mmWave Radar-Based Deep Learning Model for Scoliosis Detection (AUC-ROC)
The diagnostic performance of the mmWave radar-based deep learning classification model is assessed by the area under the receiver operating characteristic curve (AUC-ROC). The AUC-ROC is computed by comparing the model's predicted probability of scoliosis for each participant against the reference standard (clinical examination combined with Cobb angle measurement from standard-of-care radiographic imaging) on a held-out test set. The AUC-ROC is reported as a single numeric value between 0 and 1, with 95% confidence intervals.
Time frame
Assessed at the end of the data collection period, approximately 18 months after study startSecondary outcomes
1
Sensitivity and Specificity of the Deep Learning Model at the Optimal Operating Point
At the operating point that maximizes the Youden index on the validation set, the sensitivity (true positive rate) and specificity (true negative rate) of the deep learning model for classifying scoliosis status are calculated on the held-out test set. Both metrics are reported as percentages with 95% confidence intervals.
Time frame
Assessed at the end of the data collection period, approximately 18 months after study start2
Comparative Diagnostic Performance Across Deep Learning Architectures
The diagnostic accuracy (AUC-ROC) of multiple deep learning architectures, including convolutional neural networks (CNN), recurrent neural networks (RNN/LSTM), and transformer-based models, is compared on the same dataset using cross-validation. The architecture yielding the highest AUC-ROC is identified as the best-performing model.
Time frame
Assessed at the end of the data collection period, approximately 18 months after study start3
Stratified Diagnostic Performance by Scoliosis Severity (Cobb Angle Category)
The diagnostic accuracy of the deep learning model is evaluated separately for participants with mild scoliosis (Cobb angle 10 to 24 degrees), moderate scoliosis (Cobb angle 25 to 39 degrees), and severe scoliosis (Cobb angle 40 degrees or greater). Sensitivity is reported for each severity category to assess whether the model detects clinically significant curves of different magnitudes.
Time frame
Assessed at the end of the data collection period, approximately 18 months after study startSponsors and contacts
Click on the lead sponsor to view all of their trials.
Source ClinicalTrials.gov
This trial is not recruiting at the moment. You can still explore other options: