Automated human body assessment
Machine Learning
Computer Vision
Digital Twins
- Our client is a US-based healthtech startup developing an application for smart human posture diagnosis and personal movement plan creation.
- The goal of the project is to automate core functionalities with a Machine Learning system.
- Our solution uses Deep Learning and mathematical algorithms to detect posture flaws and evaluate the patient’s silhouette by processing its 3D digital twin.
- We’re capable of detecting over 20 musculoskeletal disorders.
- We deliver a full ML cycle - from the process design to model training and evaluation to integrating deployable components with the client’s infrastructure.
Recent studies show over 60% prevalence of postural disorders for healthy 20 to 50-years-olds,' and it starts to affect all demographics, moving far beyond pure aesthetics.
Is there a way to tackle this problem without regular physiotherapeutic appointments? Our client believes there is. And we help turn that belief into reality.
Revolution in motion
We were contracted by a NY-based healthtech startup that brings together the best physiotherapists and orthopedists in the US. They wanted to build a system to offer professional and fully automated movement-oriented aid via a mobile app.
The application benefits not only those with postural disorders. The client wants to help anyone in a group of risk, from those with desk jobs to pro athletes looking for movement optimization guidance they can access anytime.
The application's core combines medical knowledge, machine learning proficiency, and modern 3D-scanning technology.
Physiotherapist in your pocket
It all starts with creating a digital twin of a patient's body. The scan-based silhouette model will serve as an input for diagnosis tools further in the process.
Sparkbit's team is responsible for the next step - the digitalization of posture diagnosis. Up until now, it was performed by a physiotherapist in his office. We replaced him (and the office) with a sequence of analyses performed by deep-learning 3D analysis models and mathematical algorithms we proposed.
We aim to assess, describe and evaluate the silhouette. We're currently capable of detecting over 20 different musculoskeletal disorders. Among others:
- lordosis and kyphosis associated with spine curvature,
- duck feet and pigeon toes based on feet position,
- valgus and varus stress connected with legs outline,
- leaning forward and backward, determined by measuring a center of mass.
As the last step, a set of algorithms assigns corrective exercises to help the patient. They build up to become a personalized movement plan that evolves with the patient's progress. Continuous therapy adjustments allowed by periodically performing new scans are what predestinates the product's upcoming success. It provides a defined and actionable path to full recovery.
The science of movement
Science is all about experimenting, and so is our part of the solution. We worked closely with the client's physiotherapists to evaluate the results of various experiments. The goal was to select the most promising ML models and proceed with them to reach the highest accuracy.
The most significant challenge we faced was insufficient model training data. How did we cope with it? Initially, we went creative with the dataset itself. As we’ve already had some scans, we could rig the models to feed the deep learning with new data made from the existing ones, however, it still wasn’t enough.
The answer came with research. Since we knew how to perform a "by the book" posture assessment, we could apply mathematical methods to it. We've proven they are more accurate than predictions offered by deep learning in certain cases. This way, we mitigated the problem of the time and resource costs of expanding the dataset with an up-and-coming heuristic method.
As our approach brought new business value to the product, the stakeholders decided that we should include it in our solution, combining mathematical algorithms with deep learning-based assessments.
The client is currently applying for a patent for our algorithms as they’re to be the core IP of the solution. The whole ML solution is also a part of the FDA approval application.
MLOps - the fundamental process
We provided a complete MLOps cycle for the project. This way, the client could focus on expanding the product’s functionalities and further application development while we designed and delivered the technology, seamlessly integrating it with the system.