J&J enlists AI to streamline joint replacement surgery
“It is a way for healthcare providers to coordinate with DePuy Synthes, the orthopaedics company of Johnson & Johnson, to enhance preoperative coordination by making it smoother, simpler, and smarter,” Swanson says.
“It is a way for healthcare providers to coordinate with DePuy Synthes, the orthopaedics company of Johnson & Johnson, to enhance preoperative coordination by making it smoother, simpler, and smarter,” Swanson says.
Developing a digitally integrated solution
J&J’s ACM team developed the solution in a true agile fashion, with a shippable product available after every sprint, Swanson says.
The system — which leverages a suite of digital tools, including HL7 EMR/DICOM industry-standard integrations, a NodeJS and React web app, an open API architecture, and microservices to integrate with external partners — makes use of in-house Redshift machine-learning algorithms to predict hip and knee implant sizing using patient biometrics in addition to X-ray images. ACM also includes a robust portal that offers case analytics around product utilization, surgery metrics, upcoming case schedules, case details, and templating insights.
“The primary data source is ACM’s internal case report, though we also rely on orthopaedic supply data from Mercy, a not-for-profit Catholic healthcare organization,” Swanson says. “For ACM case reports, data transfer occurs from a high-trust data environment and an automated pipeline is built to provide the most recent data to the data science team through Redshift tables. The data includes patient biometric information such as height, weight, age, and gender. These features tend to be available in most EMR [electronic medical record] systems and thus allow for scale of the solution.”
To predict component size for TKA and THA surgeries, J&J implemented a range of machine learning techniques. “After cleaning and harmonizing the data in terms of units and metrics, we develop multi-class classification models to predict component size specific to each brand and component. The primary prediction algorithm in production is ordinal logistic regression and different techniques are used to deal with the class imbalance problem (stratified sampling, SMOTE, etc.),” Swanson says.
Completed algorithms are shared via Amazon Web Services S3 infrastructure with the SC EMR IT team. Results are visualized in a Tableau dashboard for business stakeholders to track accuracy over time. Models are retrained approximately once per business quarter.
“The encrypted case schedule and patient information are electronically transferred securely. The data is used to determine the specification of the DePuy Synthes implant range and the instrument set for each surgery,” says Swanson.
Overcoming obstacles along the way
Swanson and his team had to deal with a few challenges while developing ACM. The two major ones were getting commercial buy-in to accelerate adoption of the solution and ensuring its effective marketing as there were distinct regional go-to-market processes that challenged the backlog prioritization.
About Author
