OsteoSight™ Hip (v1)

OsteoSight Hip (v1) is an AI-enabled software application that analyzes standard hip or pelvis X-rays in patients aged 50 and older to estimate bone mineral density at the femoral neck. It generates a report to alert radiologists and physicians to patients at possible risk of low bone mineral density, prompting further clinical assessment. The device is not intended to diagnose osteoporosis, replace DXA scanning, or rule out low BMD.

OsteoSight uses supervised machine learning to process DICOM images from standard anteroposterior hip/pelvis radiographs and produces a binary output (positive/negative for low BMD). Like its predicate (Rho), it operates via cloud integration with host systems, uses DICOM sources, delivers results via PACS, and consists of containerized microservices. The key difference is OsteoSight focuses solely on the femoral neck and hip/pelvis region, whereas Rho analyzes multiple anatomical sites (lumbar spine, thoracic spine, chest, pelvis, knee, hand/wrist).

OsteoSight was developed in accordance with ISO 14971 (risk management), BS EN 62366 (usability engineering), and IEC 62304 (software lifecycle). Software verification and validation demonstrate compliance with FDA guidance on 'Content of Premarket Submissions for Device Software Functions' at the basic documentation level.

OsteoSight is substantially equivalent to Rho because both are AI-enabled software devices using supervised ML to opportunistically identify patients at risk of low BMD from routine X-rays in patients ≥50 years old, delivered via PACS in clinical settings with equivalent intended users and indications. While OsteoSight is anatomically narrower (femoral neck only vs. multiple sites), this represents a subset of intended use rather than a fundamental difference in technology, function, or safety profile. Clinical performance validation demonstrates the device meets pre-specified sensitivity and specificity goals, with high specificity (0.943) and reasonable sensitivity (0.441) in populations where results were generated, comparable to predicate device performance standards.