A novel classifier of radiographic knee osteoarthritis for use on knee DXA images is predictive of joint replacement in UK Biobank

Objectives DXA scans may offer a novel means of evaluating radiographic knee osteoarthritis (rKOA) in large population studies and through opportunistic screening. We aimed to develop and apply a semi-automated method for assessing rKOA using ∼20,000 knee DXA images from UK Biobank (UKB) and assess its face validity by checking for expected relationships with clinical outcomes. Methods Right knee DXA scans were manually annotated for osteophytes to derive corresponding grades. Joint space narrowing (JSN) grades in the medial joint compartment were determined from automatically measured minimum joint space width. Overall rKOA grade (0-4) was determined by combining osteophyte and JSN grades. Logistic regression was employed to investigate the associations of osteophyte, JSN, and rKOA grades with knee pain and hospital-diagnosed knee osteoarthritis (HES-KOA). Cox proportional hazards modelling was used to examine the associations of these variables with risk of subsequent total knee replacement (TKR). Results Of the 19,595 participants included (mean age: 63.7), 19.5% had rKOA grade ≥1 (26.1% female; 12.5% male). Grade ≥1 osteophytes and grade ≥1 JSN were associated with knee pain, HES-KOA, and TKR. Higher rKOA grades were linked to stronger associations with these clinical outcomes, with the most pronounced effects observed for TKR. HRs for the association of rKOA grades with TKR were 3.28, 8.75, and 28.63 for grades 1, 2 and 3-4, respectively. Conclusions Our DXA-derived measure of rKOA demonstrated a progressive relationship with clinical outcomes. These findings support the use of DXA for classifying rKOA in large epidemiological studies and in future population-based screening. Key messages Radiographic knee osteoarthritis (rKOA) can be semi-automatically derived from DXA images. DXA-derived rKOA shows expected relationships with clinical outcomes of knee osteoarthritis. DXA imaging presents a viable method for classifying rKOA in large-scale epidemiological research.