Work

Comparison Of A Novel Deep Learning Model With Healthcare Professionals In Chronic Wound Infection Detection

Year: 2026
Type: conference-paper
Abstract: PURPOSE: Chronic wounds affect 10.5 million Medicare beneficiaries with an annual cost of up to 50 billion dollars. With an aging population (20% >65 years by 2040) demand for wound care experts is predicted to increase. This will affect our rural population more which has a 50% higher rate of amputation. Numerous Artificial Intelligence solutions to this increased demand have been proposed but very few have been compared directly with healthcare professionals. SCARWID (Synthetic Caption Augmented Retrieval for Wound Infection Detection), a Deep Learning (DL) model, was created to identify wound infections in a remote setting without wound care expertise. We evaluated and compared SCARWID performance on infection detection in diabetic foot ulcer images (DFU) to healthcare professionals (HCP). METHODS: A total of 1,208 DFU images with wound care expert assigned ground truth labels (infected vs non-infected) were used. SCARWID was trained and evaluated with 5-fold cross-validation. The same image set was divided among HCPsmedical students (n=4), plastic surgery residents (n=5), emergency physicians (n=3), and plastic surgeons (n=4). Infection detection accuracy, sensitivity (SN), specificity (SP), and positive predictive value (PPV) were calculated for each group. Paired accuracy differences between each HCP group and SCARWID were estimated with 95% confidence intervals (CIs) and p-values. RESULTS: Across 5-fold cross-validation, SCARWID achieved 81% 1% accuracy, 85% 2% SN, 78% 1% SP, and 79% 1% PPV. On the same images HCP overall reached 64% 8% accuracy, 52% 21% SN, 74% 19% SP, and 72% 16% PPV. By subgroup, students averaged 68% 9% accuracy (SN 68% 16%, SP 70% 16%, PPV 70% 13%); residents 63% 7% accuracy (SN 35% 22%, SP 89% 14%, PPV 85% 15%); emergency physicians 57% 7% accuracy (SN 64% 6%, SP 50% 10%, PPV 54% 4%); and plastic surgeons 65% 8% accuracy (SN 48% 14%, SP 80% 12%, PPV 72% 12%). Paired accuracy differences (group − SCARWID) were −0.134 for students (95% CI −0.284 to +0.016; p=0.06), −0.188 for residents (95% CI −0.276 to −0.101; p=0.003), −0.245 for emergency physicians (95% CI −0.412 to −0.078; p=0.02), and −0.167 for plastic surgeons (95% CI −0.290 to −0.045; p=0.02). CONCLUSION: Overall, SCARWID outperformed every HCP group in accuracy of infection detection in DFU images. These results support the potential of DL models to augment infection triage where wound-care expertise is limited. *Source: https://ps-rc.org/meeting/Program/2026/53.cgi*
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7160656151
FWCI:
Cites:
Cited by:
Related to: 10