Back to 2026 Abstracts
Non-Invasive Prediction of Wound Healing Timelines Using AI-Driven Biomechanical Analysis
Robert D. Sampson
*2, Jayesh Rathi
3, Deboparna Banerjee
3, Shi Fu
1, Gurtej Singh
2, Hugh Rosshirt
1, Aleksandra Krajewski
2, Alvin Wong
2, Steven Skiena
3, Lav Varshney
3, Miriam Rafailovich
1, Sami Khan
21Department of Materials Science and Chemical Engineering, Stony Brook University, Stony Brook, NY; 2Division of Plastic and Reconstructive Surgery, Stony Brook University Hospital, Stony Brook, NY; 3AI Innovation Institute, Stony Brook University, Stony Brook, NY
PURPOSE: Clinicians lack objective tools to predict when a wound will heal, forcing reactive management that delays intervention. Existing AI wound tools trained on color features show reduced accuracy on darker skin tones, especially as underserved populations bear a disproportionate chronic wound burden. We developed an AI-driven platform that predicts healing day from biomechanical data using digital image speckle correlation (DISC), with a detection pipeline independent of skin pigmentation.
METHODS: DISC tracks displacement of natural skin features across video frames during controlled point loading to quantify force propagation through healing tissue. We automated wound boundary detection using two zero-shot AI foundation models (Grounding DINO, SAM) that rely on gradient-based intensity patterns rather than color, enabling detection across all skin tones without wound-specific training. Using a 28-day porcine burn model (six groups, two timepoints), we trained eight ML architectures on five features: wound closure, DISC-derived force propagation ratio, applied force, treatment, and application timing (n=614, 10 repeated 80/20 splits).
RESULTS: All eight models predicted healing day with approximately 90% accuracy within one week, outperforming baseline (MAE 5.55 days). Accuracy ranged from 85-91% across linear, tree-based, kernel, and ensemble methods, confirming the signal is model-agnostic. The best model (KNN) achieved MAE 3.39±0.25 days and 90.6±2.6% within 7-day accuracy. Feature importance identified wound closure (48.4%) and force propagation ratio (17.6%) as dominant predictors. Ablation confirmed force propagation adds predictive value beyond visual closure alone.
CONCLUSIONS: DISC biomechanical data predict healing day with about 90% accuracy within one week. Because the system detects tissue boundaries via mechanical displacement rather than color, it avoids known bias in medical imaging AI. These results support DISC as a platform for remote wound monitoring, burn zone prediction, debridement timing, and equitable assessment across skin tones.
Back to 2026 Abstracts
