Inside Multimodal Fusion for Wound Healing Risk: Combining Images with HbA1c, ABI, and…
healinganalytics
Image-only models miss half the story. See how multimodal fusion combines wound photos with labs and vascular data into a complete healing risk profile.
An image tells you what the wound looks like today. It does not tell you whether the patient's arterial supply, glycemic control, or nutrition will let it heal. Multimodal fusion — combining the image with structured clinical data — is where healing prediction becomes clinically useful.
The inputs
A production fusion model typically ingests:
- Image-derived features — area, depth, tissue composition, periwound status.
- Vascular — ankle-brachial index, toe pressures, transcutaneous oximetry.
- Metabolic — HbA1c, glucose variability, renal function.
- Nutritional — albumin, prealbumin, BMI trend.
- Comorbidities and meds — diabetes duration, smoking, immunosuppression, steroids.
- Wound history — duration, prior recurrences, prior amputation.
The architecture
Image features from a segmentation backbone are concatenated with a tabular encoder that handles the labs and history. A fusion head outputs a healing probability, a time-to-closure estimate, and — for lower-limb wounds — an amputation risk score.
Why fusion beats image-only
A clean-looking DFU on a patient with an ABI of 0.4 and HbA1c of 11 is a different wound than the same photo on a well-controlled patient with a palpable pedal pulse. Image-only models cannot see that. Fusion models push accuracy for twelve-week healing prediction from the mid-seventies into the mid-eighties on validated cohorts.
Where it lives in the workflow
Fusion only matters if the labs are actually in the chart. A purpose-built AI-powered EMR for wound care pulls ABI, HbA1c, and albumin from the record automatically, so the wound healing software scores every encounter without re-entry. For diabetic patients specifically, diabetic foot care workflows surface the amputation risk score at the point of decision.
The honest limits
Missing labs degrade the score. Old ABIs are worse than none. And risk stratification is a prompt to escalate — vascular referral, offloading, nutrition — not a prediction of destiny.