Using Artificial Intelligence (AI) to Model Wound Healing Prediction – a Preliminary Study

Introduction: Accurate wound healing prediction can help guide treatment decisions by early identification of patients presenting non-healing risk. Prior efforts include the development of Wound Healing Index for diabetic foot ulcers.  Such prediction and risk stratification systems demonstrate some efficacy, but they exclude vital wound image information and associated visual details.

Purpose: Using data set from a chronic wound data repository, we apply the latest developments in deep learning to develop a healing prediction model integrating wound images.

Method: The proposed deep learning framework combines wound images and patient demographics data to develop a healing prediction model utilizing Long Short-Term Memory (LSTM) network. An open source Python deep learning library was employed to develop the machine-learning framework. The model, a combination of Deep Convolutional Network and Deep Neural Network, was trained using the Adam optimizer and categorical cross entropy as the loss function. The model output is the predicted healing trajectory.

Result: We tested the proposed model on an input data source comprised of 4 weeks of longitudinal data from the chronic wound data repository. The predicted and actual healing trajectories are compared by way of the Pompeiu-Hausdorff (P-H) distance, which is a measure of trajectory similarity (where zero would denote two identical trajectories). The model tested here achieves a P-H distance of 1.6 cm2. 

Conclusion: The proposed novel LSTM network enables the use of wound image and patient demographic data to perform robust sequence learning. This early application of artificial intelligence may enable more robust healing prediction, thereby facilitating better risk stratification and, subsequently, treatment decisions. Future work will focus on improving the robustness and generalizability of the hybrid LSTM cell with more images and data augmentation.