OBJECTIVES: Our objective was to develop a prediction model for healing hard-to-heal venous leg ulcers within six months in US settings.

METHODS: The model was derived from a longitudinal retrospective cohort of patients from Net Health Wound Care web-based EHRs. Adult hard-to-heal venous leg ulcers patients were included, and patients with a cancer diagnosis, septic phlebitis, and deep vein thrombosis were excluded. We have used penalized logistic regressions shrinkage with the elastic net nested into the five-state model's health states with a cycle length of one month.

RESULTS: The final dataset contains 33,320 hard-to-heal leg ulcers and 140 variables. Dataset was split randomly into training (80%) and validating (20%) datasets. The beginning of the cycle is considered a landmark point, and regression was conducted at the patient at risk at the landmark point. Therefore, every health state and every landmark point have a unique set of predictors. Regression equations were derived for transition among all health states according to the natural history of ulcers). The main target for prediction was complete wound closure of reference wound (a cumulative probability for final state after 6-months). We have demonstrated the dynamic exchange of predictors over time and transitions among different health states. However, several predictors remain the same through time/health states, including ulcer size and duration and patient age. The area under the curve varied from 72 to 84% in the validation dataset.

CONCLUSIONS: We have demonstrated that risk prediction tools can be intuitive and clinically meaningful for practitioners even when complex dynamics of the wound healing process is modelled. Further, we have combined machine learning, biostatistical regression modelling and decision analysis, demonstrating that those approaches can be used in synergy, leading to optimal prediction performance. Further model validation in the external cohort is warranted.