Dynamic Assessment of Frailty Status in Outpatients with Diabetic Foot Ulcer and Personalized Artificial Intelligence-Driven Intervention Recommendation System

Patients with diabetic foot ulcer (DFU) often suffer from frailty, and the frailty status can evolve dynamically. However, there is a lack of simple dynamic assessment and personalized intervention tools in the outpatient setting. This work aimed to construct a dynamic assessment system for frailty status in outpatients with DFU, analyze the frailty evolution trajectories and key influencing factors, develop a personalized intervention recommendation system based on machine learning and explainable artificial intelligence, and preliminarily validate its feasibility. This study retrospectively included 372 patients with diabetic foot ulcers (DFUs) who visited the wound care clinic of a tertiary hospital between January 2023 and January 2025, The FRAIL scale (Fatigue, Resistance, Ambulation, Illnesses, Loss of weight) scores and clinical indicators were collected at baseline (T0) and three follow up time points (T1: 30-60 days, T2: 90-150 days, T3: 180-240 days). Generalized estimating equations (GEE) and latent class mixed models (LCMM) were used to analyze dynamic frailty evolution trajectories. Three algorithms (XGBoost, Random Forest, LightGBM) were compared for predicting trajectory classes, with SHAP for model interpretation. A personalized intervention recommendation system was constructed by integrating a knowledge base, and its feasibility was preliminarily verified using expert ratings based on 56 patients. During the 7 month follow up, the proportion of frail patients increased from 48.7% to 63.8%. In the development cohort (n = 316), LCMM identified four trajectories: sustained robust (11.1%), slowly worsening (32.3%), rapidly worsening (29.1%), and improving (27.5%). GEE analysis showed that age (𝛽 = 0.04), Wagner grade (𝛽 = 0.38), and dressing change pain score (𝛽 = 0.07) were independent risk factors for frailty worsening (all P < 0.05). The macro-average area under the receiver operating characteristic curve (AUC) of XGBoost on the validation set and the independent test set were 0.918 and 0.898, respectively, but the independent test accuracy and macro-average F1 dropped to 0.589 and 0.569. SHAP analysis indicated that Wagner grade, age, and dressing change pain score contributed the most to the predictions. The average expert rating for the personalized intervention recommendation system was 4.31 (95% CI: 4.23-4.39), with an intraclass correlation coefficient of 0.83 (95% CI: 0.76-0.88). These results suggest that frailty in outpatients with DFU exhibits dynamic divergent trajectories; XGBoost combined with SHAP can effectively predict frailty progression; and the personalized intervention recommendation system is preliminarily feasible, although its clinical value still requires further validation through multicenter, prospective studies.