Evaluating risk factors affecting chronic wound healing: a comprehensive analysis using logistic regression and neural network models

Abstract

Chronic wounds impose immense burdens and suffering on patients, with healing governed by a complex blend of physiological, social, and psychological components. This study aimed to comprehensively evaluate risk factors affecting chronic wound healing by integrating traditional statistical methods with machine learning techniques. We retrospectively collected demographic and clinical data from 232 chronic wound patients treated at our hospital. The Mann-Whitney U test was employed for univariate comparisons of continuous variables, while categorical variables were analyzed using the chi-square or Fisher's exact test, and ordinal variables were assessed with the Cochran-Armitage trend test. Statistically significant parameters were subsequently incorporated into binary logistic regression and multilayer perceptron neural network (MLP) models. Univariate analysis identified seven factors significantly associated with wound prognosis: number of concurrent wounds ( P = 0.020), wound size ( P = 0.043), pain score ( P = 0.005), intervention modalities ( P for trend = 0.005), NSAIDs use ( P = 0.010), hemoglobin ( P = 0.005), and albumin ( P = 0.001). In the multivariable logistic regression model (event = good prognosis), albumin (adjusted OR=1.101, 95% CI: 1.024 -1.183, P = 0.010) and intervention with two modalities (OR=4.775, 95% CI: 1.215 -18.762, P = 0.025) or three modalities (OR=6.360, 95% CI: 1.404 -28.797, P = 0.016) were independently associated with good prognosis. The MLP model's AUC was recorded at 0.771 (95% CI: 0.710 -0.832), with wound size (100%) and albumin (69.3%) showing the highest normalized importance. The DeLong test showed no significant difference between the logistic regression and MLP models ( P = 0.907). Albumin levels and multimodal interventions were independently associated with favorable chronic wound prognosis, while wound size and albumin emerged as the most influential predictors in the neural network model. These findings underscore the importance of nutritional status and comprehensive wound management, although causal inference is limited by the retrospective design.