Diabetic Foot Ulcer Classification Models Using Artificial Intelligence and Machine Learning Techniques: Systematic Review.

Abstract

Background: Diabetes-related foot ulceration (DFU) is a common complication of diabetes, with a significant impact on survival, health care costs, and health-related quality of life. The prognosis of DFU varies widely among individuals. The International Working Group on the Diabetic Foot recently updated their guidelines on how to classify ulcers using "classical" classification and scoring systems. No system was recommended for individual prognostication, and the group considered that more detail in ulcer characterization was needed and that machine learning (ML)-based models may be the solution. Despite advances in the field, no assessment of available evidence was done.

Objective: This study aimed to identify and collect available evidence assessing the ability of ML-based models to predict clinical outcomes in people with DFU.

Methods: We searched the MEDLINE database (PubMed), Scopus, Web of Science, and IEEE Xplore for papers published up to July 2023. Studies were eligible if they were anterograde analytical studies that examined the prognostic abilities of ML models in predicting clinical outcomes in a population that included at least 80% of adults with DFU. The literature was screened independently by 2 investigators (MMS and DAR or EH in the first phase, and MMS and MAS in the second phase) for eligibility criteria and data extracted. The risk of bias was evaluated using the Quality In Prognosis Studies tool and the Prediction model Risk Of Bias Assessment Tool by 2 investigators (MMS and MAS) independently. A narrative synthesis was conducted.

Results: We retrieved a total of 2412 references after removing duplicates, of which 167 were subjected to full-text screening. Two references were added from searching relevant studies' lists of references. A total of 11 studies, comprising 13 papers, were included focusing on 3 outcomes: wound healing, lower extremity amputation, and mortality. Overall, 55 predictive models were created using mostly clinical characteristics, random forest as the developing method, and area under the receiver operating characteristic curve (AUROC) as a discrimination accuracy measure. AUROC varied from 0.56 to 0.94, with the majority of the models reporting an AUROC equal or superior to 0.8 but lacking 95% CIs. All studies were found to have a high risk of bias, mainly due to a lack of uniform variable definitions, outcome definitions and follow-up periods, insufficient sample sizes, and inadequate handling of missing data.

Conclusions: We identified several ML-based models predicting clinical outcomes with good discriminatory ability in people with DFU. Due to the focus on development and internal validation of the models, the proposal of several models in each study without selecting the "best one," and the use of nonexplainable techniques, the use of this type of model is clearly impaired. Future studies externally validating explainable models are needed so that ML models can become a reality in DFU care.

Trial registration: PROSPERO CRD42022308248; https://www.crd.york.ac.uk/PROSPERO/view/CRD42022308248.