Diabetic foot disease affects 6% of people with diabetes and includes infection, sores, or destruction of tissues of the foot. The various forms of the diabetic foot disease impair patients quality of life and affect social participation. Between 0.03% and 1.5% of the population with diabetes may require some type of amputation, whereas among patients with diabetic foot ulcerations (DFU) or sores this may run as high as 25%. Most sores can be prevented with good foot care and screening for risk factors for a foot at risk of complications, such as lower leg amputation (LLA).
Scoring methods for prediction of lower-limb amputation (LLA) in patients with DFU perform relatively well in a hospital setting, but poor in a primary care setting. Furthermore, scores that predict development of LLA by site (toe, foot, lower limb below knee, lower limb above knee) do not exist. Additionally, current scores for risk of LLA are difficult to interpret. Therefore, the aim of this study is to develop a score for risk of LLA, broken down by limb (toe, foot, lower limb below knee, lower limb above knee) and mortality using traditional and sophisticated automated techniques.
The study objective is to develop prognostic scores for risk of lower limb amputation (LLA), broken down by limb (toe, foot, lower limb below knee, lower limb above knee), and mortality in patients with diabetic foot ulcerations (DFU) or patients who are at risk of developing DFU.
Two cohorts will be created: 1) patients with DFU and 2) patients at risk of developing a diabetic foot problem. In these cohorts patients aged 18 years or older with a registration of either DFU or diabetic feet at risk between April 2004 and time of data extraction will be included. The date of first registration will define start of follow-up. All patients will be followed up for the occurrence of an LLA, date of transfer out of practice area, death, end of data collection, or end of risk period (1, 5, 10, or 15 years), whichever comes first.
Both traditional methods and machine learning techniques will be applied to develop and validate the prognostic scores.
Traditional: Regression models will be fitted with a predefined set of determinants for LLA, using forward selection with a significance level of 0.05. Various measures for prognostic accuracy including sensitivity, specificity, positive predictive values, negative predictive values, and beta-coefficients of the included factors will be determined. The beta-coefficients in the final Cox models will be converted into 1-, 5-, 10-, and 15-year absolute risk scores. The models performance will be determined by assessing goodness-of-fit and by assessment of the discriminative ability. To asses internal validity a k-fold cross-validation will be done.
Machine learning: The study population will be split into a `training' cohort and a test cohort. The training cohort will be used to develop the prediction models using machine learning algorithms such as random forest (RF) and artificial neural networks (ANN). The models will be trained through 10-fold cross-validation. The test dataset will be used for model evaluation. Accuracy, sensitivity, specificity, receiver operating characteristics (ROC) curves and areas under the curve (AUC) will be determined.
Lower limb amputation (composite, and stratified by site); All-cause mortality
Frank de Vries - Chief Investigator - Utrecht University
Frank de Vries - Corresponding Applicant - Utrecht University
Johan Roikjer - Collaborator - Aalborg University Hospital
Johanna Driessen - Collaborator - Utrecht University
Johannes T.H. Nielen - Collaborator - Utrecht University
Joop van den Bergh - Collaborator - Maastricht University
Nicolaas Schaper - Collaborator - Maastricht University
Niels Ejskjaer - Collaborator - Aalborg University Hospital
Patrick Souverein - Collaborator - Utrecht University
Romin Pajouheshnia - Collaborator - Utrecht University
HES Admitted Patient Care