SRS95 - Using deep learning models (AI) on diabetic foot disease related MRI reports of the foot to identify salient predictors for limb loss

Abstract Introduction Diabetic foot disease (DFD) is a complex disease and is associated with lower limb amputation. Magnetic resonance imaging (MRI) is commonly used in patients with DFD. Aims To explore the use of deep learning models to identify predictors on MRI for limb loss in patients with DFD. Methods 419 MRI reports were reviewed to identify salient predictors for no amputation versus major versus minor limb loss from a list of features including collections, ulcer, osteomyelitis, diabetic myopathy/sarcopenia and Charcot foot. Random Forest, Extreme Gradient Boosting and a Multilayer Perceptron model was applied. Further composite analysis was undertaken to compare no amputation versus major lower limb amputation, no amputation versus all amputation and death and all amputation versus death. Results Extreme gradient boosting comparing the no amputation versus minor amputation versus major amputation group using extreme gradient boosting was the only model of statistical significance. The model had a 63.4% accuracy [95% c.i. 0.52–0.7; P-value < 0.001]. The ROC-AUC values were 0.724 (no amputation group), 0.779 (minor amputation group) and 0.439 (major amputation group). Features highlighted as potential predictors included absence of osteomyelitis and presence of Charcot as variables which influenced the outcome. Conclusions MRI reports may provide useful information in highlighting features which could potentially predict the risk of adverse outcomes. Uniformity in reporting may allow this to be explored further.