Abstract Body: Objective: Patients with peripheral artery disease (PAD) are at increased risk for major adverse limb events (MALE) and major adverse cardiovascular events (MACE). To date, machine learning models for predicting clinical outcomes in PAD have primarily relied on patient-level factors, such as demographics and medical history, and the potential additive value of image-derived features for outcome prediction in PAD remains understudied. Therefore, using a machine learning approach, we sought to evaluate if incorporation of CT imaging features (arterial calcium burden and calf muscle density) could improve 12-month outcome discrimination when combined with patient-level factors in a clinical cohort of PAD patients.
Methods: Patients with PAD (n=90) were prospectively enrolled for non-contrast CT imaging of the lower extremities. The femoral–popliteal artery was manually segmented on axial CT images, and a threshold of >130 Hounsfield units was applied within the arterial volume of interest to identify and quantify calcium density. Mean muscle densities were obtained from the gastrocnemius, soleus, and anterior tibial muscles segmented across the entire calf. Patient-level factors and 12-month outcome data were obtained through chart review. All variables were preprocessed prior to model development. An elastic net machine learning model was trained for outcome discrimination by first using patient-level factors and secondly with imaging features added. Model performance was evaluated using three-fold cross-validation, with discrimination quantified by the area under the receiver operating characteristic curve (ROC-AUC).
Results: A combined total of 21 adverse events were recorded in the 12 months after CT imaging, including 12 MALE (13.33%) and 9 MACE (10.0%). The patient-level factor only model achieved an ROC-AUC of 0.52 ± 0.08. The addition of image-derived measures of arterial calcium burden and calf muscle density to the model demonstrated a robust improvement in outcome discrimination, as reflected by an ROC-AUC of 0.72 ± 0.03.
Conclusions: Predictive modeling of risk for adverse events in patients with PAD can be improved with the addition of image-derived features including arterial calcium burden and regional skeletal muscle density.