The prevalence of atherosclerotic peripheral artery disease (PAD) is estimated around 200 million people worldwide. The clinical subcategories may vary from no symptoms to irreversible tissue, sensory, and motor loss in the legs. The main treatment is revascularization surgery: reopening the blood vessel or bypassing the diseased region. A key indicator to the success of a revascularization surgery is the perfusion i.e., the way blood flows in and out the region of interest. Near-infrared fluorescence dynamic imaging seems to offer a qualitative indication of blood perfusion [1]. This technique, however, lacks objectivity in its interpretation. We here present a digital twin that includes both mechanistic and data-driven approaches and aims at quantitatively assessing the success of a revascularization surgery in PAD patients, based on the fluorescent dynamic imaging time-varying signal. First, a mechanistic model is built by applying reduced hemodynamics and pharmacokinetic methods and aims at representing the patient’s state. The key body parts are represented as compartments connected through the arterial and veinous systems. The blood circulation is lumped in each compartment adopting the electric-hydraulic analogy; the heart being represented as a blood flow generator. Transport of the injected fluorescent tracer is simulated on top of hemodynamics. Each compartment transforms the received concentration with a certain transfer function that depends on the geometry of the vascular bed. The model reproduces the clinically observed signal behavior with good precision. The biomarkers identified by surgeons in their clinical practice as characterizing the patient’s feet perfusion are selected to assess the sensitivity of the mechanistic model to the patient’s characteristics. A variance-based global sensitivity analysis highlights the most impactful parameters: the outcome is coherent with clinical practice. Following the guidelines presented in [2], synthetic intensity signals are generated from the mechanistic model and are fed to an artificial neural network that aims at predicting the state of the patient. The chosen architecture is a long-short term memory recurrent neural network. This permits to take into account the correlation between consecutive time steps (short term) and all the previous ones (long-term memory). The neural network successfully recovers the patient’s state with satisfying precision. In particular, blood perfusion may ...