Artificial intelligence and model checking methods for in silico clinical trials

Model-based approaches to safety and efficacy assessment of pharmacological treatments (In Silico Clinical Trials, ISCT) hold the promise to decrease time and cost for the needed experimentations, reduce the need for animal and human testing, and enable personalised medicine, where treatments tailored for each single patient can be designed before being actually administered. Research in Virtual Physiological Human (VPH) is harvesting such promise by developing quantitative mechanistic models of patient physiology and drugs. Depending on many parameters, such models define physiological differences among different individuals and different reactions to drug administrations. Value assignments to model parameters can be regarded as Virtual Patients (VPs). Thus, as in vivo clinical trials test relevant drugs against suitable candidate patients, ISCT simulate effect of relevant drugs against VPs covering possible behaviours that might occur in vivo. Having a population of VPs representative of the whole spectrum of human patient behaviours is a key enabler of ISCT. However, VPH models of practical relevance are typically too complex to be solved analytically or to be formally analysed. Thus, they are usually solved numerically within simulators. In this setting, Artificial Intelligence and Model Checking methods are typically devised. Indeed, a VP coupled together with a pharmacological treatment represents a closed-loop model where the VP plays the role of a physical subsystem and the treatment strategy plays the role of the control software. Systems with this structure are known as Cyber-Physical Systems (CPSs). Thus, simulation-based methodologies for CPSs can be employed within personalised medicine in order to compute representative VP populations and to conduct ISCT. In this thesis, we advance the state of the art of simulation-based Artificial Intelligence and Model Checking methods for ISCT in the following directions. First, we present a Statistical Model Checking (SMC) methodology based on hypothesis testing that, given a VPH model as input, computes a population of VPs which is representative (i.e., large enough to represent all relevant phenotypes, with a given degree of statistical confidence) and stratified (i.e., organised as a multi-layer hierarchy of homogeneous sub-groups). Stratification allows ISCT to adaptively focus on specific phenotypes, also supporting prioritisation of patient sub-groups in follow-up in vivo clinical trials. Second, resting on a representative VP population, we design an ISCT aiming at optimising a complex treatment for a patient digital twin, that is the virtual counterpart of that patient physiology defined by means of a set of VPs. Our ISCT employs an intelligent search driving a VPH model simulator to seek the lightest but still effective treatment for the input patient digital twin. Third, to enable interoperability among VPH models defined with different modelling and simulation environments and to increase efficiency of our ISCT, we also design an optimised simulator driver to speed-up backtracking-based search algorithms driving simulators. Finally, we evaluate the effectiveness of our presented methodologies on state-of-the-art use cases and validate our results on retrospective clinical data.