Abstract: Model-based approaches to safety and efficacy assessment of pharmacological drugs, treatment strategies, or medical devices (In Silico Clinical Trial, ISCT) aim to decrease time and cost for the needed experimentations, reduce animal and human testing, and enable precision medicine. Unfortunately, in presence of non-identifiable models (e.g., reaction networks), parameter estimation is not enough to generate complete populations of Virtual Patient (VPs), i.e., populations guaranteed to show the entire spectrum of model behaviours (phenotypes), thus ensuring representativeness of the trial.We present methods and software based on global search driven by statistical model checking that, starting from a (non-identifiable) quantitative model of the human physiology (plus drugs PK/PD) and suitable biological and medical knowledge elicited from experts, compute a population of VPs whose behaviours are representative of the whole spectrum of phenotypes entailed by the model (completeness) and pairwise distinguishable according to user-provided criteria. This enables full granularity control on the size of the population to employ in an ISCT, guaranteeing representativeness while avoiding over-representation of behaviours.We proved the effectiveness of our algorithm on a non-identifiable ODE-based model of the female Hypothalamic-Pituitary-Gonadal axis, by generating a population of 4 830 264 VPs stratified into 7 levels (at different granularity of behaviours), and assessed its representativeness against 86 retrospective health records from Pfizer, Hannover Medical School and University Hospital of Lausanne. The datasets are respectively covered by our VPs within Average Normalised Mean Absolute Error of 15%, 20%, and 35% (90% of the latter dataset is covered within 20% error).

Abstract: Advances in the use of neuroimaging in combination with A.I., and specifically the use of machine learning techniques, have led to the development of brain-reading technologies which, in the nearby future, could have many applications, such as lie detection, neuromarketing or brain-computer interfaces. Some of these could, in principle, also be used in forensic psychiatry. The application of these methods in forensic psychiatry could, for instance, be helpful to increase the accuracy of risk assessment and to identify possible interventions. This technique could be referred to as ‘A.I. neuroprediction,’ and involves identifying potential neurocognitive markers for the prediction of recidivism. However, the future implications of this technique and the role of neuroscience and A.I. in violence risk assessment remain to be established. In this paper, we review and analyze the literature concerning the use of brain-reading A.I. for neuroprediction of violence and rearrest to identify possibilities and challenges in the future use of these techniques in the fields of forensic psychiatry and criminal justice, considering legal implications and ethical issues. The analysis suggests that additional research is required on A.I. neuroprediction techniques, and there is still a great need to understand how they can be implemented in risk assessment in the field of forensic psychiatry. Besides the alluring potential of A.I. neuroprediction, we argue that its use in criminal justice and forensic psychiatry should be subjected to thorough harms/benefits analyses not only when these technologies will be fully available, but also while they are being researched and developed.