Prevention of Suicidal Relapses in Adolescents With a Smartphone Application: Bayesian Network Analysis of a Preclinical Trial Using In Silico Patient Simulations.

BACKGROUND: Recently, artificial intelligence technologies and machine learning methods have offered attractive prospects to design and manage crisis response processes, especially in suicide crisis management. In other domains, most algorithms are based on big data to help diagnose and suggest rational treatment options in medicine. But data in psychiatry are related to behavior and clinical evaluation. They are more heterogeneous, less objective, and incomplete compared to other fields of medicine. Consequently, the use of psychiatric clinical data may lead to less accurate and sometimes impossible-to-build algorithms and provide inefficient digital tools. In this case, the Bayesian network (BN) might be helpful and accurate when constructed from expert knowledge. Medical Companion is a government-funded smartphone application based on repeated questions posed to the subject and algorithm-matched advice to prevent relapse of suicide attempts within several months. OBJECTIVE: Our paper aims to present our development of a BN algorithm as a medical device in accordance with the American Psychiatric Association digital healthcare guidelines and to provide results from a preclinical phase. METHODS: The experts are psychiatrists working in university hospitals who are experienced and trained in managing suicidal crises. As recommended when building a BN, we divided the process into 2 tasks. Task 1 is structure determination, representing the qualitative part of the BN. The factors were chosen for their known and demonstrated link with suicidal risk in the literature (clinical, behavioral, and psychometrics) and therapeutic accuracy (advice). Task 2 is parameter elicitation, with the conditional probabilities corresponding to the quantitative part. The 4-step simulation (use case) process allowed us to ensure that the advice was adapted to the clinical states of patients and the context. RESULTS: For task 1, in this formative part, we defined clinical questions related to the mental state of the patients, and we proposed specific factors related to the questions. Subsequently, we suggested specific advice related to the patient's state. We obtained a structure for the BN with a graphical representation of causal relations between variables. For task 2, several runs of simulations confirmed the a priori model of experts regarding mental state, refining the precision of our model. Moreover, we noticed that the advice had the same distribution as the previous state and was clinically relevant. After 2 rounds of simulation, the experts found the exact match. CONCLUSIONS: BN is an efficient methodology to build an algorithm for a digital assistant dedicated to suicidal crisis management. Digital psychiatry is an emerging field, but it needs validation and testing before being used with patients. Similar to psychotropics, any medical device requires a phase II (preclinical) trial. With this method, we propose another step to respond to the American Psychiatric Association guidelines. TRIAL REGISTRATION: ClinicalTrials.gov NCT03975881; https://clinicaltrials.gov/ct2/show/NCT03975881.

Integrating expert's knowledge constraint of time dependent exposures in structure learning for Bayesian networks.

Learning a Bayesian network is a difficult and well known task that has been largely investigated. To reduce the number of candidate graphs to test, some authors proposed to incorporate a priori expert knowledge. Most of the time, this a priori information between variables influences the learning but never contradicts the data. In addition, the development of Bayesian networks integrating time such as dynamic Bayesian networks allows identifying causal graphs in the context of longitudinal data. Moreover, in the context where the number of strongly correlated variables is large (i.e. oncology) and the number of patients low; if a biomarker has a mediated effect on another, the learning algorithm would associate them wrongly and vice versa. In this article we propose a method to use the a priori expert knowledge as hard constraints in a structure learning method for Bayesian networks with a time dependant exposure. Based on a simulation study and an application, where we compared our method to the state of the art PC-algorithm, the results showed a better recovery of the true graphs when integrating hard constraints a priori expert knowledge even for small level of information.

Probabilistic graphical models for genetic association studies.

Probabilistic graphical models have been widely recognized as a powerful formalism in the bioinformatics field, especially in gene expression studies and linkage analysis. Although less well known in association genetics, many successful methods have recently emerged to dissect the genetic architecture of complex diseases. In this review article, we cover the applications of these models to the population association studies' context, such as linkage disequilibrium modeling, fine mapping and candidate gene studies, and genome-scale association studies. Significant breakthroughs of the corresponding methods are highlighted, but emphasis is also given to their current limitations, in particular, to the issue of scalability. Finally, we give promising directions for future research in this field.

Visualization of pairwise and multilocus linkage disequilibrium structure using latent forests.

Linkage disequilibrium study represents a major issue in statistical genetics as it plays a fundamental role in gene mapping and helps us to learn more about human history. The linkage disequilibrium complex structure makes its exploratory data analysis essential yet challenging. Visualization methods, such as the triangular heat map implemented in Haploview, provide simple and useful tools to help understand complex genetic patterns, but remain insufficient to fully describe them. Probabilistic graphical models have been widely recognized as a powerful formalism allowing a concise and accurate modeling of dependences between variables. In this paper, we propose a method for short-range, long-range and chromosome-wide linkage disequilibrium visualization using forests of hierarchical latent class models. Thanks to its hierarchical nature, our method is shown to provide a compact view of both pairwise and multilocus linkage disequilibrium spatial structures for the geneticist. Besides, a multilocus linkage disequilibrium measure has been designed to evaluate linkage disequilibrium in hierarchy clusters. To learn the proposed model, a new scalable algorithm is presented. It constrains the dependence scope, relying on physical positions, and is able to deal with more than one hundred thousand single nucleotide polymorphisms. The proposed algorithm is fast and does not require phase genotypic data.

A hierarchical Bayesian network approach for linkage disequilibrium modeling and data-dimensionality reduction prior to genome-wide association studies.

BACKGROUND: Discovering the genetic basis of common genetic diseases in the human genome represents a public health issue. However, the dimensionality of the genetic data (up to 1 million genetic markers) and its complexity make the statistical analysis a challenging task. RESULTS: We present an accurate modeling of dependences between genetic markers, based on a forest of hierarchical latent class models which is a particular class of probabilistic graphical models. This model offers an adapted framework to deal with the fuzzy nature of linkage disequilibrium blocks. In addition, the data dimensionality can be reduced through the latent variables of the model which synthesize the information borne by genetic markers. In order to tackle the learning of both forest structure and probability distributions, a generic algorithm has been proposed. A first implementation of our algorithm has been shown to be tractable on benchmarks describing 105 variables for 2000 individuals. CONCLUSIONS: The forest of hierarchical latent class models offers several advantages for genome-wide association studies: accurate modeling of linkage disequilibrium, flexible data dimensionality reduction and biological meaning borne by latent variables.