JMIR Publications

ABSTRACT

Background:

Recently, Artificial Intelligence technologies and learning machine methods offer 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 diagnosis process and to suggest treatment rational options in medicine. But data in psychiatry are related to behaviour and clinical evaluation, they are more heterogenous, 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 specific case, Bayesian Network (BN) might be helpful and accurate, when constructed from expert knowledge. We present MEDICAL COMPANION, a government funded smartphone application based on repeated questions to the subject and algorithm-matched advices to prevent relapse of suicide attempts within several months.

Objective:

(1) to present our methodology, based on American Psychiatric Association (APA) digital healthcare guideline, designed to develop a BN algorithm as a medical device and, (2) to provide results from a pre-clinical phase.

Methods:

Experts were psychiatrists working in university hospitals, experienced and trained in the management of suicidal crisis. As recommended to build a BN, we divided the process into two tasks, (a) structure determination, representing the qualitative part of the BN, and (b) parameter elicitation, with the conditional probabilities, corresponding to the quantitative part. In task one, we determine contextual variables which could be mandatory data in any clinical study (age, gender), then we defined clinical question related to the mental state of the patients and finally we propose specific factors related to the questions .The factors were chosen for their known and demonstrated link with the suicidal risk in the literature (clinical, behavioral and psychometrics). Finally, we propose specific advices related to the patient’s state. We obtain a structure for the BN with graphical representation of causal relations between variables. Task 2 allows to ensure that the advices were adapted to the clinical states of patients and to the context. We run simulation (usecase) in a four-step process: at the first step, simulations of the first set of data for mental state provided from expert. At the second step, back to expert with simulation results and validation of accuracy. At the third step, simulation of the first set of data for advices provided from expert with incremental updates with usecases. At the fourth step, back to expert with simulation results and validation of accuracy of advices.

Results:

As we expect, a such a priori expert model is valuable to test preclinical situation and result’s simulation with various patient profiles (usecase) allows us to build a valid BN.

Conclusions:

BN is an adapted methodology to build an algorithm for a digital assistant dedicated to suicidal crisis management. Digital psychiatry is an emerging field, but it needs valid evaluation and testing before using with patients. As any medical device or by analogy to psychotropics, a phase II (preclinical) trial is necessary. With this method, we propose another step to respond to the APA guidelines.