Diagnóstico diferencial baseado em variáveis autonômicas e estrutura do discurso entre transtornos neuropsiquiátricos utilizando aprendizagem de máquina

Abstract

The study of the adaptive nature stress response shows the major participation of physiological mechanisms associated with the hypothalamic-pituitary-adrenal endocrine axis and the Autonomic Nervous System (ANS), in its sympathetic and parasympathetic divisions, as well as the immune system. Individuals with different neuropsychiatric disorders present signs and symptoms that suggest dysregulation of the ANS (dysautonomia) and the expression or form of oral speech. These occurrences are present in other disorders that are part of the psychophysiological spectrum that includes post-traumatic stress disorder (PTSD), anxiety disorder (AD) and obsessive-compulsive disorder (OCD). In this context, the objective of this study was to develop a diagnostic classification mathematical model for PTSD, based on the analysis of autonomic variables and discourse structure. We investigated 298 males aged 22 to 48 years allocated into four groups: PTSD (n = 76), AD (n = 77), OCD (n = 73) and Control (n = 72). The PCL-5, BAI and YBOCS questionnaires were used to obtain the psychometric data related respectively to PTSD, AD and OCD. The SpeechGraphs® software was used to analyze the representation of the word trajectory and to quantitatively characterize the speech complexification. An ECG signal (ADInstruments model PowerLab®) was used to analyze heart rate variability (HRV) and skin conductance (RGP). Machine learning techniques (Decision Tree and Naive Bayes) were employed to obtain the mathematical model. The model generated for PTSD classification based on HRV measurements presented accuracy of 92.3% (p <0.0001) and Kappa index of 89.7% with the generation of a decision algorithm using parasympathetic axis measurements (SDNN and RMSSD) and sympathetic (LF). The model generated for PTSD classification based on autonomic skin conductance (μS) measurements presented 96.6% accuracy (p <0.0001) and 95.4% Kappa index with the generation of a decision algorithm using measurements checked on the second one and 180 of the five-minute. The model generated for PTSD classification based on speech trajectory measurements presented accuracy of 80.9% (p <0.0001) and Kappa index of 71.4% with the generation of a decision algorithm using lexical diversity measures (Nodes), recurrence (RE, PE) and connectivity (LSC, LCC and L3). The classification in levels of severity of the disorder allowed the identification, by the k-means method, of 3 classes (degrees) of elevation for each variable. Models generated with autonomic measurements have better accuracy for PTSD classification and present the potential to be used as a more efficient method for diagnosis, future investigations into risk stratification, severity categorization and follow-up of the clinical evolution of this disorder.