Seminários em Neurociências 2011.2
Quarta-feira 30/11/2011 10:50h Sala 3 – Escola de Ciência e Tecnologia
Graph analysis of psychotic speech: differential diagnosis between schizophrenia and mania
Natália Mota (Instituto do Cérebro & Hospital Universitário Onofre Lopes, UFRN)
The differential diagnosis between manic and schizophrenic patients presents a substantial challenge during acute psychotic crises. Successful diagnosis requires long-term training in the identification of symptoms assessed by a qualitative analysis of speech. Here we sought to quantify differences in the speech graph structure of schizophrenic and manic psychotic subjects.
We recorded oral interviews with 24 subjects (8 schizophrenics, 8 maniacs and 8 controls) and applied the SCID DSM IV, PANSS and BPRS scales to identify psychotic symptoms. Patients were asked to report on recent dreams. The reports were transcribed, parsed into semantics units (SU) and represented by a directed graph in which each node corresponded to a SU and each edge represented the link between consecutive SU. Twelve graph attributes were calculated (nodes, edges, self-loops, parallels edges, largest connected component, largest strongly connected component, average total degree, wake nodes, wake edges, loops with one, two and three nodes), and non-parametric statistical tests were used to assess significant differences. A naive Bayes classifier was trained with different combinations of the graph attributes as inputs. The output was a binary decision in the form “is this graph from a given group or not”. To quantitatively compare the schizophrenic, manic and control groups, we calculated the sensitivity and specificity of classification and a receiver operating characteristic (ROC) curves were built based on the output of the classifier, using the area under the ROC curve (AUC) as a metric of classification quality (AUC = 0.5 means classification at chance level and AUC = 1 means maximum classification quality). We also calculated the kappa statistic to assess the agreement between psychiatric diagnosis and group classification based on speech graphs. (Values > 0.6 mean good agreement and > 0.8 mean great agreement).
Manic reports contained more words than schizophrenic group (p = 0.0067) and almost all attributes were significantly higher. When the data were normalized by the total number of words in each report, graphs from the manic group still displayed more parallel edges (p = 0.0050) than graphs from the schizophrenic group, with more loops than control group (p = 0.0019), reflecting the “logorrhea” symptom typical of maniacs. Conversely, schizophrenic reports presented more nodes (p = 0.0114) and a higher average degree (p = 0.0074) than graphs from maniacs, reflecting “poor speech”. Manic patients had a significantly higher rate of interruptions of the dream report to comment on unrelated waking events. This effect persisted when the data were normalized by the total number of words (p = 0.0196), and seems to reflect the symptom of “flight of thoughts”. The classifier based on graph attributes sorted schizophrenic from manic group with 93% of sensitivity and specificity (kappa: 0.88, AUC: 0.88). Schizophrenics were sorted from control group with 93% of sensitivity and specificity (kappa: 0.875, AUC: 0.97) and maniacs from control group with 81% of sensitivity and specificity (kappa: 0.625, AUC: 0.94). None of these graph attributes were correlated with BPRS and PANSS total score, which indicates that our approach is not redundant with psychiatric scales, but rather measures complementary features such as structural speech symptoms.
Altogether, the analyses reveal quantitative differences between speech graphs from schizophrenic and manic psychotic patients, which may reflect classical symptoms not well grasped by standard psychiatric scales. Quantitative speech analysis can therefore help the differential diagnosis of psychosis.
Local: Sala 3 – 1º. andar - Escola de Ciências e Tecnologia / UFRN
Campus Universitário Lagoa Nova, 59078-970 - Natal - RN
Tel: +55 (84) 9193 – 6311 (Instituto do Cérebro)