Detecting and forecasting mental health phase transitions is crucial for effective psychotherapeutic
treatment and intervention. Complex dynamics and chaos theory have emerged as useful tools for understanding the mental processes of people suffering from mental problems. They provide insight into the presence of phase transitions and chaotic patterns, whereby the phase transition (PT) is a phenomenon in psychotherapy, which represents a point of change in client’s multivariate time series and can be
psychologically manifested as, for example, a change in the level of depression [1].
The dataset used, consisting of client diary entries, was collected during psychotherapeutic
process in inpatient psychiatric care. The methods explored in the thesis aim
to enhance PT detection and forecasting accuracy by integrating questionnaire-based
features (QF) with text-derived features extracted from unstructured client diary entries.
Leveraging ML and NLP techniques, the study aims to address the limitations
of traditional quantitative-only approaches by incorporating rich textual data.
NLP methods were used to preprocess the textual data and extract features from the
diary texts, such as sentiment analysis (identifying emotions in the text) and different
linguistic patterns (repetitive words etc.). Afterwards, ML methods were used together
with the Pattern Transition Detection Algorithm (PTDA) [2] to identify PTs in the
time series consisting of features extracted by NLP methods. PTDA uses complex
dynamics principles to analyze different dynamic aspects of the data and detect significant
changes or shifts in patterns. By estimating dynamic features such as average
change and periodicity, PTDA provides insight into the occurrence of PTs.
Finally, a forecasting model was developed which was not only able to detect, but also
to forecast with some accuracy when PTs will occur in the future. The performance
of the detection and forecasting model was assessed on the basis of ground truth data
on PT, derived from the expert assessment of the psychotherapeutic time series.
Key findings demonstrated that ML-extended methods outperform traditional PT detection
approaches. A notable 52% improvement in detection accuracy (measured
by a decrease in MSE) was achieved when combining QF and text-derived features.
Forecasting experiments revealed that QF features excel in short-term predictions,
while text-derived features enhance longer-term forecasting. The highest accuracy was
achieved with a balanced combination of QF and text features, highlighting the complementary
nature of these data types.
The aim of this study is to demonstrate the potential of NLP methods in identifying
and forecasting PT in mental health health based on diary entries. By combining
complex dynamics with NLP methods, researchers and health professionals can gain a
more comprehensive understanding of the processes in mental health and develop more
effective treatments and interventions.
|
