logo SBA


Digital archive of theses discussed at the University of Pisa


Thesis etd-05082014-123621

Thesis type
Tesi di dottorato di ricerca
Thesis title
Data Collection Platforms for Mental Health Research
Academic discipline
Course of study
tutor Pioggia, Giovanni
  • biomedical signal processing
  • mobile application
  • stress classification
Graduation session start date
The modern society exposes the individual to a lifestyle often frantic. This turn in a persistent harmful state of stress for the body. It hasn’t evolved to be subjected at persistent stressful situations. For this reason, in recent years there has been an increase in diseases due to stress. An example are cadiovascular pathologies.
The aim of my PhD was to validate and provide tools to assess (coming from physiologic signals) the person’s Level of Stress during everyday life. This information if acquired noninvasively (without annoy users) is fundamental in treatment, thanks to Cognitive Behavioural Therapy (CBT) of patients with pathological stress situations. It’s also very important in prevention, giving to users real-time feedback on their own situation, and through special techniques of concentration and relaxation. It’s demostrated how this treatments help to normalize parameters that indicate a high level of mental stress.
To accomplish this result various steps needed. The first was to find out together with experts, State-of-the-art method that could be used to collect preliminary data. The literature suggested to look at Experience Sampling Method, a method which gave a better ability to contextualize Psychological data during the ordinary life. I chose to go beyond the State of the art, by implementing the first application for Smartphone capable to gathering Psychological and releted Physiological informations (ECG-Accelerometer). Developed application called PsychLog.
Meanwhile PsychLog was realized, it was necessary to choose most suitable sensor for data acquisition. Moreover it could be able connect Smartphone. The chosen sensor was Shimmer2 ®. The reasons were, its dimensions, reduced weight and quality of the signal.
Next step was the validation of whole Smartphone + Wearable Sensor System. ECG data were collected at 250 Hz, a lower sample-rate than 1000 Hz (frequency of a clinical Holter). It was therefore necessary to find out important differences between tachograms obtained by both devices. It’s well known from literature that fatures that better describe autonomic system (and therefore more stress-related) are from tachogram signal (differences between ECG RR spikes). First of all it was necessary to evaluate the QRS detection algorithm performance, implemented inside the Smartphone. Two subjects dressed both Holter and Sensor simultaneously. Sensitivity and Specificity were calculated (of detected beats), achieving very high values for both (Sens = 99.97% and Spec = 99.94%). Furthermore Relative Error Percentage was evaluated according to CEI 2-47 ISO60601-2-47. Error has been always lower than the 10% (maximum admitted). It was also evaluated the absence of any Bias between Holter and sensor data. This first described stage has been useful to evaluate the QRS detection algorithm (fundamental to calculate the most important signal where to extract stress-related features). I did also an evaluation of power Spectral Density (PSD) from both Holter and Sensor Tachogram. The measured error presents the following statistic: Subj1 (m = 0.0839% σ = 0.3143%) and Subj2 (m = 0.3457% σ = 0.1435%). The PSD from sensor has to be as close as possible to Holter’s PSD, because tachogram is used as source of features both in the time and frequency domain. These analyses demostrated that the Wearable + Smartphone System was suitable to address a bigger measure campaing.
A bigger data collection campaign was realized in order to evaluate physiological and psychological data correlation. In particular if the Psychological data could be used as Label for Physiological. Considered people belonged at two professional categories (Nurses and Teachers). Emotional state information is gained through questionnaires presented by the Smartphone. It was important to look at the possibility to reduce number of questions, and use new scores for stress level’s assessment, without losing information of gold standard scores (all proposed questionnaires was deliveded by expert clinicians). Six subjects’ data was collected for a week each. Auto-Perceived Stress and Gold Standard Indexes showed a positive correlation (R = 0.348) that accoding to clinitians opinion is a good result. Moreover I did a deeper analysis for a particular subject, This showed as the ratio (LF/HF) had an inversely proportional trend respect level of stress, as reported in literature.
The next step was to create a large measure campaign. The sessions considered for subsequent evaluations were 726. With collected data I wanted to achieve an automatic decision model for Stress Level’s evaluation from physiological data (avoiding wasted time to fill questionnaires). The literature proves that most important features were 15 (9 in the time domain and 6 in frequency domain). Not all that features could be used as inputs for automatic decision model. It was therefore necessary to select the most important ones. In First step I chose to apply the ReliefF method (choosen for its higher robustness against noise in data). In this way I got a ranking of features’ importance about classification. The next step was to choose the correct number of features. The Davies-Bauldin Index was calculated iterating number of features (each step added a new one), having trend of this parameter. Davies-Bauldin Index has not been used as an index of classification’s quality, but as a discriminator of most important features within the data set. Lower value was obtained with number of features amounting 6.
The features are selected and the final step was to define the model that best categorize collected data (DSS). The features were used as input to different classifiers. The number of defined classes was 4 (No Stress, Low, Medium, High Stress). These features have been normalized in linear way. By Leave One Subject Out Cross Validation the K-nearest got a misclassification of 17.31%. To improve the classification, I decided to use a different approach at the standardization of the incoming data. Variables were fuzzyfied. For each feature, I defined 3 Membership Functions, under guidance of medical staff and looking at literature, obtaining, de facto 3 new features for each parameter. I again tested each classifier. The only one that showed an improvement was the SOM (Kohonen Self Organizing Map), obtaining an average classification error equal to 11.83%, with capability to discriminate between the 4 defined classes, close to 90%.
The last step was to implement a Server Side application able to connect to Database through WSDL functions. At this point I created a prediction model for Stress Level within a hardware/software architecture, and customized for each user. The functioning is divided into 2 parts. The Training stage is one week long, and mobile device acquires both physiological and psychological data. These data are saved in the DB through WSDL functions. Meantime DSS system reads the data and trains the SOM. Then there is the Test phase. For each session measurement,the system automatically calculates Level of Stress from physiological data, and notifies the user his Level of Stress
The system, validated and developed, can be used for many purposes. One example is aggregated data collector. This work for the first time combines behavioral and autonomic system information. If used as a Black Box, the system can convey a therapy aimed to lowering stress. Moreover this work represents the starting point for a research thread in the field of Correlations between Mind and Autonomic System, thanks to the large amount of data that has been recorded and analyzed. The architecture is easily extensible and a challenge for future is the introduction of new features gained with less invasive systems in order to realize more precise knowledge based model .