As many research results suggest that the psychological aspects of treatment are closely related to the treatment outcomes, investigation of these aspects – along with implementation of appropriate measures – is an essential element in striving towards optimal medical treatment service.
Various approaches, both qualitative and quantitative, were employed for studying the psychological aspects of treatment. The first objective of the present work was to examine these approaches in the specific area of infertility treatment: the thesis presents characteristics, strengths and weaknesses of the approaches for assessment of anxiety, depression, aggressiveness and stress in patients undergoing infertility treatment. The second objective was to establish the position/role of text mining in comparison to other approaches for assessment of the psychological aspects of treatment. For this purpose, the messages posted on the online support group "Infertility" (Neplodnost) of the med.over.net web-forum in the period 2002-2016 were analysed. The analysis consisted of keyword extraction, text clustering, text classification and sentiment analysis.
Keyword extraction was performed using different statistical and linguistic methods as well as network approach based on Rapid Algorithm Keyword Extraction (RAKE). Keywords were extracted manually from a smaller subset of randomly chosen forum messages by two independent annotators (raters). The inter-rater agreement was calculated. The final list of keywords included the keywords selected by at least one annotator. Keywords were then extracted automatically using the abovementioned keyword extraction methods. The analysis was performed using the base forms (lemmas) of adjectives and nouns. Comparison of manually and automatically extracted keywords showed that three methods – term frequency based keyword extraction, normalized term frequency (tf-idf) and RAKE – performed best; their efficacy was similar. However, compared to the first two methods, RAKE showed the advantage of extracting not only single-word but also multiple-word items as keywords. The list of manually determined keywords comprised almost 25% of multiple-word items. Consequently, RAKE was eventually selected for automatic identification of keywords in all other messages. The hundred most frequent identified keywords were depicted as a word cloud.
The word cloud suggested that the online community participants were on one hand looking for informational and emotional support, and on the other hand provided such support. The finding was further supported by retrieval of latent topics that were extracted by the topic modelling. Text pre-processing included removal of stop words, punctuation and numbers, lemmatisation (putting words in their base form) and part-of-speech tagging. Only base forms of nouns, adjectives and verbs were used for the analysis. Topic modelling was performed using the LDA (latent Dirichlet allocation) method. Each individual message was assigned to one of the ten key topics. For each topic words most likely to be associated with it (topic keywords) were identified.
The ten topics were manually merged into two general topics: topic of the emotional support and topic of the informational support. The emotional support topic consisted of three subtopics: encouragement, congratulations, and empathy / sympathy. The informational support topic included nine topics. These comprised discussions about the stages of assisted reproductive treatment, from ovarian puncture to possible pregnancy, and the problems and symptoms associated with each individual stage. The users shared information about official procedures such as referrals, appointment scheduling and waiting times for an appointment. They discussed medicines, their availability, functioning and administration. Participants conversed about the impact of the infertility treatment on their work due to frequent sick leaves. They proved to be reluctant to discuss their condition and treatment with their relatives, friends and co-workers. The participants informed each other about the investigations and surgical procedures that are part of the infertility treatment, about the health system and infertility treatment-related legal regulations, the menstrual cycle, the stages of infertility treatment and possible alternative approaches to treatment
The latent topics that were obtained were content-validated using a random sample of 100 messages from each topic (1000 messages in total) that were human-inspected regarding the content appropriateness. The inspection suggested appropriate content homogeneity of the topics, although some topics seemed somewhat more homogeneous than the others.
Subsequently, the reliability (repeatability) of the findings was verified by three replicates of the LDA method with different random initial assignment of topics to documents (messages) and words to topics. Content measurement was found to be very reliable for eight of the ten topics and less reliable for the other two.
Further analysis included the classification of users in two "pregnancy" groups: (a) pregnancy achieved or reported and (b) pregnancy not achieved or reported. The analysis included the last three messages from the users who contributed at least 100 words to the forum. Different classifiers have been tested on randomly selected and manually annotated messages. The logistic regression classifier performed best and was used to classify all users (their last three messages). The procedure was then repeated with all users who were assigned to the group of non-pregnant women. The text of users classified as pregnant has been manually reviewed to ensure the homogeneity of the group.
Users' sentiment analysis was performed using the Kadunc and Robnik Šikonja sentiment dictionary (based on the Hu and Liu sentiment dictionary). The dictionary contains a list of words with positive and negative sentiment. The positive to negative words ratio per user was calculated. The analysis included description of the overall users’ sentiment and the sentiment of the two pregnancy groups. Negative sentiment words appearing in the forum text were classified into four categories – anger / aggression, fear / anxiety, sadness / depression, undefined – by three independent raters, following the example of the LIWC (Linguistic inquiry and word count) dictionary. Inter-rater agreement was calculated. The final list consisted of words that achieved the majority agreement between raters.
For each topic, sentiment was calculated as the positive to negative words ratio to find out which treatment aspect(s) users perceived as most negative. The proportion of each emotional category words among all negative words was calculated for each topic, providing identification of the dominant negative emotions for each particular topic. This proportion was also calculated for all users and for each pregnancy group.
Findings of the analysis were compared to other related research, both in Slovenia and abroad as the third objective of the thesis was validation of the results obtained using the text mining technique. Text-mining analysis of online support communities’ has already been used in other health-related research areas, but hardly ever in the present study field and not at all in Slovenia. Therefore, the present results were compared to the results of other research approaches in the studied field. Comparison with regard to contents showed similarities to the results of other studies in the research field. For validation purposes, the present results were also compared with the results of an online survey analysis, with administrator's permission published on the online Infertility support group discussion board in 2018. The results of the text-mining approach were comparable to those of this survey.
The advantages of the approach presented in this work are its non-intrusiveness, time- and cost-effectiveness and the benefit of "disinhibition effect" (the sharing of sensitive and personal data as a result of the anonymity offered by the Internet), while its disadvantages are the non-representative sample and the associated conclusion bias, lack of basic demographic and anamnestic data on the forum participants, and the passive researcher’s role in obtaining information. As with many research approaches, the credibility of the conclusions based on the findings and confidence in them are greater if they are similar, regardless of the research approach used. Combination of research approaches is thus a key feature of the endeavour to acquire an extensive knowledge about patient's psychological experience during their medical treatment process.