Name of participant: Amelie Wührl
Project’s name: Patienten-Sprachanalyse mithilfe von NLP zur Erfassung erwartungsbasierter Behandlungsergebnisse (PLACEBO)
Project description:
The treatments’ success is very closely connected to a patient’s attitude and their expectation towards the treatment (Bingel et al., 2011; Aulenkamp et al., 2023; Rief et al., 2016, i.a.). This is true for both positive and negative treatment outcomes, commonly referred to as the Placebo and Nocebo effect. Understanding treatment expectations and their psychological and neurobiological mechanisms require carefully designed medical trials as well as surveys to understand individual patients’ mindsets. While of high quality, these studies are time-consuming and due to their costly setup often limited to a small number of participants.
Both limitations can be mitigated by accessing social media data. Within the area of Natural Language Processing (NLP), social media health mining specializes on automatically extracting, processing and modeling information from large social media data. So far however, there is no research on detecting descriptions and assessments of treatment expectations. As patients frequently turn to online platforms to discuss their medical journey, this data has the potential to understand factors that impact treatment success on a large scale. Moreover, we may identify novel factors not captured by traditional survey-based studies from this type of content.
Therefore, the goal of the project is to build data and AI models that enable us to automatically detect descriptions of treatment expectations and to analyze first-hand accounts of a large range of patients, patient groups and demo- graphics.
To this end, the first objective is to understand the characteristics of treatment expectations on social media. Based on this we design and conduct an annotation study with which we create a training and evaluation dataset for our task. Then, we investigate which modeling approach best facilitates classifying documents and text spans with respect to the task. We will consider state-of-the-art large language models (LLMs), as well as more light-weight custom models. Based on our models, we investigate the extent to which we can leverage such models to discover novel information and analyze treatment expectations in large scale social media data. Together with the industry partner, we start to develop effective ways to address patient concerns based on this knowledge, either through communication or treatment development efforts.
Upon completion of the project, we will have a dataset annotated for treatment expectation detection. Further, we will have computational models for treatment expectation detection and detailed understanding of the capabilities of state-of-the-art language models for the task. Moreover, we will evaluate a prototype application. Such a system could for example provide industry stakeholders with prevalent expectations and patients with targeted pre-treatment information to mitigate any negative expectations.
Software Campus Partner: Universität Stuttgart and Merck KGaA
Implementation period: 01.05.2025 – 30.04.2027
