AI-Based Language Models for Improving Diagnosis, Monitoring, and Outcomes of Depression and Anxiety
Researchers in the project want to develop AI-based language models for use in clinical contexts to support diagnosis and monitoring of depression and anxiety.
Depression and anxiety are among the most widespread diseases of our time. They create enormous suffering for many people, along with high costs. For instance, mental illness now accounts for nearly half of all sick leave in Sweden. Correct and early diagnosis, accompanied by monitoring of anxiety and depression over time, are vital in order to begin, change or conclude the right treatment and the right time. This will in turn lead to speedier recovery and reduce suffering.
At present, anxiety and depression are largely diagnosed by means of subjective assessments made by physicians, based on interviews between care provider and patient, sometimes with the help of rating scales. Validated decision support is often absent. Language is the natural way for people to communicate about their state of mind. But there is a lack of reliable and scientifically standardized tools to measure how patients describe their mental ill-health in clinical contexts.
The research team has previously conducted basic studies into how patients can describe mental ill-health in words, which are analyzed using semantic language models, and how these have high validity for rating scales for depression and anxiety. The project now aims to make a substantial innovative contribution by showing how the accuracy of these measuring instruments can be markedly improved by:
- Combining different response formats such as free text and rating scales, as well as different types of question in an optimal manner
- Using “intelligent questionnaires” to select the questions that yield the most information about a prospective diagnosis
- Using AI-based deep language models that understand grammar and semantics in different languages, along with machine learning algorithms that weigh up different factors in an optimal way
- Studying the attitudes of patients and clinics, and their perceptions of the benefits, accuracy and reliability of using these measuring instruments in clinical contexts
- Using a randomized clinical trial (RCT) to evaluate how the decision support the team has developed impacts the length of time that patients are sick and how long they are on sick leave as compared with current practice
The proposed methods radically improve the precision in diagnosis of depression and anxiety diagnosis. Preliminary data show that:
- The proportion of misdiagnosis can be halved as compared with use of rating scales by combining multiple texts and rating scales with AI
- The method relying solely on free text and word responses predicts rating scales with greater accuracy than earlier literature suggests (r = .85)
- Word responses are more accurate than rating scales at classifying self-perceived emotional states or emotional states reflected by facial expressions
In summary, the researchers intend to create new, practicable methods that give patients a natural way of using word responses to describe their mental health. The researchers believe this will improve the reliability of diagnosis and monitoring of anxiety and depression. They will be using RCT to measure how this leads to speedier recovery and shortens sick leave.
Affiliated with WASP-HS
This research project is affiliated with WASP-HS and genereously funded by Marianne and Marcus Wallenberg Foundation.
Professor, Lund University
PostDoc, Lund University