Psychotherapy

Under the leadership of Prof. Dr. Bernhard Humm (FB I) from the Research Center for Applied Computer Science (FZAI) a machine learning algorithm is being developed in cooperation with Vacay GmbH to predict possible psychotherapy discontinuations of suicidal patients. The database consists of a large number of questionnaires that patients fill out several times a day. The algorithm is based on a neural network whose quality depends heavily on this training data. The quality of the prognosis is determined with the help of the F1-measure, related to the feature "therapy discontinuation". A value of about 0.7 is aimed for, as it is known from previous research that this is a realistic and achievable value. For the study planning, a statistically sound case number planning is to be carried out in order to find out how many patients will probably at least be needed in order to achieve the targeted F1 value. However, economic and ethical aspects also play a role here. A well-founded case number is therefore mandatory, since the approval of an ethics committee is also required.

A very large number of networks were trained by resampling the existing patient data in different sample sizes and their F1 values were calculated. These were used to investigate different types of functions for learning curves by looking at the 95% confidence intervals of their extrapolations to use them to determine a minimum number of patients for the target F1 value of 0.7. Different sampling strategies were also used.

Projekt manager

Prof. Dr. Bernhard Humm (h_da, FB I und FZAI)

Contact persons at DISO

• Prof. Dr. Antje Jahn
• Florian Junge

Cooperation partner

Martin Schüller (Vacay GmbH)

Time period

June 2018 to July 2020

• June 2018: First inquiry and initial discussion with Tilman Deuschel and Stephan Gimbel
• November 2018: Presentation of the concept in a large round at DISO (at that time: CCSOR) with hints for improvement and revision
• February 2020: New contact by Prof. Dr. Humm for methodological advice on extrapolation of learning curves
• March 2020: New meeting and workshop with Prof. Dr. Humm and his project partner Martin Schüller; further work by Florian Junge and Martin Schüller
• June 2020: Discussion of the results in a workshop; discussion of the final steps
• July 2020: Final document for inclusion in corresponding DFG application; inclusion of DISO as cooperation partner for further steps of the project