Filtern
Dokumenttyp
- Masterarbeit (2) (entfernen)
Sprache
- Englisch (2) (entfernen)
Schlagworte
Institut
- Fakultät Management und Technologie (2) (entfernen)
In the study, predictive models for predicting therapy outcome are created using the dataset from E-COMPARED project, which belongs to the so-called type 3 models that use data from the intervention and preintervention phases to predict treatment outcomes, which can help to adapt intervention to maximize treatment. The predictive models aim to classify patients into two groups, improved and nonimproved. Since it is important to determine whether the models contribute to improvement of treatment, research questions that can contribute to the usage of type 3 models are established. The study focuses on the following three questions: (1) How accurately can the therapy outcome be predicted by various machine learning algorithms? Answering this question can let the people concerned obtain information about the reliability of contemporary predictive models. In addition, if the predictive power of the models is good, it is more likely to be used to assist therapists’ decisions. (2) Which kind of data is more important in predicting the therapy outcome? The answer to this question can show which dataset should be considered first to make better predictive models. Therefore, it can be helpful for researchers who want to make predictive models in the future and eventually help to facilitate personalized therapy. (3) What are the features with strong predictive power? The answer to this question can affect the people concerned, especially therapists. Therapists can use the most influential features revealed to adjust and improve future treatments.
In past digital health interventions, an issue has been that participants drop out over time which is referred to as the "law of attrition" (Eysenbach, 2005). Based on this, the study proposes that though initially, participants respond to the intervention, there is a hypothesized second diminishing effect of an intervention. However, the study suggests that on top, there is a third effect. Independent of the individual notification or nudge, people could build the knowledge, skills and practice needed to independently engage in the behavior themselves (schraefel and Hekler, 2020). Using behavioral theory and inspired by prior animal computational models of behavior, the thesis proposes a dynamical computational model to allow for a separation of intervention and internalization. It is targeted towards the specific case of the HeartSteps intervention that could not explain a diminishing immediate effect of the intervention, second hypothesized effect, while a person’s overall steps remained constant, third effect (Klasnja et al., 2019). The study incorporates a habituation mechanism from learning theory that can account for the immediate diminishing effect. At the same time, a reinforcement mechanism allows participants to internalize the message and engage in behavior independently. The simulation shows the importance of a participant’s responsiveness to the intervention and a sufficient recovery period after each notification. To estimate the model, the study uses data from the HeartSteps intervention (Klasnja et al., 2019; Liao et al., 2020), a just-in-time adaptive intervention that sent two to five walking suggestions per day. The study runs a Bayesian estimation with Stan in R. Additional validation tests are needed to estimate the accuracy of the model for different individuals. It could however serve as a template for future just-intime adaptive interventions due to its generic structure. In addition, this model is of high practical relevance as its derived dynamics can be used to improve future walking suggestions and ultimately optimize notification-based digital health interventions.