Artikel
Prediction of chronic stress in practice assistants of general practices (small-business): comparison of four machine learning approaches with a classical statistical procedure
Suche in Medline nach
Autoren
-
Arezoo Bozorgmehr
- Institute of General Practice and Family Medicine, University Hospital Bonn, University Bonn, Bonn, Deutschland
-
Anika Thielmann
- Institute of General Practice and Family Medicine, University Hospital Bonn, University Bonn, Bonn, Deutschland
-
Birgitta Weltermann
- Institute of General Practice and Family Medicine, University Hospital Bonn, University Bonn, Bonn, Deutschland
| Veröffentlicht: | 25. September 2020 |
|---|
Gliederung
Text
Background and current state of (inter)national research: Occupational stress is associated with adverse outcomes. Previous studies documented high chronic stress of personnel in primary care. We aim to derive the best prediction model for chronic stress by comparing four common machine learning (ML) approaches to a classical statistical procedure.
Research questions and objectives: Which of four ML approaches and one classical statistical method shows the best prediction accuracy when exploiting complex interactions between practice characteristics and chronic stress?
Methods or hypothesis: We compare four ML algorithms (support vector machines, K-nearest neighbors’, random forest, and artificial neural network) and one of the most frequently used classical statistical procedure (logistic regression) to develop prediction models for chronic stress in practice assistants. The operating curve’ (AUC) is used to assess the predictive accuracy of the models. For the best model, the five key factors associated with chronic stress are determined. The data set comprises personnel from 185 teaching practices which used the TICS (Trierer Inventar for chronic stress) as outcome (Viehmann, Plos One 2017).
Results: The data set comprised 550 practice assistants (98.5% females). The TICS was analysed as binary outcome. The accuracy of the classical logistic regression and four machine learning algorithms to predict chronic stress was as follows: logistic regression (AUC 0.598, 95% CI), random forest (AUC 0.770, 95% CI), neural networks (AUC 0.696, 95% CI), support vector machines (AUC 0.693, 95% CI), and K-nearest Neighbours (AUC 0.616, 95% CI). Using the variable frequencies at the decision nodes of the random forest model, these five factors were most important: High work pressure, high demand to concentrate, time pressure, complicated things at work, age.
Discussion: Machine learning approaches showed significantly better prediction accuracy compared to a classical logistic regression, In particular, random forest performed the best, with a predictive accuracy of 77% when compared to three other ML methods.
Practical implications: ML can help to understand complex data from primary care.
