Machine Learning methods to build prediction models for mental well-being indicators in healthcare workers
Year: 2024
Type: dissertation
Abstract: The main objectives of the work are to build prediction models with machine learning methods to predict dropouts from online intervention and to cluster the dropouts to provide more information about them. The dropouts are predicted from a web-based Personal COVID intervention. The intervention consists of nine core modules and three complementary modules. Before starting the intervention, the participants answered different questionnaires. The participants were selected to the intervention based on their answers to specific questionnaires. The data used in this work consists of preintervention data and module data. The preintervention data refers to the data collected before starting the intervention (demographic data and answers to the questionnaires related to mental health, sleep and professional life quality), and the module data refers to the actions made by the participant during the intervention. Exploratory data analysis was performed to get a good overview of the data, but also to find meaningful information about the dropouts. The data set is small, it contains in total 64 participants, and there is a high imbalance between dropouts and non-dropouts. Exploratory data analysis also contained statistical analysis of the preintervention variables and principal component analysis. The prediction models were built for modules 1-4 because they had most dropouts. Models were built by trying out different data set combinations and different machine learning algorithms (logistic regression, decision tree, gradient boosting, support vector machine and random forest). Due to a high number of different combinations, 180 models were trained with moderate results considering the small data size. Overfitting occured especially in the models of modules 2 and 4 where there were fewer dropouts, so threshold adjustment was applied to these models. Two of the best-performing models of each module were selected for hyperparameter optimization, but in general it didn't significantly affect their performance. Only one model's performance improved noticeably. The average balanced accuracy of the best-performing model in each module is 0.735 and the average sensitivity is 0.743. These results are reasonable considering the characteristics of the data. All four models use the preintervention data for prediction, indicating that it is enough for predicting dropouts. Clustering was performed on modules 1 and 3 due to their higher dropout rates. The data were classified into dropouts and non-dropouts using the best-performing model of the module, and then both groups were clustered. Two different methods were tried, k-means and hierarchical clustering. Based on the statistical analysis, the clusters are different enough but k-means resulted in clusters with more internal statistical differences. For these clusters, profiling information was provided by calculating average and standard deviations for each variable.
Author: Mähönen, Helka
Language: English
Subfield: Health Information Management
Field: Health Professions
Domain: Health Sciences
Keywords: psychological well-being, healthcare workers, COVID-19 pandemic, online intervention, class imbalance, hyperparameter optimization, threshold adjustment, K-means clustering, hierarchical clustering, mental health screening, sleep quality, user profiling, small datasets, overfitting, random forest, gradient boosting, support vector machine
Open Access status: green
OpenAlex ID: W7133000526
