Welchowski, Thomas: Advances in Machine Learning Approaches for Biostatistical Learning. - Bonn, 2025. - Habilitation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-83173
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-83173
@phdthesis{handle:20.500.11811/13164,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-83173,
doi: https://doi.org/10.48565/bonndoc-586,
author = {{Thomas Welchowski}},
title = {Advances in Machine Learning Approaches for Biostatistical Learning},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2025,
month = jun,
note = {This habilitation thesis summarized current state-of-art advances in machine learning for biomedical applications. The first contribution was the development of a framework for tuning KDSN to increase prediction performance (Welchowski and Schmid, 2016). KDSN are a computational efficient alternative to backpropagation-based artificial neural network techniques with comparable prediction performance on biomedical tabular data that allow layer-wise closed form solutions. The proposed model-based tuning framework is much shorter in terms of computation time than grid-based search strategies. This work was extended to SKDSN that includes variable selection, dropout and regularization to make KDSN more flexible (Welchowski and Schmid, 2019). SKDSN modifications improved upon the performance of KDSN, but could not match the performance of ensemble methods applied to biomedical tabular data sets, especially when the number of covariates was high. IML methods provide tools to gain further insights from those black-box models. A case study in ecology highlighted strength and weaknesses of IML methods that quantify magnitude of effects and their interactions (Welchowski et al., 2022). In particular, graphical tools showed their limits to investigate higher order interaction effects. Previous approaches for inference of model-agnostic interaction effects were limited to few comparisons of covariates sets due to computational runtime intensive resampling and prediction model refitting. The follow-up article Welchowski and Edelmann (2024) then developed a model-agnostic interaction hypothesis test to detect interaction effects to address these shortcomings. Simulations showed control of type I error and reasonable power levels were achieved with approximately few hundred observations. Furthermore due to the derived asymptotic distribution the test is far more computational runtime efficient than previous approaches and can be flexibly specified to covariate sets of interest.},
url = {https://hdl.handle.net/20.500.11811/13164}
}
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-83173,
doi: https://doi.org/10.48565/bonndoc-586,
author = {{Thomas Welchowski}},
title = {Advances in Machine Learning Approaches for Biostatistical Learning},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2025,
month = jun,
note = {This habilitation thesis summarized current state-of-art advances in machine learning for biomedical applications. The first contribution was the development of a framework for tuning KDSN to increase prediction performance (Welchowski and Schmid, 2016). KDSN are a computational efficient alternative to backpropagation-based artificial neural network techniques with comparable prediction performance on biomedical tabular data that allow layer-wise closed form solutions. The proposed model-based tuning framework is much shorter in terms of computation time than grid-based search strategies. This work was extended to SKDSN that includes variable selection, dropout and regularization to make KDSN more flexible (Welchowski and Schmid, 2019). SKDSN modifications improved upon the performance of KDSN, but could not match the performance of ensemble methods applied to biomedical tabular data sets, especially when the number of covariates was high. IML methods provide tools to gain further insights from those black-box models. A case study in ecology highlighted strength and weaknesses of IML methods that quantify magnitude of effects and their interactions (Welchowski et al., 2022). In particular, graphical tools showed their limits to investigate higher order interaction effects. Previous approaches for inference of model-agnostic interaction effects were limited to few comparisons of covariates sets due to computational runtime intensive resampling and prediction model refitting. The follow-up article Welchowski and Edelmann (2024) then developed a model-agnostic interaction hypothesis test to detect interaction effects to address these shortcomings. Simulations showed control of type I error and reasonable power levels were achieved with approximately few hundred observations. Furthermore due to the derived asymptotic distribution the test is far more computational runtime efficient than previous approaches and can be flexibly specified to covariate sets of interest.},
url = {https://hdl.handle.net/20.500.11811/13164}
}