dc.description.abstract | Diabetes is one of the most common chronic diseases in Taiwan and is often associated with various complications.
Among them, diabetic nephropathy is one of the most frequent ones.
It is also a disease with high morbidity and mortality.
Because symptoms of kidney-related diseases are usually not readily observable at an early stage,
most patients are unaware of it until the condition has progressed.
By the time the kidney damage has already occurred,
however, it is usually too late, and the patients will need hemodialysis as a treatment method for survival.
If the patients can be informed of the possibility of the disease beforehand,
it may allow them to pay more attention to their health conditions.
In this sense, providing effective temporal information for prediction results is an important influencing factor in the study of longitudinal data.
Therefore, this study will explore the influence of different time series data processing methods on the results based on the existing laboratory data.
In this study, machine learning models with different architectures are trained on biochemical data,
which include the learning model XGBoot that is based on tree structure,
the multilayer perceptron built by tensorflow,
and the Jacobian matrix learning model (JMLM).
In general, JMLM is a more interpretive model compared to other models because it first uses clustering algorithm to group each data point and then uses Taylor series expansion to approximate the data points.
In addition, this study compares multiple feature selection methods and analyzes the impact of features on the results.
Ultimately, with the accuracy and sensitivity reaching 0.857 and 0.854, respectively,
the multi-layer perception and self-selected features have the best effect on cross-validation. | en_US |