80 lines
5.0 KiB
TeX
80 lines
5.0 KiB
TeX
%! Author = alex
|
|
%! Date = 3/6/25
|
|
|
|
\section{Methodology}\label{sec:methodology}
|
|
|
|
% why did I select tft over other methods -> include examples of time series and why I belief a complex model could help
|
|
% you did I apply it
|
|
% implementation details
|
|
|
|
|
|
\subsection{Data Collection \& Preprocessing}\label{subsec:data_collection_preprocessing}
|
|
\subsubsection{Data Collection}\label{subsubsec:data_collection}
|
|
The dataset used in this work was collected by \textit{VivoSensMedical GmbH}, a medical technology company based in Leipzig, Germany,
|
|
specializing in fertility tracking devices and applications.
|
|
Each entry in the dataset is derived from temperature measurements recorded by the \textit{OvulaRing} wearable
|
|
device\cite{alexander_fertilitatsmonitoring_2014}, which continuously measures core body temperature every 5 minutes.
|
|
The sensor is worn intra-vaginally and is designed for extended use, requiring removal only for data synchronization.
|
|
This enables high-resolution, long-term temperature tracking, providing a detailed view of menstrual cycle patterns.
|
|
|
|
% add a few example cycles
|
|
% TODO: stats about dataset
|
|
The dataset used for training and evaluation consists of 45000 cycles,
|
|
|
|
In addition to raw temperature values, the dataset includes user-specific metadata, such as health status, weight, and age.
|
|
Users can also manually input time-dependent markers, indicating events such as menstruation, ovulation, and intercourse,
|
|
which serve as contextual features for the model.
|
|
While these additional data points can enhance predictive accuracy, it is important to note that they are self-reported
|
|
and may be subject to errors or biases.
|
|
|
|
\subsubsection{Data Preprocessing}\label{subsubsec:data_preprocessing}
|
|
Not all cycles in the dataset can be used for training.
|
|
Cycles that are too short are excluded, as they usually do not contain enough information to make reliable predictions.
|
|
Additionally, cycles that are too long are also excluded, as they might contain multiple ovulation events or
|
|
pregnancies, which would make the prediction task ambiguous.
|
|
\subsubsection{Feature Engineering}\label{subsubsec:feature_engineering}
|
|
|
|
%TODO: show that a model without marker context and one with them is trained, to show the impact of the markers
|
|
|
|
%TODO: show that for initial testing and fine tuning the dataset was reduced to 10% of size to speed up early impressions of performance
|
|
%TODO: show that initial feature set with only temperature as observable and only pregnancy "chance" as target
|
|
% did not show promise on small 10% dataset, so the features got extended
|
|
% show the modifications that were made to account for multi target prediction
|
|
% show, the mothod of circumventing the non-padding-implementation by using a dedicated feature
|
|
% show, that a different loss / temporal loss wheighting or position aware loss could help
|
|
% show, the problem of the model using its past targets, even though they are not observable and how I dealt with this -> explain and use exogenous
|
|
% introduce change of quantiles to single variable output and why, quantiles loss other than .5 was almost always just 0 or 1, so no significant predicitve quality in this, and this only makes predictions harder for model
|
|
% explain why class imablances are there, and what I did to counteract this, also explain, why special weighting is useful for something like the pregnancy risk prediction
|
|
% show the work that needed to be done for inference -> padding values for future observables
|
|
|
|
|
|
|
|
% show, that for very regular cycles, no sophisticated methods are necessary
|
|
|
|
|
|
Notes for training:
|
|
- one head 512 hidden size make very "smoothed" out curves, which result in ver conservative predictions
|
|
|
|
Intermediate training results:
|
|
- first full training:
|
|
- first feature config with ov over and fertility did not show much promise, as it detected the ov too late, essentially when it was already over
|
|
- either there is no patterns detectable in advance, or there are not enough features yet
|
|
- soft of "soft ceiling" at 0.3 loss, many models get there quickly and then "bounce" around
|
|
|
|
\subsection{Time-Series Modeling Approach}\label{subsec:time-series_modeling_approach}
|
|
\subsubsection{Limitations of Traditional Time-Series Models}\label{subsubsec:limitations_of_traditional_time-series_models}
|
|
\subsubsection{Temporal Fusion Transformer}\label{subsubsec:temporal_fusion_transformer}
|
|
\subsubsection{Input \& Output Modeling}\label{subsubsec:input_output_modeling}
|
|
|
|
% build in the probability curve for the pregnancy chance
|
|
|
|
\subsection{Model Training}\label{subsec:model_training}
|
|
\subsubsection{Training Setup}\label{subsubsec:training_setup}
|
|
\subsubsection{Hyperparameter Tuning}\label{subsubsec:hyperparameter_tuning}
|
|
\subsubsection{Training Details}\label{subsubsec:training_details}
|
|
|
|
\subsection{Evaluation}\label{subsec:evaluation}
|
|
\subsubsection{Evaluation Metrics}\label{subsubsec:evaluation_metrics}
|
|
\subsubsection{Baseline Comparisons}\label{subsubsec:baseline_comparisons}
|
|
\subsubsection{Explainability \& Interpretability}\label{subsubsec:explainability_interpretability}
|