temperature-based-fertility-prediction-thesis/thesis/sections/background.tex

%! Author = alex
%! Date = 3/7/25


\section{Background}\label{sec:background}

\subsection{Physiological Background}\label{subsec:physiological_background}

\subsubsection{Menstrual Cycle}\label{subsec:menstrual_cycle}
The menstrual cycle describes the physiological changes in the female body that prepare it for pregnancy.
It is divided into two phases: the \textbf{follicular phase} and the \textbf{luteal phase}.
\\
During the follicular phase, the ovarian follicles mature, and the endometrium (the inner lining of the uterus) thickens
in preparation for a potential implantation of a fertilized egg.
Around day 14 of a typical cycle, ovulation occurs, marking the transition to the luteal phase.
Ovulation refers to the rupture of the mature ovarian follicle and the release of an egg cell into the fallopian tube.

Ovulation is triggered by a surge in \textbf{luteinizing hormone (LH)}
and \textbf{follicle-stimulating hormone (FSH)}, following a peak in estradiol levels.
As ovulation occurs, estradiol levels drop, progesterone levels begin to rise and a slight increase in body temperature
(typically around 0.5°C) can be observed.
This marks the beginning of the luteal phase.

During the luteal phase, the endometrium thickens further, creating an optimal environment for embryo implantation.
LH and FSH levels decrease, while progesterone remains elevated to support endometrial maintenance.
If fertilization does not occur, progesterone levels drop, leading to the shedding of the endometrial lining along
with the unfertilized egg.
This process, known as menstruation, marks the beginning of a new cycle.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.6\textwidth]{background_menstrual_cycle_physiology}
    \caption{Physiological changes during the menstrual cycle~\cite{pedroso_menstrual_2022}.
    The figure shows the relationship between hormonal fluctuations, physiological changes, and body temperature, with a clear temperature rise around ovulation.}

    \label{fig:background_menstrual_cycle_physiology}
\end{figure}

The menstrual cycle typically lasts around 28 days, with ovulation occurring near the midpoint.
However, variations, particularly in the follicular phase length, are common and can be influenced by factors such as stress, diet, exercise and age~\cite{silberstein_physiology_2000}.
Figure~\ref{fig:background_menstrual_cycle_physiology} provides a detailed overview of the hormonal and physiological changes throughout the menstrual cycle.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.9\textwidth]{background_labeled_cycle}
    \caption{Body core temperature curve across a menstrual cycle. The red line represents a locally smoothed temperature trend.
    Menstruation, ovulation, and the fertile phase are indicated in red, blue, and green, respectively.}
    \label{fig:background_labeled_cycle}
\end{figure}

Figure~\ref{fig:background_labeled_cycle} shows the temperature curve over the course of a menstrual cycle with the
menstruation, fertile phase and ovulation marked.
It starts with a menstruation and ends just before the next menstruation.
The follicular phase starts at the beginning and goes on until the ovulation.
The luteal phase begins at the ovulation and continues until the next menstruation.

Not every cycle results in ovulation—a phenomenon known as anovulation—which leads to a monophasic temperature pattern.
Anovulation can have various causes, including hormonal imbalances, stress, or underlying health conditions~\cite{rosenfield_adolescent_2013}.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.9\textwidth]{background_anovulatory_cycle}
    \caption{Example of a cycle without an ovulation and the resulting absence of a temperature rise}
    \label{fig:background_anovulation}
\end{figure}


Anovulation is reflected in temperature data as either an absence of a clear temperature rise or a rise
that is insufficient in magnitude or duration to be considered a reliable indicator of ovulation.
Distinguishing between ovulatory and anovulatory cycles is challenging, as the only definitive confirmation of
successful ovulation in a clinical sense is a positive pregnancy test.
Even ultrasound imaging can only confirm that an egg was released from its follicle—not whether it was successfully implanted or fertilized.
Figure~\ref{fig:background_anovulation} shows a cycle that does not have an ovulation, and thus no resulting temperature rise.


To illustrate the diversity of real-world menstrual cycles, Figures~\ref{fig:background_long_cycle} and~\ref{fig:background_short_cycle}
show examples of cycles that are significantly longer or shorter than a normative 28-day cycle.

\begin{figure}[htb]
    \centering
    \includegraphics[width=0.9\textwidth]{background_long_cycle}
    \caption{Example of a long cycle with a length of 111 days}
    \label{fig:background_long_cycle}
\end{figure}

\begin{figure}[htb]
    \centering
    \includegraphics[width=0.9\textwidth]{background_short_cycle}
    \caption{Example of a short cycle with a length of 22 days}
    \label{fig:background_short_cycle}
\end{figure}

These irregularities appear not only on a per-cycle basis, but also across time within the same individual.
Figures~\ref{fig:background_irregular_cycles} and~\ref{fig:background_regular_cycles}
show examples of a woman with an irregular and a regular menstrual cycle pattern, respectively.
Raw body core temperature readings are shown in light blue, with a red line indicating smoothing by local regression.
Vertical dotted black lines mark the beginning of each cycle.
The irregular example highlights how multiple parameters can vary between individuals:
cycle length, timing of ovulation, temperature shift magnitude between phases, and intra-phase temperature variability.
This multidimensional variability underscores the need for adaptive, data-driven models capable of learning personalized patterns---
rather than relying on population-wide assumptions.
For a regular cycle pattern, sophisticated analysis or predictions are often not necessary, as the last ovulation day
can reliably be used as the next.

\begin{figure}[htb]
    \centering
    \includegraphics[width=0.9\textwidth]{background_irregular_cycle_example}
    \caption{Example of a woman with irregular menstrual rhythm}
    \label{fig:background_irregular_cycles}
\end{figure}

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.9\textwidth]{background_regular_cycle_example}
    \caption{Example of a woman with regular menstrual rhythm}
    \label{fig:background_regular_cycles}
\end{figure}

\subsubsection{Fertility Prediction}\label{subsubsec:fertility_prediction}
Throughout the menstrual cycle, the chance of fertilization varies significantly.
An egg cell released from the ovary during ovulation, can be fertilized for up to 24 hours.
However, since male sperm cells can survive up to 6 days inside the female reproductive tract,
the fertile window is typically defined as the five days before ovulation until one day after ovulation~\cite{dunson_day-specific_1999}.
Research by~\citeauthor{dunson_day-specific_1999} has shown that the highest chance of fertilization is around one day before ovulation,
as illustrated in Figure~\ref{fig:background_pregnancy_chance}.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.7\textwidth]{background_pregnancy_chance_over_time}
    \caption{Chance of fertilization depending on the day of the menstrual cycle.
    The highest chance is around one day before ovulation~\cite{dunson_day-specific_1999}.}
    \label{fig:background_pregnancy_chance}
\end{figure}

It is important to note that fertility prediction is inherently dependent on ovulation prediction.
Since the probability of conception is tightly linked to ovulation timing, the accuracy of fertility prediction methods
is constrained by the precision of ovulation detection.
This relationship underscores the necessity of developing reliable ovulation prediction models,
as even small inaccuracies can significantly impact fertility assessments.

\subsubsection{Physiological Signs of Ovulation}\label{subsubsec:physiological_signs}
Several physiological signs correlate with ovulation and can be used for prediction.
As shown in Figure~\ref{fig:background_menstrual_cycle_physiology}, these include hormonal fluctuations (LH and FSH surges) and
changes in body temperature.
Additionally, variations in cervical mucus consistency, salivary ferning patterns,
and electrical resistance of the skin and vaginal mucosa have been observed~\cite{silberstein_physiology_2000}.

Among these, ultrasonography provides the most accurate confirmation of ovulation by detecting follicular changes,
but it is costly and requires specialized equipment.
Hormone measurements in urine and blood are widely used and available in at-home test kits,
but they require frequent testing.
Temperature-based methods, particularly body temperature tracking, offer a non-invasive alternative by
detecting the slight temperature rise that follows ovulation.
Advances in wearable technology have further enabled continuous and automated temperature monitoring,
improving accessibility and usability~\cite{alexander_fertilitatsmonitoring_2014, luo_detection_2020, yu_tracking_2022}.

\subsection{Data Source and Characteristics}\label{subsec:data_background}
This study is based on a dataset collected from users of the \emph{OvulaRing}~\cite{noauthor_ovularing_nodate},
an intravaginal wearable sensor developed by VivoSensMedical GmbH, located in Leipzig, Germany~\cite{noauthor_vivosens_nodate}.
The device continuously records intravaginal core body temperature at 5-minute intervals.
The sensor itself measures approximately 1\,cm $\times$ 1\,cm $\times$ 2\,cm and is embedded in a silicone ring with a diameter of 5\,cm for ease of use.
It pairs with a smartphone via Bluetooth to synchronize and upload recorded data to a secure database.
Figure~\ref{fig:background_ovularing} shows an image of the ring attached to its silicone ring.

The product has been on the market for over a decade, resulting in an extensive longitudinal dataset of menstrual cycles.
Cycle boundaries are defined by self-reported menstruation, which users manually log in the accompanying app to mark the beginning of each cycle.

Thanks to a battery life of at least six months, the device supports continuous monitoring of long and irregular cycles, enabling the capture of highly variable menstrual patterns.

A known limitation of manual cycle annotations is the potential for misalignment.
Intermediate bleeding events unrelated to menstruation (e.g., ovulatory spotting or irregular shedding) or missing menstruation entries can lead to ambiguous cycle definitions.
Therefore, all user-entered cycle starts undergo manual review to reduce annotation errors.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.3\textwidth]{ovularing}
    \caption{The OvulaRing sensor attached to its silicone ring~\cite{noauthor_ringpng_nodate}.}
    \label{fig:background_ovularing}
\end{figure}

In addition to temperature measurements, the database includes contextual metadata such as age, height, weight, and optional user-entered markers.
These markers provide further physiological context and may include information about intermediate bleeding, sexual intercourse, or positive pregnancy tests.

At the time of writing, the dataset contains approximately 65{,}000 annotated cycles, comprising more than 350 million individual temperature measurements.

\subsubsection{Dataset Summary}
For the present study, the dataset was reduced to approximately 40{,}000 cycles after filtering out entries that were incomplete,
contained hardware-related anomalies, or fell outside a reasonable cycle length range.

Very short cycles typically result from incorrect cycle start entries or premature termination of temperature recordings.
Extremely long cycles are often due to data entry errors or pregnancy-related recordings,
where the sensor was worn continuously throughout gestation—sometimes producing sequences up to nine months long.

While such cases may still contain useful information, they were excluded from this analysis to avoid complications in preprocessing and labeling.
In most instances, only a small portion of these extended cycles contributes meaningfully to the study objectives.
The cutoff values for cycle length are 10 and 150 days, respectively.

The cleaned dataset has 40{,}266 menstrual cycles from 6{,}245 users.
The median number of cycles per user is 4 (IQR: 2--8) and the median cycle length is 28 days (IQR: 26--32).

The average data density—defined as the fraction of available measurements out of the theoretical maximum of 288 measurements per day—is 0.90.
This corresponds to an average data availability of 90\% per cycle, with an average loss of 10\%.

It should be noted, that both ovulation day and anovulation estimates are based on retrospective algorithmic inference, not human labels.
Further details are provided in section~\ref{sec:methodology}.

\subsubsection{Irregularities and Confounding Factors}
Core body temperature is influenced by various factors unrelated to the menstrual cycle.
Illnesses—especially those involving fever—can significantly affect temperature patterns.
This poses a challenge for any analysis relying on temperature data, as one of the key physiological indicators of ovulation is a post-ovulatory temperature rise (see Section~\ref{subsubsec:physiological_signs}).

Figure~\ref{fig:background_fever_cycle} shows an example of a cycle where an illness caused a marked increase in temperature.
This event is particularly problematic because the fever-induced rise occurs just before the expected ovulatory shift, potentially confounding ovulation detection.
Distinguishing illness-related changes from cycle-related ones requires models that are sensitive to context and robust to outliers.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.9\textwidth]{background_fever_cycle}
    \caption{Example of a cycle affected by illness, showing a fever-induced temperature rise shortly before the expected ovulatory shift.}
    \label{fig:background_fever_cycle}
\end{figure}

Another common irregularity arises when users temporarily remove the sensor during menstruation, typically for hygiene reasons—even though the device is safe to wear continuously.
This behavior frequently results in missing data at the start of each cycle.
Depending on the total cycle length, this gap can represent a significant portion of the cycle's data.

Figure~\ref{fig:background_menstruation_data_gap} shows an example cycle with a data gap during menstruation.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.9\textwidth]{background_menstruation_data_gap}
    \caption{Example of a cycle with a data gap at the beginning, caused by sensor removal during menstruation.}
    \label{fig:background_menstruation_data_gap}
\end{figure}

%TODO: more stats!

\subsubsection{Privacy}
The dataset used in this study contains sensitive personal health information and is handled with strict privacy safeguards.
All data is pseudonymized and processed exclusively on encrypted devices, ensuring that no identifiable information can be traced back to individual users.
VivoSensMedical does not share user data with third parties; the data is used solely for internal research and product improvement efforts that directly benefit users at no additional cost.

\subsection{Technical Background}\label{subsec:technological_background}

\subsubsection{Time Series Analysis}\label{subsubsec:time_series_analysis}
Time series analysis is a fundamental tool for studying sequential data that evolves over time.
Unlike other data types, time series data has an inherent temporal order, where each data point is associated
with a timestamp, capturing its dependence on past values.
Time series analysis typically serves two main goals:
Understanding the underlying mechanisms that lead to the observed data and predicting future data points based on the
historical information and potentially external factors~\cite{cryer_time_2008}
\\
Time series analysis encompasses various methods, ranging from simple statistical models to complex deep learning architectures.
Classical methods


In the following, we will introduce the two most common approaches used for machine learning on time series data,
LSTMs and transformers.

\subsubsection{RNN and LSTM Networks}\label{subsubsec:lstm_networks}
Recurrent Neural Networks (RNNs) are extensions of classical neural networks that incorporate cyclic connections between neurons.
These recurrent connections allow the network to retain information from previous inputs by feeding the hidden state
from a prior time step into the current one, enabling a form of temporal memory.

In practice, this means that input data is processed sequentially, one step at a time.
At each step \(t\), the input \(x_t\) is combined with the previous hidden state \(h_{t-1}\) to produce a new
hidden state \(h_t\), which contributes to the output \(o_t\).
This process allows the network to learn temporal dependencies and model sequential data effectively~\cite{medsker_recurrent_1999}.

However, this approach has the downside that the model cannot explicitly control how it remembers or forgets information at each step,
limiting its ability to manage long-term dependencies.
During training via backpropagation, the weights of a neural network are updated based on the partial derivatives of the loss function.
The more propagations required (e.g., in deeper networks), the more multiplications are needed to compute gradients for earlier weights.

RNNs are typically trained using \textit{backpropagation through time} (BPTT), in which gradients are propagated through many time steps,
leading to numerical instability.
If the gradients shrink exponentially, the model suffers from the \emph{vanishing gradient} problem;
if they grow exponentially, it results in \emph{exploding gradients}~\cite{hochreiter_vanishing_1998}.
In both cases, learning is significantly impaired.

Exploding gradients can often be mitigated using techniques such as \emph{gradient clipping},
where the magnitude of the gradient is capped---typically within a range of \([-1, 1]\)---to stabilize training.

Figure~\ref{fig:rnn_unfolded} illustrates the unfolded structure of an RNN across three time steps.
This technique, known as \emph{unfolding}, clarifies how sequential inputs update the hidden state and generate
outputs at each time step.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.8\textwidth]{recurrent_neural_network_unfold}
    \caption{Schematic diagram of the unfolded structure of a recurrent neural network~\cite{fdeloche_english_2017}}
    \label{fig:rnn_unfolded}
\end{figure}

To address the vanishing gradient problem and enable better long-term memory,
\emph{Long Short-Term Memory} (LSTM) networks were introduced by \citeauthor{hochreiter_long_1997}~\cite{hochreiter_long_1997}.
LSTMs extend the RNN architecture by incorporating a memory cell and a series of gates that regulate the flow of
information: the \emph{forget gate}, the \emph{input gate}, and the \emph{output gate}.

\begin{itemize}
    \item The \emph{forget gate} determines which information from the previous cell state should be discarded.
    \item The \emph{input gate} controls what new information is added to the cell state.
    \item The \emph{output gate} selects relevant parts of the current cell state to produce the output and the next hidden state.
\end{itemize}

Each gate employs a sigmoid activation function to regulate the flow of information,
allowing LSTMs to preserve and update memory over long sequences.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.6\textwidth]{background_lstm_structure}
    \caption{Architecture of a memory cell \( c_j \) and its gate units \( \text{in}_j \) and \( \text{out}_j \).
    At the center, the self-recurrent linear unit (with weight 1.0) maintains the internal state~\cite{hochreiter_long_1997}.
    See text for more details.}
    \label{fig:lstm_architecture}
\end{figure}

The LSTM unit maintains an internal state \( s_{c_j}(t) \) for each cell $c_j$ that evolves over time to capture long-term dependencies.
Figure~\ref{fig:lstm_architecture} shows the structure of a memory cell along with its input and output gates.
The \emph{input gate} activation \( y_j^{\text{in}}(t) \) controls the influence of the new candidate input \( g(\text{net}_{c_j}(t)) \) on the state update (Eq.~\eqref{eq:state_update}):

\begin{equation}
    s_{c_j}(t) = s_{c_j}(t - 1) + y_j^{\text{in}}(t) \cdot g(\text{net}_{c_j}(t))
    \label{eq:state_update}
\end{equation}

The \emph{output gate} \( y_j^{\text{out}}(t) \), computed from \( \text{net}_{\text{out}_j}(t) \), determines how much of the transformed internal state is passed on as visible output (Eq.~\eqref{eq:cell_output}):

\begin{equation}
    y_j^{c}(t) = y_j^{\text{out}}(t) \cdot h(s_{c_j}(t))
    \label{eq:cell_output}
\end{equation}

\(g\) and \(h\) are differentiable functions, here \emph{tanh}.
The gate activations themselves are derived from net inputs formed as weighted sums of previous outputs \( y^u(t-1) \), as given by:

\begin{align}
    \text{net}_{\text{out}_j}(t) &= \sum_u w_{\text{out}_j,u} \, y^u(t - 1) \label{eq:net_out} \\
    \text{net}_{\text{in}_j}(t) &= \sum_u w_{\text{in}_j,u} \, y^u(t - 1) \label{eq:net_in} \\
    \text{net}_{c_j}(t) &= \sum_u w_{c_j,u} \, y^u(t - 1) \label{eq:net_c}
\end{align}

During training, the weights associated with the inputs \( x_t \) and previous outputs \( h_{t-1} \),
as well as the biases for each gate in a cell, are learned through backpropagation.
To build more expressive models, memory cells can be stacked in multiple layers, and their outputs concatenated or passed sequentially to higher layers.

LSTMs are widely used in biomedical applications due to their capacity to handle sequences of variable length and complexity.
In the context of ovulation prediction, where hormonal patterns exhibit periodicity but also irregularity,
LSTMs are well-suited to learn relevant time-dependent signals from sequential physiological measurements.

While powerful, LSTMs can be computationally intensive and sensitive to hyperparameter tuning.
Therefore, they are often compared with alternative architectures,
including simpler feedforward networks and more recent attention-based models,
to evaluate trade-offs in performance, interpretability, and computational cost.

The next section introduce the \emph{Transformer} architecture, a more recent alternative that forgoes
recurrence in favor of attention mechanisms.

\subsubsection{Transformer Models}\label{subsubsec:transformer_models}

Transformer models are a class of neural architectures that use \emph{self-attention}
to model dependencies in sequential data without relying on recurrence~\cite{vaswani_attention_2017}.
Unlike recurrent neural networks (RNNs), Transformers process input sequences in parallel,
allowing them to model relationships between any pair of input tokens or timesteps directly.
This mitigates the limitations of recurrent models, such as long-term memory constraints
and vanishing gradients.

Originally introduced for machine translation, Transformers have proven broadly applicable to
various sequence modeling tasks due to their flexibility, scalability, and strong performance
on complex temporal patterns.

At the core of the Transformer is the attention mechanism, which enables the model to compute
context-aware representations by weighing the importance of different input positions for each output.
This is achieved through \emph{scaled dot-product attention}, where queries, keys, and values are
linearly projected from the input and used to compute attention scores.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.4\textwidth]{background_transformer_architecture}
    \caption{The Transformer - architecture for an encoder-decoder model~\cite{vaswani_attention_2017}}
    \label{fig:background_transformer_architecture}
\end{figure}


Figure~\ref{fig:background_transformer_architecture} illustrates the original encoder-decoder model introduced by~\citeauthor{vaswani_attention_2017}.

The Transformer architecture consists of two components: an \emph{Encoder} and a \emph{Decoder}.

\paragraph{Encoder:}

The encoder is responsible for encoding the input into a contextualized representation.
In the case of machine translation, this input would be a sentence in the source language.


The input tokens are first mapped to dense continuous vector representations (embeddings).
Since the attention mechanism permutation-invariant---that is, it does not inherently encode the order of tokens in the sequence---
\emph{positional encodings} are added to the token embeddings to provide information about the token positions in the sequence.


Without positional encoding, repeated tokens such as `The` would be indistinguishable
to the model regardless of their location, even if they play different syntactic or semantic roles.
Positional encodings, often based on sinusoidal functions, inject a unique position-dependent signal
into each token, enabling the model to distinguish between identical tokens in different positions.

In this work, we use since and cosine functions of different frequencies:
\begin{align}
    PE_{\text{pos}, 2i} &= \sin\left(\frac{\text{pos}}{10000^{\frac{2i}{d_{\text{model}}}}}\right), \\
    PE_{\text{pos}, 2i+1} &= \cos\left(\frac{\text{pos}}{10000^{\frac{2i}{d_{\text{model}}}}}\right)
\end{align}
where \(i\) is the dimension of the input and \(pos\) is the position in the sequence.
This was done according to the original paper~\cite{vaswani_attention_2017}.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.8\textwidth]{background_attention_mechanism}
    \caption{(left) Scaled Dot-Product Attention.
        (right) Multi-Head Attention consisting of several attention layers running in parallel~\cite{vaswani_attention_2017}.}
    \label{fig:background_attention_mechanism}
\end{figure}

Inside each encoder block, \emph{Multi-Head Attention} is applied to the input sequence.
This mechanism allows the model to jointly attend to information from different representation subspaces at different positions.
The process begins by linearly projecting the input sequence $X \in \mathbb{R}^{n \times d_{\text{model}}}$ into three distinct matrices:

\begin{align}
    Q &= XW^Q \\
    K &= XW^K \\
    V &= XW^V
\end{align}

where $W^Q$, $W^K$, and $W^V$ are learned projection matrices.
Each row of $X$ corresponds to the embedding of one token, and the projections produce the queries ($Q$), keys ($K$), and values ($V$) for each token.

The core operation of attention is then computed as:

\begin{equation}
    \text{Attention}(Q, K, V) = \text{softmax}\left( \frac{QK^\top}{\sqrt{d_k}} \right) V
\end{equation}

The dot product $QK^\top$ measures the pairwise similarity between queries and keys, scaled by the dimensionality $d_k$ to control the magnitude.
The softmax normalizes the scores across each row, producing an attention distribution over all input positions.

Multi-head attention runs this mechanism in parallel across $h$ different heads.
Each head uses different learned projections:

\begin{equation}
    \text{head}_i = \text{Attention}(XW_i^Q, XW_i^K, XW_i^V)
\end{equation}

The resulting vectors from all heads are concatenated and projected:

\begin{equation}
    \text{MultiHead}(X) = \text{Concat}(\text{head}_1, \ldots, \text{head}_h)W^O
\end{equation}

This output is then passed through a position-wise feed-forward network (FFN), applied independently to each token:

\begin{equation}
    \text{FFN}(x) = \max(0, xW_1 + b_1)W_2 + b_2
\end{equation}

To stabilize training, the Transformer uses residual connections and layer normalization:

\begin{align}
    x_1 &= \text{LayerNorm}(X + \text{MultiHead}(X)) \\
    x_2 &= \text{LayerNorm}(x_1 + \text{FFN}(x_1))
\end{align}

In summary, each encoder block consists of:
\begin{itemize}
    \item A multi-head attention mechanism that projects the input into queries, keys, and values.
    \item A feed-forward network that processes the attention output.
    \item Residual connections and normalization applied after both components.
\end{itemize}

\paragraph{Decoder:}

In a sequence-to-sequence Transformer, the decoder generates the target sequence autoregressively, one token at a time, conditioned on the input sequence processed by the encoder.

At inference time, decoding begins with a special \emph{start-of-sequence} token. Each token is embedded and augmented with positional encodings to incorporate information about token order, just as in the encoder.

Each decoder block consists of three main sublayers:

\begin{enumerate}
    \item \textbf{Masked Multi-Head Self-Attention}:
    The decoder first applies multi-head self-attention over its own output sequence.
    To prevent positions from attending to future tokens (which would leak information during training),
    a causal \emph{look-ahead mask} is applied to the attention weights.
    Formally, for input $Y$ to the decoder:

    \begin{equation}
        \text{MaskedAttention}(Q, K, V) = \text{softmax}\left( \frac{QK^\top}{\sqrt{d_k}} + M \right)V
    \end{equation}

    where $M$ is a mask matrix with $-\infty$ in positions that should be masked (future positions), and $0$ elsewhere.

    \item \textbf{Cross-Attention (Encoder-Decoder Attention)}:
    The output from the masked self-attention is used to attend to the encoder’s representations.
    This allows the decoder to incorporate context from the entire source sequence.
    The queries come from the decoder hidden states, while the keys and values come from the encoder outputs:

    \begin{equation}
        \text{CrossAttention}(Q, K, V) = \text{Attention}(YW^Q, XW^K, XW^V)
    \end{equation}

    where $X$ is the encoder output, and $Y$ is the decoder’s self-attention output.

    \item \textbf{Feed-Forward Network and Normalization}:
    The result of the cross-attention layer is passed through a feed-forward network (FFN),
    followed by residual connections and layer normalization, mirroring the encoder architecture:

    \begin{align}
        y_1 &= \text{LayerNorm}(Y + \text{MaskedMultiHead}(Y)) \\
        y_2 &= \text{LayerNorm}(y_1 + \text{CrossAttention}(y_1, X)) \\
        y_3 &= \text{LayerNorm}(y_2 + \text{FFN}(y_2))
    \end{align}
\end{enumerate}

Finally, the output of the last decoder block is projected into the vocabulary space using a learned weight matrix $W_{\text{vocab}}$
and passed through a softmax function to produce a distribution over the next possible token:

\begin{equation}
    P(\text{next token} \mid \text{previous tokens}) = \text{softmax}(y_3 W_{\text{vocab}}^\top + b)
\end{equation}

During inference, the model generates tokens one at a time, feeding each predicted token back into the decoder to generate the next.
This continues until a special \emph{end-of-sequence} token is produced.

Multiple decoder blocks can be stacked to increase model depth and expressivity.
While originally developed for machine translation, the Transformer decoder has since been applied to a wide range of tasks,
including time-series forecasting and biomedical modeling~\cite{wu_deep_nodate,zeng_are_2022}.
Its ability to model complex, long-range dependencies without recurrence makes it well-suited to domains like biomedical time-series,
where signals are often irregular and span diverse temporal resolutions.


\subsubsection{Convolutional Layers as Temporal Feature Extractors}
For high-resolution time-series data, the input dimensionality can become large,
especially in models like Transformers that process the entire sequence in parallel.
This can lead to increased memory consumption and slower training.
To mitigate this and retain as much information as possible, convolutional layers can be used
to reduce the sequence length while preserving important local patterns.

In this context, one-dimensional convolutions act as learnable filters that slide over the input sequence to extract temporal features.
Each filter is parameterized to respond to specific local structures in the data, such as peaks, slopes, or short-term motifs.
By adjusting the \emph{stride}—the step size of the convolution—the model can control the degree of downsampling,
effectively reducing the number of time steps passed to subsequent layers.

Additional dimensionality reduction can be achieved using pooling operations, such as \emph{max pooling}, which retains only the maximum value within a given window.
These techniques reduce the computational load while maintaining salient information for downstream processing.

Figure~\ref{fig:background_convolution_example} illustrates a simple one-dimensional convolution applied to a sequence using a filter of size 3.
The stride determines how far the filter moves at each step, affecting both the resolution and length of the resulting feature map.

\begin{figure}[htbp]
    \centering
    \includegraphics[width=0.6\textwidth]{background_convolution_example}
    \caption{Example of a simple 1-D convolution on an input sequence.}
    \label{fig:background_convolution_example}
\end{figure}