Download as pdf or txt
Download as pdf or txt
You are on page 1of 1

I. M ETHODOLOGY number of predicted horizons.

This approach offers two no-


The primary objective is to predict patients’ vital signs in table benefits: it shortens the time series length, simplifying
order to provide doctors with advanced indicators of potential long-term predictions, and it optimizes storage capacity and
future changes. These indicators will aid doctors in assessing computational complexity.
the likelihood of sepsis or septic shock occurrence. To achieve
this, data were leveraged from the eICU database, specifically
focusing on critical parameters such as blood pressure. The
approach involves segmenting the data into 9-hour intervals.
Within these intervals, the initial six hours were utilized to
predict the vital signs for the subsequent three hours. The
methodology encompasses the utilization of the following
models:
Fig. 3. N-Beats structure.

C. Temporal Fusion Transformers (TFT)


TFT combines elements of Gated Recurrent Unit (GRN),
Long Short Term Memory (LSTM), and Attention mecha-
nisms. TFT offers a range of advantages:
• Static Covariate Encoder: This element encodes context
vectors, which are then shared across various segments
Fig. 1. Pipeline structure.
of the network.
• Feature Selection via Gating Mechanism: This mecha-

A. Neural Hierarchical Interpolation for Time Series Forecast- nism minimizes the influence of irrelevant inputs.
ing (N-Hits) • Sequence-to-Sequence Layer: It facilitates local process-
ing of time-varying variables, encompassing both past
The N-Hits model were applied due to its ability to effec- and future known variables.
tively handle time series data. Unlike conventional statistical • Temporal Self-Attention Decoder: This component is cru-
models, N-Hits employs deep learning techniques to analyze cial for capturing long-term dependencies within datasets.
various components of the time series, including trends, peri- Additionally, TFT supports interpretability, aiding users
ods, and noise. The structure of the N-Hits model comprises in identifying global important features, temporal pat-
stacked blocks, where each block’s purpose is to capture terns, and significant events.
residual patterns from the previous block. These blocks consist
of interconnected fully connected layers, with each block
yielding both forward and backward predictions. To enhance
nonlinearity, an activation function is applied before outputting
predictions.

Fig. 4. TFT structure.

R EFERENCES
[1] Challu, C., Olivares, K. G., Oreshkin, B. N., Ramirez, F. G., Canseco,
Fig. 2. N-Hits structure. M. M., Dubrawski, A. (2023, June). NHITS: Neural Hierarchical
Interpolation for Time Series Forecasting. In Proceedings of the AAAI
Conference on Artificial Intelligence (Vol. 37, No. 6, pp. 6989-6997).
[2] Lim, B., Arık, S. Ö., Loeff, N., Pfister, T. (2021). Temporal fusion
B. Neural basis expansion analysis for interpretable time transformers for interpretable multi-horizon time series forecasting.
series forecasting (N-beats) International Journal of Forecasting, 37(4), 1748-1764.
[3] Oreshkin, B. N., Carpov, D., Chapados, N., Bengio, Y. (2019). N-
N-Beats model were also be experimented, which bears BEATS: Neural basis expansion analysis for interpretable time series
a resemblance to N-Hits but introduces some distinct ele- forecasting. arXiv preprint arXiv:1905.10437.
ments. Notably, the H-Beats block within the N-Beats model
incorporates an additional layer of max pooling to facilitate
down sampling. This maneuver is followed by interpolation
processing, which aligns the number of outputs with the

You might also like