Abstract
Trajectory forecasting in healthcare data has been an important area of research in precision care and clinical integration for computational methods. In recent years, generative AI models have demonstrated promising results in capturing short and long range dependencies in time series data. While these models have also been applied in healthcare, most of them only predict one value at a time, which is unrealistic in a clinical setting where multiple measures are taken at once. In this work, we extend the framework temporal fusion transformer (TFT), a multi-horizon time series prediction tool, and propose TFT-multi, an end-to-end framework that can predict multiple vital trajectories simultaneously. We apply TFT-multi to forecast 5 vital signs recorded in the intensive care unit: blood pressure, pulse, SpO2, temperature and respiratory rate. We hypothesize that by jointly predicting these measures, which are often correlated with one another, we can make more accurate predictions, especially in variables with large missingness. We validate our model on the public MIMIC dataset and an independent institutional dataset, and demonstrate that this approach outperforms state-of-the-art univariate prediction tools including the original TFT and Prophet, as well as vector regression modeling for multivariate prediction. Furthermore, we perform a study case analysis by applying our pipeline to forecast blood pressure changes in response to actual and hypothetical pressor administration.
Introduction
In the Intensive Care Unit (ICU), clinicians rarely interpret vital signs in isolation. A drop in blood pressure is clinically significant when viewed alongside changes in heart rate or oxygen saturation. However, most current AI forecasting tools are univariate, predicting only one variable at a time. This approach ignores critical physiological correlations and often struggles with the irregular, "messy" nature of ICU data, where some vitals are recorded frequently and others are sparse.
TFT-multi addresses these gaps by treating the patient's physiological state as an interconnected system. By forecasting multiple vital signs simultaneously, the model leverages high-frequency signals to "anchor" the predictions of more volatile or sparsely sampled variables.
Methods
The researchers developed TFT-multi by extending the Temporal Fusion Transformer (TFT) architecture. Key technical innovations include:
- Multivariate Output: The final layers are modified to output simultaneous quantile predictions (10th, 50th, and 90th percentiles) for Mean BP, Pulse, SpO2, Respiratory Rate, and Temperature.
- Observation Masking: To handle missing data, the joint loss function only accounts for time points where real measurements exist, preventing the model from learning artifacts from simple data-filling methods.
- Counterfactual Forecasting: The model accepts "known future" inputs, such as planned medication doses, enabling clinicians to simulate how a patient might respond to different treatment scenarios.
Key Results
TFT-multi was validated on the MIMIC-IV dataset (9,638 patients) and an independent institutional cohort (35,286 patients). It consistently outperformed state-of-the-art baselines, including the original univariate TFT and Facebook’s Prophet.
- Predictive Accuracy: In external validation, TFT-multi achieved significantly lower Mean Absolute Error (MAE) across 4 out of 5 vitals. For instance, Mean BP error was reduced to 6.2, compared to 10.1 for the original TFT.
- Handling Missingness: The improvement was most pronounced in variables with high missingness, such as SpO2, where univariate models often failed to capture meaningful trends.
- Calibration: Bland-Altman analysis confirmed the model is exceptionally well-calibrated for Mean BP and Pulse, with predicted values closely matching ground truth across clinical ranges.
Discussion and Future Work
While TFT-multi represents a significant advancement in multivariate forecasting, some challenges remain. The model currently struggles to track rapid, frequent fluctuations in highly sparse data like SpO2. Additionally, while the model can simulate medication effects, future iterations will need to incorporate specific dosages rather than binary (on/off) logic to enhance clinical utility.
The ultimate goal for this framework is integration into real-time early warning systems, providing ICU staff with a holistic view of potential patient instability hours before it occurs.
Summary updated