HealthMap: Transforming Clinical Time-series into Visual Representation for Predictive Modeling

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Deep Learning

Background:

• Accurate mortality prediction in intensive care units (ICUs) is vital for patient care.
• Recurrent neural networks (RNNs) are commonly used for time series data but struggle with complex ICU data.
• Limitations exist in capturing long-term and heterogeneous temporal patterns with traditional sequence models.

Purpose of the Study:

• To develop and evaluate a vision-based deep learning model for ICU mortality prediction.
• To transform ICU time series data into heatmap representations for analysis.
• To compare the performance of the vision-based model against traditional RNN-based methods.

Main Methods:

• A deep learning model utilizing convolutional neural networks (CNNs) was developed.
• ICU time series data was converted into heatmap representations.
• The model employed spatial feature extraction to capture temporal trends.

Main Results:

• The vision-based model achieved superior performance compared to RNN-based methods.
• Key performance metrics included an AUROC of 0.82 and AUPRC of 0.39.
• The model demonstrated high overall classification accuracy for mortality prediction.

Conclusions:

• Heatmap representations effectively encode time series dynamics for mortality prediction.
• Vision-based models offer a promising alternative to sequence models for ICU risk stratification.
• This approach provides clinicians with an intuitive visual tool for identifying patient deterioration patterns.