TIDE: An Early Prediction System for ICU Delirium Based on Graph Neural Networks

This post introduces the TIDE project (original project title: TIDE: Predicting ICU Delirium Onset from Time Series Using a Patch-Based Graph Neural Network). The project applies the T-PatchGNN architecture to early prediction of ICU delirium, using physiological monitoring data from the first 24 hours after patients are admitted to the ICU in the MIMIC-IV database to achieve end-to-end prediction. Using only structured EHR data, the system reaches an AUROC of 0.829, providing interpretable intelligent support for early clinical intervention. The project is maintained by stephenyang5 and published on GitHub (link: https://github.com/stephenyang5/TIDE) on June 13, 2026.

ICU delirium is a common neuropsychiatric complication in intensive care units, affecting approximately one-third of critically ill patients. It is directly associated with prolonged hospital stays, increased mortality, and long-term cognitive impairment. Early prediction faces multiple technical challenges:

1. Irregularity of data collection: Significant differences in recording frequencies of indicators such as vital signs and laboratory tests;
2. Class imbalance: The prevalence of delirium is only 7%-13%, so simply predicting "no delirium" has no clinical value;
3. Complexity of label definition: It requires meeting both CAM-ICU positivity and RASS score ≥ -3, which demands the model to understand the evaluation time window and logical relationships.

The TIDE project develops an end-to-end machine learning system, adapting to irregular multivariate time series data based on the T-PatchGNN architecture.

Data Pipeline Engineering

Covers more than 26,000 ICU admission records, including:

• Queue screening module: Includes patients with first ICU admission, age ≥18 years, and hospital stay ≥24 hours; excludes early deaths and specific comorbidities;
• Feature engineering module: Extracts 57-dimensional features (vital signs, laboratory tests, sedatives/vasoactive drugs), using label-based drug ID parsing and Last Observation Carried Forward (LOCF) imputation strategy;
• Data construction module: Provides a command-line interface to implement end-to-end feature set construction.

Patch Data Conversion

Core innovation: Convert 24-hour data into a (V, P, L) tensor structure (V = variable dimension, corresponding to 57 monitoring indicators; P = time block dimension, 24-hour window divided into 3 8-hour blocks; L = number of observation points in each block), implementing a three-level masking mechanism: point-level masking distinguishes between real observations and forward-filled values, block-level masking handles missing 8-hour periods, and hospital-level masking marks complete records that do not meet standards.

TIDE uses the T-PatchGNN architecture for binary classification adaptation, which is more lightweight than DeLLiriuM (a large language model with 345 million parameters) but has comparable performance.

Data Processing Flow

Original ICU observation data (irregular, multivariate) is first split into 8-hour time blocks for each variable. For a 24-hour prediction window, each variable generates 3 time blocks.

Encoder Design

• TTCN meta-filter encoder: Processes variable-length time blocks and observation masks, adaptively handling different sampling frequencies of variables and avoiding information loss from traditional timestamp alignment;
• Transformer layer: Captures temporal dependencies within sequences, learning the evolution patterns of the same monitoring indicator across different time blocks;
• Adaptive Graph Convolutional Network (GCN): Models correlations between variables, automatically learning which physiological indicators tend to change simultaneously and which combinations of indicators are most indicative for delirium prediction.

Key Design Decisions

• Honest normalization: Only uses training set to calculate normalization parameters to prevent data leakage;
• Class-weighted loss: Uses weighted binary cross-entropy loss for the 13% positive rate;
• Pre-LOCF mask: Generates point-level masks from pre-imputation data, allowing the model to distinguish between real observations and filled values.

The project was evaluated on 26,345 ICU inpatients, of which 3,460 were delirium-positive (prevalence 13.1%).

Main Performance Metrics

A single stratified split (random seed 42) and Bootstrap method with 200 iterations were used to calculate the 95% confidence interval:

Configuration	Test AUROC	95% Confidence Interval
Full feature set (57 dimensions)	0.829	[0.81, 0.85]
Excluding CAM/RASS features	0.780	[0.75, 0.81]

Comparison with Existing Research

The DeLLiriuM (2025) study reported approximately 78.1% AUROC for structured EHR deep learning and approximately 82.5% AUROC for a 345 million-parameter LLM in external validation. TIDE reaches an AUROC of 0.829 using only structured EHR data, close to the LLM benchmark but with a significantly smaller model size. Notably, TIDE's conservative baseline (~0.80) is comparable to structured EHR methods, indicating the efficiency advantage of graph neural network architecture in medical time series tasks.

The project currently has the following limitations:

1. Single database limitation: Trained and validated only on MIMIC-IV, lacking external validation. Differences in data distribution across hospitals may affect model generalization;
2. Computational resource constraints: Due to resource limitations, multi-random seed cross-validation was not performed, which limits the evaluation of result stability;
3. Interpretability research: The project is developing an interpretability tool suite to verify whether the model learns clinically meaningful signals rather than data artifacts.

The core value of the TIDE project lies in proving the effectiveness of lightweight graph neural network architecture in complex medical prediction tasks. Compared to large language models requiring 345 million parameters, TIDE achieves comparable prediction performance with lower computational overhead. For clinical workflows, this efficiency advantage means:

• Faster inference speed: Can be deployed in real-time or near-real-time scenarios;
• Lower hardware requirements: Suitable for medical environments with limited resources;
• Better interpretability: Smaller models are usually easier for feature importance analysis. More importantly, the system can generate risk scores based on routine monitoring data from the first 24 hours after patient admission, providing clinicians with an early warning window to support targeted prevention and intervention measures.