💻 computer science

Risk Stratification for ICU Delirium using Pervasive Ambient Sensing Information

This study demonstrates that passive ambient sensing of sound and light, particularly sound pressure levels analyzed via efficient sequential neural networks, can effectively predict ICU delirium risk across multiple time horizons, offering a practical and interpretable tool to enhance early detection and prevention strategies.

Original authors: Jiaqing Zhang, Sabyasachi Bandyopadhyay, Miguel Contreras, Jessica Sena, Yuanfang Ren, Andrea Davidson, Ziyuan Guan, Tezcan Ozrazgat-Baslanti, Subhash Nerella, Azra Bihorac, Parisa Rashidi

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiaqing Zhang, Sabyasachi Bandyopadhyay, Miguel Contreras, Jessica Sena, Yuanfang Ren, Andrea Davidson, Ziyuan Guan, Tezcan Ozrazgat-Baslanti, Subhash Nerella, Azra Bihorac, Parisa Rashidi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Intensive Care Unit (ICU) as a high-stakes, 24-hour control room where patients are fighting for their lives. Sometimes, these patients develop a condition called delirium—a sudden state of confusion, disorientation, and hallucination. It's like their brain's "operating system" crashes. This is dangerous, costly, and hard to predict before it happens.

Currently, doctors try to spot delirium by asking patients questions (like a pop quiz on attention and memory). But this is like trying to catch a thief only after they've already broken the window; it's too late to prevent the damage, and it only happens a few times a day, leaving huge gaps in monitoring.

This paper asks a different question: Can we predict a brain crash by listening to the room and checking the lights, without ever asking the patient a single question?

The Experiment: Eavesdropping on the ICU

The researchers set up a "listening and watching" experiment in 9 different ICUs. They didn't look at medical charts or blood tests. Instead, they treated the hospital room like a stage and recorded two things:

The Sound: How loud was the room? Was it a constant hum, sudden beeps, or chaotic noise?
The Light: How bright was the room during the day versus the night?

They gathered this "ambient" data from 309 patients over several days. Think of this data as a continuous diary of the room's atmosphere, written in numbers rather than words.

The Detective Work: Teaching Computers to Listen

The team fed this diary of sounds and lights into four different types of "digital detectives" (neural networks). These computers were trained to look for patterns in the noise and light that usually happen before a patient gets confused.

They tested the computers on 10 different timeframes, asking: "If we look at the data from today, can you predict if the patient will get delirium tomorrow? In 3 days? In a month?"

The Big Findings

Here is what the digital detectives discovered, using some simple analogies:

1. The Sound is the Star Player
If the ICU room were a sports team, sound is the star quarterback. The computer learned that the noise levels in the room were the strongest predictor of delirium.

The Pattern: The model found that delirium often happens when there is a specific mix of sounds: a low, steady background hum (like a quiet night) combined with sudden, sharp foreground noises (like a beeping alarm or a door slamming). It's like a calm lake suddenly getting splashed by a rock; that contrast seems to trigger the brain's confusion.
The Result: Using just sound data, the computer could correctly identify delirium risk about 80% of the time.

2. Light is the Supportive Coach
Light was helpful, but not as powerful as sound. It was like a supportive coach who gives good advice but doesn't make the big plays.

The Pattern: The computer noticed that the amount of light during the day mattered more than the light at night.
The Result: Light alone was okay at predicting risk, but not as accurate as sound.

3. The Power of the "Sound + Light" Combo
When the researchers combined sound and light, the computer didn't get more accurate at guessing the final score (the overall accuracy stayed similar). However, it got faster.

The Analogy: Imagine a weather forecaster. Sound alone might tell you "It will rain next week." But adding light data is like adding a radar; it tells you, "It's going to rain tomorrow morning."
The Result: The combined model was better at flagging high-risk patients immediately after the sensors started recording. It gave an earlier warning signal.

How the Computer "Thought"

The researchers used a special tool called SHAP to peek inside the computer's brain and see why it made its decisions.

It found that daytime noise was the biggest clue. Specifically, the difference between the quietest moments and the loudest moments during the day was a major red flag.
At night, the loudest single noise (like a sudden crash or alarm) was the biggest warning sign.

The Bottom Line

This study is the first to try predicting ICU delirium using only the environment (light and sound) rather than medical records.

What worked best: Listening to the room's noise.
What helped: Adding light data to get an earlier warning.
The takeaway: The environment of the ICU isn't just background noise; it's a signal. By passively listening to the sounds and watching the lights, we might be able to spot a patient's confusion before it even starts, offering a new, practical way to keep patients safe.

Note: The paper emphasizes that this is a proof-of-concept study. While the results are promising, the team notes that they need more data and real-world testing before this can be used as a standard medical tool in hospitals.

Technical Summary: Risk Stratification for ICU Delirium using Pervasive Ambient Sensing Information

Problem Statement
Delirium is a prevalent and serious complication in Intensive Care Units (ICUs), linked to increased morbidity, prolonged hospital stays, and significant healthcare costs (estimated at $4–16 billion annually). Current diagnosis relies on the Confusion Assessment Method ICU (CAM-ICU), which is administered by nurses, leading to sparse measurements (as low as 38% of admissions) and retrospective identification only after symptoms manifest. While environmental factors like ambient light and sound are known to contribute to circadian disruption and delirium, no prior study has developed a risk assessment model specifically utilizing these environmental variables (sound pressure and light intensity) for prospective prediction.

Methodology
The study utilized data from two prospective, single-center studies (PAIN and ADAPT) conducted at the University of Florida Shands Hospital between May 2019 and September 2024.

Cohort: The final ICU-ENV dataset comprised 309 patients (247 in development, 62 in testing) who were over 18 years old, admitted to the ICU, and expected to stay at least 24 hours.
Data Collection:
- Ambient Sensors: Light intensity was collected via ActiGraph GTX3+ (PAIN study) and Thunderboard Sense 2 (ADAPT study). Sound pressure levels were collected via iPod Touch with AudioTools (PAIN) and Thunderboard Sense 2 (ADAPT).
- Ground Truth: Delirium status was derived daily from electronic health records using a composite algorithm based on the Richmond Agitation Sedation Scale (RASS), Glasgow Coma Scale (GCS), and CAM-ICU scores.
Preprocessing:
- Light Data: Averaged into "daytime" (07:00–18:59) and "nighttime" (19:00–06:59).
- Sound Data: Converted into seven statistical descriptors: Maximum ( $L_{max}$ ), Minimum ( $L_{min}$ ), and percentiles ( $L_{99}, L_{90}, L_{50}, L_{10}, L_{01}$ ).
- Windowing: Data was zero-padded to a 7-day observation window. Ten prediction horizons were defined (1–7 days, 14, 21, and 28 days post-observation).
Model Architecture: Four sequential neural network models were evaluated: Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), Convolutional Neural Network (CNN), and Transformer. The CNN model demonstrated superior performance.
Interpretability: Shapley Additive Explanations (SHAP) analysis was employed to determine feature importance and the direction of influence.

Key Results

Model Performance: The Convolutional Neural Network (CNN) achieved the strongest discrimination across all datasets.
- Sound-Only Model: Achieved an AUC of 0.80, with peak performance in the 28-day prediction window.
- Light-Only Model: Showed moderate performance, peaking at a 5-day prediction window.
- Combined (Sound + Light) Model: Achieved the highest risk stratification scores at a 7-day prediction window. While the combined model slightly reduced overall classification accuracy compared to the sound-only model, it significantly improved early risk detection.
Risk Stratification: The combined model assigned the highest relative risk (RR) immediately after the sensing period (Day 1), with an RR 1.8 times higher than the sound-only model and 2.6 times higher than the light-only model.
Feature Importance (SHAP Analysis):
- Sound: Sound features were the dominant predictors. Specifically, daytime $L_{90}$ (sustained foreground sound) and nighttime $L_{max}$ were significant positive predictors. Conversely, daytime $L_{min}$ (background sound) was a negative predictor.
- Light: Daytime light intensity was more important than nighttime light intensity.
- Interaction: In the combined model, sound features remained more informative than light features.

Significance and Claims
The authors claim this is the first study to investigate the feasibility of predicting ICU delirium incidence and stratifying risk using only ambient environmental factors (light and sound).

Clinical Utility: The study demonstrates that passive ambient sensing, particularly sound, provides a clinically meaningful and interpretable signal for delirium risk estimation.
Early Intervention: Integrating sound with light data offers a practical pathway to enrich multimodal ICU strategies by assigning higher risk scores at earlier time points (immediately post-sensing), potentially facilitating earlier interventions.
Limitations: The authors acknowledge the modest sample size (309 patients) limits the robustness of the models and prevents training with large parameter counts to avoid overfitting. They also note potential batch effects between sensors, particularly regarding light intensity distributions, which may have contributed to the lower performance of light-only models.

Conclusion
This work establishes a foundation for using pervasive sensing in the ICU to monitor delirium risk. While sound pressure levels proved to be the primary driver of predictive accuracy, the integration of light data enhances the temporal resolution of risk assessment, offering a non-invasive, continuous monitoring approach to complement existing clinical assessments.